Life

Over the past two weeks, we visited Yellowstone and Glacier National Parks, where we saw an incredible variety of wildlife—including bears, bison, deer, and bighorn sheep. I highly recommend making the trip, especially during the summer, when these national parks are at their best. It is a wonderful opportunity to step away briefly from the noise of the markets and the relentless flood of AI-related information, return to nature, take a deep breath, and reset.

The people who work and live in these national parks may seem far removed from AI, yet in some ways they are also very close to it. At times, I find myself admiring their way of life. Market news has little impact on their daily lives and may not matter to them at all. What truly matters is whether we appreciate the grace God gives us each day—nature, people, animals, and everything around us.

I am still learning how to strike that balance myself: how to pursue new technologies and information while maintaining a calm mind, thinking for the long term, and making grounded decisions in the present.

Market: MRVL Offers More Compelling Upside Potential

MRVL Earnings Call

This quarter’s financial results and guidance both exceeded the company’s previous outlook. We believe Marvell’s partnership with Google and the ramp-up of its scale-up business could provide meaningful upside over the next year. The bear case centers on modest gross-margin pressure over the coming quarters. However, we expect continued operating-margin expansion to support a higher valuation.

Financial Performance

Revenue increased to $2.739 billion, up 37% year over year and 13% quarter over quarter, compared with the Q1 outlook of $2.7 billion for Q2. Non-GAAP EPS reached $0.94, with both metrics exceeding the midpoint of guidance.

The company expects Q3 revenue to rise 15% sequentially to $3.15 billion, representing year-over-year growth of more than 50%.

Key Technology and Business Drivers: Google and Optical Connectivity as the Primary Sources of Upside

Google TPU Partnership

Based on the disclosed figures, cumulative revenue could reach $120 billion over six years if all milestones are achieved, equivalent to approximately $18.5 billion annually. Compared with the company’s current annual revenue base of only $15–18 billion, this represents substantial upside. The most significant volume ramp is expected to begin in FY2029, corresponding to CY2028, and continue thereafter.

The partnership primarily involves attach silicon for Google’s TPU ecosystem, including CXL and DPU solutions. The contract covers a broad product portfolio spanning AI inference accelerators, storage controllers, custom NICs, memory-interface controllers, and near-memory computing solutions.

CXL Memory Expansion and Scale-Up Optics

High Potential to Exceed Current Ramp Expectations

Marvell has invested in CXL—or Compute Express Link—development for several years. This quarter, the company announced that it had begun receiving meaningful volume orders.

The key reason is that CXL has proven particularly well suited to memory expansion for AI inference workloads.

Marvell’s CXL memory-expansion chips are currently being deployed at significant scale by several U.S. hyperscalers. Combined with the PCIe Gen 6 and CXL 3.1 switching technologies acquired through XConn, Marvell can offer a complete end-to-end solution consisting of CXL switches, memory expanders, and retimers.

Operating Leverage Expansion

Potentially Misunderstood by the Market

The dilution in gross margin is indeed being driven by the strong ramp of lower-margin custom silicon revenue. However, the company also stated that operating expenses are growing at only half the rate of revenue. This should generate significant operating leverage and could bring the operating margin into its long-term target range of 38%–40% as early as Q4.

NVDA Earnings Call

This earnings call is well worth examining in depth. The key takeaway is that the circular financing cycle can continue—meaning the AI productivity flywheel remains intact. However, when industry leaders must rely on accelerating revenue growth rather than expanding gross margins to drive EPS, and operating margins can no longer improve, their valuation multiples will inevitably face pressure.

Financial Performance and Outlook

1Q27: Near-Term Guidance and Stable Gross Margin

In 1Q27, NVIDIA provided only one quarter of revenue guidance, projecting $91 billion for 2Q27. It also expected full-year gross margin to remain in the mid-to-high 70% range.

2Q27: Significant Pressure From Memory Costs

Faced with exceptionally high HBM and SOCAMM memory prices, NVIDIA lowered its 3Q27 gross-margin guidance to 74% and projected that gross margin would decline to a cyclical low of 71%–72% in 4Q27.

A “Floor” for FY2028 Revenue Guidance

To offset the decline in gross margin, management unexpectedly introduced a strong baseline forecast in 2Q27, calling for FY2028 revenue growth of 70%. The company also emphasized that, absent supply-chain constraints, underlying customer demand would imply a doubling of revenue year over year.

Products and Supply Chain: From Wafer and Packaging Shortages to Physical-Infrastructure Constraints

1Q27: Focus on Architecture and an Entirely New CPU TAM

The primary focus was the rapid adoption of Blackwell systems, including GB300 and NVL72. NVIDIA also highlighted the Vera CPU, which is designed to orchestrate agentic-AI workloads. The company expects Vera to open a new $200 billion total addressable market and guided to $20 billion in CPU revenue for the year.

2Q27: New Architecture Ramps as Physical Bottlenecks Intensify

Vera Rubin officially entered volume production and began shipping in August 2026. It is expected to rapidly grow to approximately 20% of data-center revenue in 3Q27.

We believe NVIDIA’s revenue opportunity per gigawatt of data-center capacity has increased substantially—from $18 billion for Hopper and $25 billion for Blackwell to more than $40 billion for Vera Rubin.

Management also noted that supply-chain bottlenecks have expanded beyond wafers and advanced packaging to physical infrastructure, including electricity, land, server enclosures, liquid cooling, and rack assembly. These constraints are expected to persist through at least the end of FY2028.

Business Model and Ecosystem: The Rise of ACIE and NVIDIA’s Growing Role in Circular Financing

The Rise of Non-Hyperscaler Customers, or ACIE

In 1Q27, NVIDIA broke out its data-center submarkets, grouping AI cloud providers, industrial customers, enterprises, and sovereign customers into the ACIE segment. By 2Q27, ACIE revenue had surged to $40 billion, up 25% sequentially, and is expected to eventually represent approximately half of NVIDIA’s data-center business.

“Hyperscalers will remain a major growth driver, but non-hyperscaler growth—our AIC and E segment, spanning sovereign and regional neoclouds, enterprise edge, and air-gapped data centers—will represent roughly half of our data-center business.”

Launching a Large-Scale Third-Party Financing Platform

Frontier AI laboratories such as OpenAI and Anthropic have enormous infrastructure demand, but their growth is constrained by their balance sheets and financing capacity.

In 2Q27, NVIDIA announced that, in addition to investing nearly $50 billion directly in frontier laboratories, it had partnered with six major financial institutions—Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR—to establish a financing platform capable of raising more than $500 billion in third-party capital. The platform is designed to help customers secure long-term, low-cost financing for infrastructure projects.

Introducing Credit Guarantees and Revenue-Sharing Arrangements

To support sovereign-cloud and regional neocloud operators, NVIDIA will provide take-or-pay commitments covering a portion of a facility’s capacity. These minimum-revenue guarantees give lenders greater confidence in underwriting the projects.

In return, NVIDIA can share financial risk while receiving a portion of the neocloud operator’s rental revenue. This adds highly recurring, usage-linked income on top of the company’s one-time hardware sales.

“NVIDIA provides a take-or-pay commitment on a portion of the facility’s capacity—a minimum revenue guarantee that gives lenders the confidence to underwrite the project. In exchange, we share in a portion of the neocloud’s revenue earned above that floor. We get paid twice: once on the hardware sale and again through our share of rental revenue.”

The financial forecasts below are available exclusively to paid subscribers.

Insightology View

MRVL Financial Forecast: …

Read more

Life

• Paid articles are expected to launch around the end of August: this week at work, I've been gradually organizing my past research and my watchlist. I plan to start rolling out paid articles in the next 1–2 weeks. These will include deeper analysis of individual stocks, my views on industries, the macro economy, and U.S. business, and I'll also be sharing more of my own results on quantitative strategies. If you're interested, feel free to DM me or leave a comment about what you'd like to see.

• During this deleveraging move, I've heard about quite a few young investors who — because they were over-leveraged — lost 80–90%, and some were even forcibly liquidated. And that's actually the mild version. What we should really worry about is that if AI demand starts to slow, many individual stocks could fall by more than 80–90%. That's what we need to prepare carefully for over the next 1–2 years.

Content

• Do we need to worry about CSP cash flow?

• NPO back in focus

• What I read: Risk and Reward

Market

Do we need to worry about CSP cash flow?

The major CSPs have all reported Q2 earnings and held their conference calls. The market’s biggest concern is that these companies are burning cash so fast that one day it might all blow up — especially now that the big players’ free cash flow could turn negative next year, with heavy financing used to cover all that Capex. We took a close look at this.

Apart from Microsoft, Google, Amazon, Meta, and Oracle will all see free cash flow turn negative from 2026 to 2028, with the trough in 2027.

Should we be worried?

All of this Capex is built only after contracts are in place. Looking at it from the RPO angle: Google’s cloud order backlog grew from $462 billion to $514 billion within a single quarter; Oracle’s RPO (remaining performance obligations) stands at $638 billion; and Amazon says that “most” of its 2027 AWS compute capacity has already been reserved by customers.

Current bond market appetite and spread widening are still within a manageable range.

In 2025, the five companies issued a combined $108 billion in debt (24% of Capex). At that time, apart from Oracle, everyone’s FCF was still positive — they started borrowing while cash flow was still positive, raising funds in advance to lock in the money while conditions were good.

For 2026/2027, debt issuance is estimated at $257 billion / $419 billion (31% / 35% of Capex).

For 2028/2029, it’s $421 billion / $432 billion — roughly flat, no longer accelerating.

Although total debt is accelerating — ballooning from $297 billion in 2025 to $1.57 trillion in 2029, more than fivefold — the denominator (earnings power) is also growing (EBITDA up 18–24% per year). As a result, net leverage (on an EBITDA basis) is only 0.5x even in 2029: net debt equals just half a year of EBITDA. For a typical investment-grade company, 1–3x is normal, so 0.5x is very light.

Other financial metrics

Net debt / shareholders’ equity of 18%: the capital structure is still equity-driven.

Interest coverage ratio dropping from 43x to 16x: it’s falling fast, but “earning $16 to pay $1 of interest” is still very safe.

CFO / debt falling from 2.0 to 1.0: even at 1.0, it means one year of operating cash flow could pay off all the debt.

What does the market care about?

Bond market appetite: with just these few issuers flooding the market with debt, will the investment-grade bond market struggle to digest it (an issuer-concentration problem)?

Spread widening: the market is already charging a higher price — 2–4 year spreads went from 30 to 40 bps, 5–7 year from 50 to 60 bps, and 20-year-plus from 108 to 118 bps. It’s not yet at a dangerous level, but borrowing is indeed getting more expensive.

Insightology View:

All of this debt and borrowing rests on the assumption of EBITDA growth of around 20% per year — in other words, on AI actually being monetized, and monetized quickly. If monetization starts to slow in the future (the entry of Chinese LLMs into the competition is one risk worth watching), then the entire financial structure would look very bad. From third-party sites, we can see that Anthropic’s and OpenAI’s ARR are still rising, but if their growth slope starts to flatten, the impact on the market would be far greater than this round of deleveraging.

TickerTrends Research

Anthropic Is Tracking Above OpenAI in ARR, but Its Growth Is Beginning to Decelerate

TickerTrends’ latest private-market tracking puts Anthropic at roughly $74.1 billion in annualized revenue run rate, versus $41.3 billion for OpenAI…

Read more

a month ago · 7 likes · TickerTrends

NPO back in focus

In terms of major industry news over the past week: according to the latest disclosures in August 2026 from outlets such as The Information and Korea’s Chosun Ilbo, Nvidia is considering a significant cut to the high-bandwidth memory (HBM) capacity of “Rubin Ultra,” its next-generation AI chip slated for a 2027 launch — a move the industry has described as entering a “Memory Diet” phase.

We believe that while this does affect memory, the real beneficiary is actually NPO. With less memory, each individual GPU has less data readily on hand, so running large models becomes slower. It's like shrinking the personal toolbox each worker carries around — work efficiency takes a hit. Nvidia's compensating approach isn't to add the toolbox back (that would only be more expensive), but to let more workers communicate in real time and share tools within the same space.

Two ways to make an AI server bigger:

So interconnect is back on the table and deserves more attention. AI server interconnect can be divided into two types:

Scale-up: link a group of GPUs together fast enough that they can be treated as one giant GPU. The NVL72 is exactly this — 72 GPUs in one rack, used as a single machine.

Scale-out: many such machines are then combined into a cluster, but communication between racks is much slower — like making a phone call across floors.

NVL576 favors MoE computation: In its GTC 2026 roadmap, Nvidia showcased the NVL576 — connecting 8 NVL72 racks together via scale-out. This benefits the efficiency of MoE (mixture-of-experts) architectures, because MoE models are made up of many "expert sub-models" that divide the work. These experts are spread across different GPUs and need to exchange data frequently, so the larger the open space and the smoother the communication, the more the drawback of reduced memory gets partly offset.

NVL576 requires optical connections: Inside a single rack, copper wiring (NVLink) between GPUs is enough. But copper has physical limits: as distance grows, the signal attenuates and power consumption spikes, so it basically can’t hold up across racks. To connect 8 racks into one unit, the only option is to go optical — fiber transmits over long distances, offers high bandwidth, and has low loss. So an optical scale-up fabric is the prerequisite for the NVL576 to work.

Within optics, NPO is currently the reliable medium-term path, and CPO will be the ultimate architecture — but the one ramping up first is NPO.

There are two ways to fit the optical engine (OE — the module responsible for converting between electrical and optical signals) into the system:

• CPO (co-packaged optics): the optical engine is packaged directly next to the switch chip, almost soldered together. The upside is the lowest power consumption and lowest latency — in theory the fastest and best architecture; the downside is that advanced packaging is too complex and yields are currently low.

• NPO (near-packaged optics): the optical engine sits on a substrate near the chip, connected via a socket. It captures most of the benefits of optics, but with far better manufacturability and deployment flexibility — if it breaks, you can still swap it out. In short, it’s a modular engine that’s easier to assemble and repair.

Source：https://www.naddod.com/blog/optical-interconnect-technology-analysis-lpo-npo-cpo

Insightology View:

Nvidia is expected to adopt NPO for scale-up in the Rubin Ultra and Feynman platforms. Assuming a 20% penetration rate in 2027, we estimate OE shipments for Nvidia’s platform of 5.6–6.0 million units in 2027 and 8.0–8.2 million in 2028. Adding in the in-house chip platforms of other CSPs, the overall OE market comes to 10–12 million units in 2027 and 40 million in 2028 — YoY growth of nearly 300%.

Details on the value shift to TIA/Driver: NPO removes the DSP (digital signal processor) portion from the optical module, handing the signal-conditioning work over to the TIA (transimpedance amplifier) and the laser driver. Together, these two components are worth tens of dollars in each 3.2T optical engine — considerably more than in a traditional pluggable optical module. The TAM for TIA/Driver comes to $800–900 million in 2027 and $3.0–3.1 billion in 2028, again close to 300% annual growth.

Ranking of beneficiaries: Marvell is the primary beneficiary (best positioned in high-speed TIA/Driver), with Semtech secondary. Among others, LITE, SMTC, TSEM, Bizlink, and Browave will also benefit from the NPO ramp.

What I read：Risk and Reward

This book is perfect to pick up and read every time the market drops or spikes — it settles the FOMO mindset, and the urge to sell right at the very bottom.

I finished this book once in the first quarter of this year, and I reviewed it again last month. Each time, it felt different.

This time, Chapters 12 and 14 left the deepest impression on me.

Volatility Is a Feature, Not a Bug

The stock market is the world’s best “reverse casino”

• Casino vs. stock market: In a real casino, because of the “house edge,” the more a player plays, the more they lose. The stock market is the opposite — the longer you stay in, the better your odds.

• 93% of the time you’re watching your assets shrink: This means that for as much as 93% of the time, an investor’s account is in a “drawdown state,” sitting below its all-time high.

• Detailed drawdown-frequency statistics (1950–2024):

  • Drawdown of 10% or more: 36.1% of the total time (more than a third of the time).

  • Drawdown of 20% or more: 16.4% of the total time.

  • Drawdown of 30% or more: 5.4% of the total time.

  • Drawdown of 40% or more: 2.3% of the total time.

  • Drawdown of 50% or more: 0.1% of the total time.

  • If an investor can’t tolerate their account being underwater so much of the time, they’re destined not to survive in the stock market.

The illusion of the “average return”: stock market returns are extremely uneven and lumpy

• The mean is just a mathematical abstraction: The average annual return of U.S. stocks over the past century is roughly 10%. But in reality, the market never obligingly hands you 10% every year.

• Extreme years make up the average: Across the 97 years of history from 1928 to 2024, only 3 years had an actual return between 9% and 11%.

• The real distribution of annual U.S. stock returns over 97 years:

  • 73% of the time (71 years): the annual return was positive.

  • 27% of the time (26 years): the annual return was negative.

  • 70% of the time (68 years): the annual return was a double-digit surge or plunge.

  • 58% of the time (56 years): the annual return was a double-digit surge (>10%).

  • 43% of the time (42 years): the annual return was an extreme surge or plunge of 20% or more.

  • This means the market’s normal state is “extreme surges and plunges alternating (lumpy returns),” rather than mild, average growth.

Life

I’ve recently been learning how to play bridge with friends. The last time I played bridge intensively was back in high school, and coming back to this game as an adult has given me deeper reflections. What makes bridge so fascinating is that you reason (through bidding) without being able to see all the cards, and you exchange information through bidding—while at the same time both your opponents and your partner learn your information. Beyond calculating probabilities, you also have to work with your partner’s bidding habits and read your opponents’ thinking. Even if you have good cards, you won’t necessarily win. What’s different from Texas Hold’em is that bridge requires more cooperation—you have to build trust with the partner across from you, and much of the information communication is transparent. Texas Hold’em, on the other hand, is more about individual gaming, where information is hidden not only in betting patterns but also in facial expressions and timing.

Buffett is also famously fond of bridge. Beyond both requiring you to infer the underlying truth from limited clues (bidding, card play, market signals, financial statements), bridge rewards those who consistently make correct decisions across a long series of many hands, rather than those who play the hero on a single lucky hand. Although luck heavily influences any single hand, skill emerges over the long run. And because bridge is a partnership game, you have to set aside individual heroism and coordinate with and trust your partner through a limited shared language.

Market

Besides AI, the U.S. auto market has been one of the very strong industries over the past few months.

Looking at new orders for Motor Vehicle Bodies, Trailers, and Parts from the latest U.S. Census Bureau M3 survey, not only did the value of new orders reach a record high, but the YoY figure also rose close to double digits. Over the past 10 months, 9 had positive month-over-month growth, showing strong continuity of momentum. The reason behind this is precisely one of the results of last year’s “tariffs.”

If we break down the underlying data more carefully, we find that the reason new order values hit record highs is because of【price increases】rather than【volume increases】. Behind the price is actually the tariff surcharge. After the 25% Section 232 tariff was imposed on imported parts, Tier-1 suppliers resold the tax-inclusive sub-assemblies onward, and the tariff amount directly inflated the shipment value.

Judging from the earnings calls of major automakers and ODMs, U.S. automakers have gradually begun localizing assembly and procurement. The M3 survey measures shipments of U.S.-based manufacturers. The 25% Section 232 made imported parts structurally more expensive, and OEMs’ response was not to shrink procurement but to shift procurement from plants in Mexico, Japan, and Korea back to U.S. plants—for example, GM will raise its U.S. capacity to over 2 million vehicles starting next year.

Taiwan supply chain companies raise their capital expenditure

ASE Technology Holding (3711) — 7/30 earnings call, third upward revision this year
2026 capex revised up to about $10.5B (from $8.5B), the second upward revision this year, the largest scale in its history, with funds split roughly evenly between plants and equipment. It also purchased AUO’s plant (NT$6.3 billion) and land/plant (NT$5.67 billion). ASE views glass substrates and silicon photonics as the key to breaking through hardware bottlenecks, and capex is very likely to be directed toward LEAP.

Delta Electronics (2308) — 7/30 earnings call
Capital expenditure this year will increase to about NT$70 billion, up roughly 50% from last year’s NT$46.6 billion, with continued plant expansion in Taiwan, China, Thailand, and the U.S. The second half looks better than the first half. 800V DC equipment begins mass production in Q3, but shipments before year-end will be limited, with the real large-scale shipments coming next year. The company also mentioned it has begun shortening the depreciation period for related equipment from 5 years to 3 years, so that if a project changes, risks can be reflected more quickly—it can’t wait for orders to arrive before building plants.

Unimicron (3037) — 7/29 earnings call, third upward revision this year
The board resolved to increase 2026 capital expenditure from the originally planned ~NT$34 billion by NT$19.7 billion to NT$53.7 billion, and simultaneously increased the booking amount for long-lead-time equipment by NT$17.6 billion. Of this, as much as 80% to 85% is focused on ABF.

UMC (2303) — this week’s earnings call
Raised 2026 capital expenditure from the original $1.5 billion to $2 billion (+33%), a rare significant increase for a mature-process foundry. 22/28nm continues to hit record highs, capacity utilization returned to 85%, and this month it completed delivery of its first batch of 12-inch silicon photonics mass-production wafers, with plans to launch a silicon photonics platform available to more customers in 2027.

Quanta: Raised up to $2.2B (NT$71.2 billion) via GDS in Luxembourg, the largest overseas fundraising by a Taiwanese company in nearly 20 years, used for building plants and procuring component materials.

AUO : Investing NT$8.64 billion ($267M) in capex to build a FOPLP glass substrate advanced packaging pilot production line. The three core technologies are Glass Core, RDL, and TGV, with applications targeting LEO satellite antennas, Micro LED, and optical communications.

King Yuan Electronics / Ardentec: NT$28.5 billion ($880M) to expand advanced testing capacity in Longtan .

Linde: Investing $1 billion in Arizona to supply ultra-high-purity nitrogen, oxygen, and argon to two new fabs; its Taiwan joint venture is separately investing $800 million to supply the same customer’s new plant in Taiwan (presumably TSMC).

Insightology View:
This week, besides major players Microsoft, Amazon, and Google successively announcing upward revisions to capital expenditure, Taiwanese supply chain companies also announced continued upward revisions to capex, with many making their second or third revision within the year. Overall, the entire capex cycle has already spread from the wafer end to packaging and testing, power, passive components, panel makers pivoting to packaging, and further upstream to industrial gases.

A long-term warning signal from European interest rates

This is a topic that fewer people in the market mention. In this round of interest rate decision meetings across countries, Europe announced on June 11 that its three key rates would each rise by one quarter-point. Effective June 17, the main refinancing rate, deposit facility rate, and marginal lending rate reached 2.40%, 2.25%, and 2.65% respectively—the first rate hike in nearly three years (since September 2023).

To me this is quite a warning sign, reminding me of Trichet’s rate hike in July 2008. Back then, the EU also hiked rates during a European economic downturn, and likewise hiked due to exogenous cost pressures. It also showed the same policy divergence with the Fed. And now Europe’s economic resilience is not what it was before 2008. If in the future the U.S. begins to see a cyclical slowdown in capital expenditure and the global economy cyclically slows, then EU policy may well be one of the straws adding fuel to the fire.

On risk management & leverage

The biggest news this week was the Situational Awareness (Leopold Aschenbrenner) deleveraging event, in which Citadel took over his stock book at an undisclosed discount.

When it comes to risk management in the markets, if past experience can tell us anything, it’s that when most people feel very “safe” and think they can apply historical experience to how the market works every time, that kind of thinking is extremely dangerous. Because market corrections have never promised to stop only after prices rise or fall to some specific multiple of the PE ratio.

This time, semiconductor supply chain-related stocks in the U.S., Korea, Taiwan, and elsewhere all rose enormously, and many young “stock gods” made a lot of money from this wave (many friends around me saw their net worth jump very quickly in June while still under thirty). But if leverage is used incorrectly, the scariest thing is that it will amplify some extreme event to the point of knocking you out or wiping you out to zero. In every major crash, I hear about different friends, or friends of friends, whose assets came very close to zero—and this time was no exception. Just like the passage from Munger quoted in Berkshire’s 2022 annual report often uses an analogy: a string of beautiful investment return numbers, once multiplied by a single “zero” somewhere in the middle, ends up as zero.

There is no such thing as a 100% sure thing when investing. Thus, the use of leverage is dangerous. A string of wonderful numbers times zero will always equal zero. Don’t count on getting rich twice.
— Charlie Munger, as compiled in Warren Buffett’s 2022 shareholder letter

The cycles of stock prices and fundamentals move at different speeds—one fast, one slow—and combining them while expecting good short-term results is inherently quite difficult. Leverage lowers your margin for error even further. And in my experience, the higher someone’s IQ—even very smart people—the more they tend to misuse leverage, believing it can accelerate wealth accumulation, but often being unable to control its destructive power when it works in reverse.

History tells us that leverage all too often produces zeros, even when it is employed by very smart people.
— Warren Buffett, 2010 shareholder letter

The grandfathers’ wisdom is still very useful!

Life

This week my wife and I went in for a dental checkup — our first time seeing a dentist here in the US — and it really hit home just how much more expensive dental work and extractions are than I’d expected. For the average person without coverage through an employer, a school, or their own extra insurance, one bad tooth can end up costing a small fortune.

The insurance companies here are making an absolute killing. What's interesting is that Delta Dental — the biggest pure-play dental insurer — isn't even publicly traded. Look closely at their business model: every dental plan comes with an "Annual Maximum" (usually only $1,000–$2,000 a year in payouts), plus long waiting periods and steep coinsurance percentages on the expensive procedures like extractions and implants. In other words, the insurer's downside risk is locked in — it's basically guaranteed profit. Honestly, it makes me want to invest in them myself (kidding).

This extraction bill also gave me a real, firsthand feel for the power of "claim caps" and "sky-high medical device pricing" in the US healthcare industry. Even though the pure-play dental insurers aren't publicly listed, that steady, non-negotiable demand is quietly feeding giants like Align (ALGN) and Henry Schein (HSIC) on the dental device and distribution side. Maybe this is another defensive sector — beyond semiconductors — with that "long runway, deep snow" quality worth keeping on our radar.

On the tech research side, I spent a good chunk of time getting my head around the technology behind Kimi 3 — dug through the company's old papers and read up on the founders' backgrounds. It was a week of deep, heads-down thinking. I've also been chewing on where the semiconductor cycle stands right now. The outlook is bright, but between the share-price swings and all the deleveraging, we're starting to see a lot of setups lately that we think are genuinely worth digging into and positioning around.

I'm really thankful that God kept our work, our lives, and our family safe and sound all week.

A Few Thoughts on Kimi 3

The short version: Kimi 3 changed the algorithm, and the real impact is on "cost per task," not the "model's benchmark score." In our own testing on the same task, Kimi 3 came out 2.3x cheaper than Anthropic Fable 5 — the trade-off being that it burns more tokens and runs slower. But since the unit price is only about a third, the total cost still wins. That makes it a good fit for scheduled, run-in-the-background jobs, not real-time conversation.

We think Kimi 3 saves memory, but that doesn't mean your overall memory usage drops. Kimi 3 mixes Kimi Delta Attention with traditional MLA at a 3:1 ratio, so it only saves up to 75% on the KV Cache part — the rest of the inference process still eats memory as usual. And since it burns twice the tokens, that doubled inference run needs the memory just the same.

Finally, Kimi 3 trains with MXFP4 rather than NVIDIA's NVFP4, which means future fine-tuning or the next generation isn't necessarily tied to NVIDIA chips — other ASICs are on the table.

What I Read in this week ?

Intel

• Intel posted its strongest quarterly growth in 15 years: revenue of $16.1B, a beat (consensus $14.3B). The Q3 guide is even stronger — revenue of $16.3B (vs. consensus $15.1B), gross margin of 42%.

• Foundry is picking up speed: our channel check , 18A yields are running above internal targets; Panther Lake yields are around 80%, in line with expectations; and 18A-P has now entered risk production too. We see 18A-P hitting risk production as a key inflection point — it's a genuinely competitive option for external customers (think the Apple M series).

• 14A timeline is now clear: PDK (that's the Process Design Kit) 0.9 lands around October for pre-volume, with risk production plus external-customer volume in 2H28.

  • The chart below is some of our own guesswork on Intel's future customer lineup.

• They're pushing 2026 capex hard to >$20B, with 2027 higher still. Management expects to turn profitable in 2H27 (on the back of yields plus scale).

  • Intel also announced an extra €5B to expand Xeon 6 on Intel 3 (even as 18A-P is already ramping), which points to capacity being tight and Intel acting as the swing producer. The Bowers campus is also adding Mask Operations.

Google

• This one already has plenty of KOLs covering it, so I won't go too deep into the financials. We read through everyone's takes, and the market seemed caught off guard by the negative FCF — but we'd argue that's actually a lagging indicator.

• Our view, from a long-term standpoint, is that in the AI era software has taken on a manufacturing-like character — it's been "manufacturing-ized." In the old days, adding more usage cost almost nothing; now more usage means more token cost, just like COGS in manufacturing. So as Google shifts from being more of a software/services business toward a manufacturing one, negative FCF is no surprise!

• It's not unusual for a manufacturer to swing to negative FCF when there are good growth opportunities to invest in — TSMC has had negative FCF too. So whether or not you like Google leaning more manufacturing is one thing, but negative FCF while there are investment opportunities is nothing to be alarmed about.

• The single most important metric to watch in GOOG's results is the growth rate of the Google Cloud Platform (GCP) segment. Here's why: GCP's YoY growth is currently the most important pillar supporting GOOG's valuation.

  Google Cloud is now a solid S-curve business — revenue keeps accelerating, up to 82% YoY, which is six straight quarters of accelerating YoY growth. OPM has also kept improving QoQ, edging toward 40%.

• Looked at on a marginal/incremental basis, the YoY/QoQ incremental OPM comes in at 46%/54% — both well above the current ~35% level. That tells us Cloud's OPM still has scale effects left to unlock. If the incremental margin holds around 50%, OPM has a real shot at pushing above 40%, closing in on the 45% of the core Google Services business.

• On the backlog: keep an eye on how a slowing YoY growth rate can compress the valuation. It hit $514B at the end of 2Q26 (vs. an expected $488B, a beat of roughly $26B); QoQ that's $462B → +$52B, +11% (1Q went from $240B → $462B, nearly doubling), with a bit over 50% to be recognized within 24 months.

• On capital intensity: based on the midpoint of $200B in capex guidance, we expect GOOG's capital intensity to keep climbing through 2026–2027, landing around 40% and 45% respectively, though it likely peaks in 2027. Management keeps stressing that 2027 capex will expand meaningfully again over 2026, which hints that the odds of a 2027 FCF recovery are shrinking — pushing the timing of GOOG's FCF recovery closer to 2028.

  

TXN

• This company doesn't get nearly as much buzz in the market, but we think it's the one that most deserves the time to dig into. TXN is about to enter its "harvest" phase for profits. Unlike the other CSPs, over the past two to three years — right through the semiconductor downcycle — TXN went against the grain and poured a ton of capex into building out its 300mm (12-inch) fabs. The cleanrooms and capacity are already fully in place.

• Now that TXN's capex is contracting sharply, free cash flow (FCF) is set for an explosive ramp. In today's environment, where so many tech giants are still spending like crazy on capex, TXN's FCF expansion looks genuinely scarce and pretty compelling.

  we do have cleanroom space available that we can equip and ramp, so we can support a wide range of scenarios from a customer demand standpoint. - Haviv Ilan

• Utilization recovery and margin repair: for an IDM (integrated device manufacturer) of this scale, the things to watch closely are "days of inventory" and the "utilization rate." Over the past few quarters, TXN kept utilization low to manage inventory, which meant heavy depreciation costs badly eating into gross margin. If demand is now broadening as they described on the call (spreading from data center into industrial and automotive), utilization will gradually climb, which brings down per-unit production cost and drives a strong margin recovery over the next few quarters.

• Higher power-IC content value driven by AI data centers: TXN doesn't sell AI compute chips (GPUs/ASICs), but AI servers draw enormous amounts of power and need more — and higher-end — power management chips (power semis). Traditional servers may only need so many analog chips, but the power-architecture upgrade in AI data centers gives TXN's power management and analog products a brand-new tailwind where both volume and price rise together.

  Now in the longer term, and that's going to take some time once we have the SST established, data center input will be at a DC voltage of 800 volts. And then I envision maybe direct conversion from 800 directly to 12 or 6 volts. And -- but that's coming later. And this is where the TAM continues to show strength. The higher voltage, the more opportunity we see.

From the transcript (TSMC 2Q26 earnings call, CFO Wendell Huang): “HPC increased 20% quarter over quarter to account for 66% of our second-quarter revenue.”

Insightology View:

The gross margin story is better than it looks

In this piece we want to spend extra time on gross margin, because that is where the market disagrees the most.

The 2Q margin beat wasn’t just about cost improvements and higher utilization. There was a demand component too: N3 capacity is extremely tight, and some customers paid rush order fees, lifting N3 average selling prices meaningfully from the prior quarter.

Looking ahead, the 3Q guide slips only slightly from last quarter despite 3 to 4 percentage points of dilution from the N2 ramp. That is early validation of our view that this capex upcycle will not send gross margins off a cliff.

Based on our research, TSMC is negotiating a broad foundry price increase with customers for 2027 (we estimate 10% to 15%), which should offset the margin dilution from the N2 ramp and the overseas fabs. Management’s Q&A remarks line up with this: C.C. Wei said he is “really jealous” of memory makers’ gross margins, but TSMC will not suddenly raise prices by 4x or 5x. The goal is to keep margins in a range that sustains long term expansion for both TSMC and its customers.

From the transcript (CFO Wendell Huang): “We have just guided our third quarter gross margin to decrease by 1.7 percentage point to 66% at the midpoint, primarily as we expect the steep ramp-up of our 2-nanometer technology to dilute our gross margin by about 3 to 4 percentage points.”

From the transcript (Chairman & CEO C.C. Wei, Q&A): “So we don’t suddenly increase our price by which I like to have 4x or 5x. […] we earn our value and we make sure that our profit, our gross margin is enough for our long-term sustaining expansion. […] So yes, I’m really jealous about memory company’s 86% gross margin.”

The bigger picture: this capex cycle looks longer than any before it

Management upgraded its language on the next three years of capex, from “significantly higher” than the past three years to “even more significantly higher.” The reasons: customers keep pushing for faster capacity expansion, and equipment makers are raising prices to pass through higher input costs.

From the transcript (CFO Wendell Huang, Q&A): “Last time, we said our CapEx in the next three years will be significantly higher than the CapEx in the past three years. Now the CapEx in the next three years will be even more significantly higher than the past three years.”

Put differently: once A14 (risk production in 2027, volume production in 2028, with transistor density roughly 20% above N2; A13 and the backside power A12 follow in 2029) enters its ramp over 2028 to 2030, capex should at minimum plateau at a high level. This upcycle could run longer than any previous one, putting fab construction and leading edge equipment suppliers on an unprecedented long runway.

Construction is already accelerating. Arizona Fab21 P3, Hsinchu Fab20 P3, and Kaohsiung Fab22 P4 have all been pulled forward to break ground in the second half of 2026. Capacity reshuffling is speeding up as well: existing N5 lines are gradually converting to N3, part of Fab14’s 12 inch mature capacity is being repurposed for silicon interposers to support advanced packaging, and mature nodes are gradually exiting the general purpose logic market, with the unserved demand spilling over to UMC and Vanguard (VIS).

ASML: Beat and Raise, Plus One Number That Disappeared

ASML’s July 15 call was a textbook beat and raise: Q2 revenue of €9.33 billion and gross margin of 54.0%, both ahead of guidance, with net income of €2.92 billion. EUV system sales reached €3.8 billion, including one High NA system (ASML press release). Q3 guidance jumps straight to €11 to €12 billion. Full year 2026 revenue guidance was raised sharply, from €36 to €40 billion previously to €43 to €45 billion, and full year gross margin guidance from 51% to 53% up to 54% to 56%. By segment: EUV sales are expected to grow more than 45% this year (roughly 65 Low NA systems shipped), memory related system sales more than 75%, and advanced logic more than 25% (official transcript).

From the transcript (CFO Roger Dassen): “We now expect to ship around 65 Low NA EUV systems this year, resulting in year-over-year EUV net system sales growth of over 45 percent.”

From the transcript (CEO Christophe Fouquet): “As a result, we anticipate our Memory-related net system sales to grow by over 75 percent this year.”“We now expect advanced Logic foundry-related net system sales to grow over 25 percent this year.”

Insightology View:

One number never showed up on this call: net bookings. ASML has replaced its quarterly order disclosure with visibility language. The market loses a high frequency timing signal, but what it gets instead is actually longer visibility: 2027 is close to fully booked, which locks in growth a year and a half ahead.

Look closely: the order intake figure ASML used to publish every quarter appears nowhere in this quarter’s press release or transcript. The last disclosed number was €13.2 billion in Q4 2025. In its place, qualitative language: Low NA EUV orders for 2027 are “close to fully covered,” a “significant number” of Low NA orders are already in hand for 2028, and the backlog keeps growing. Pair that with the capacity plan: Low NA EUV capacity up around 30% in 2027 from this year’s base of roughly 65 systems, DUV immersion up 30% from roughly 130, and a further 30% increase for each under evaluation for 2028.

From the transcript (CFO Roger Dassen): “For 2027, we are now close to being fully covered with orders for Low NA EUV and we are planning to increase our low NA EUV capacity by around 30 percent.”“Looking ahead to 2028, we have already received a significant number of Low NA EUV orders. Strong demand forecasts from our customers have led us to investigate a further 30 percent capacity increase for that year.”“Similarly, for our immersion systems, we intend to increase capacity by 30 percent in 2027 and are investigating a potential further 30 percent expansion for 2028.”“As a result, our backlog continues to increase with a broad mix of customers.”

One detail on pricing power worth flagging: our read is that ASML failed to push a price increase through to TSMC and only raised prices 5% to 10% on smaller customers. Even the most upstream monopolist in the industry is still a price taker when facing its largest customer.

The WFE upcycle keeps extending

The five major WFE buyers (TSMC, Samsung Electronics, SK Hynix, Micron, and Intel) account for the bulk of equipment purchasing, and all five have now confirmed capex upgrade cycles for 2026 through 2028, opening the door to consecutive upward revisions in 2027 and 2028.

Take TSMC as an example. Based on our channel checks, we expect TSMC’s WFE purchasing budget to climb from roughly 70% to 75% of capex over 2026 to 2028. Add to that TSMC’s own capex guidance moving up again, from the top of the prior $52 to $56 billion range to $60 to $64 billion, and the case that this WFE upcycle runs longer than any previous one keeps getting stronger.

Putting the Two Calls Together: Six Things We’re Watching

1. Late July IDM and memory earnings calls. The next checkpoint. TSMC says AI demand is running stronger than it expected at the start of the year; ASML says memory system sales are growing more than 75%. The Intel and memory maker calls starting in late July will mark both claims to market.

From the transcript (C.C. Wei, asked whether the five year AI CAGR guidance from January had changed): “let me give you not a number, but it’s stronger and stronger and stronger. So we don’t give you the number today because it continues to increase.”

1. TSMC’s 3Q gross margin landing, and the currency. Is the 65% to 67% guide another conservative lowball? And if the Taiwan dollar runs stronger than the 32 assumption, it eats further into margins.

2. Intel 18A’s real yield and customer list. We wait for Intel’s call and customer side evidence. This is one variable in the TSMC pricing power narrative, though we think the actual impact on TSMC’s pricing power is limited.

3. Price hikes landing. TSMC is negotiating a broad 10% to 15% foundry price increase for 2027, while management’s public language stops at regular annual price adjustments. When the negotiation shows up in official guidance will determine whether 2027 gross margin holds its ground.

4. ASML’s new visibility metrics. With bookings gone, we track two things instead: whether the “fully covered” language extends into 2028, and whether the 30% capacity plans get raised again.

5. The advanced packaging bottleneck. Management confirmed the CoPoS pilot line is built and needs about a year to mature before capacity buildout begins. We expect volume production in the second half of 2028 at the earliest, with initial capacity of 5,000 to 10,000 wafers per month, aimed at large form factor packages of 14x reticle size and above; the technical path (glass carrier versus square silicon substrate) is not yet settled. The glass core substrate supply chain is taking shape: Innolux on front end glass metallization, Ibiden and Unimicron on back end ABF buildup, with volume production expected around 2028 to 2029. Packaging remains the metronome for AI chip shipments, and the source of the “front end beats, back end shortages” risk.

Life

• It’s been six months since the last Weekly Shot, and I’m finally back. Work kept me busy this past half year, and a lot of what I was writing and sharing just wasn’t the right fit to put out publicly. But I’ve found a balance that works: the more valuable material goes behind the paid tier, and I’ll get that paid tiering sorted out over the next few days.

• Through all of it I’ve kept doing the same research on tech stocks and the broader economy, and I’ve spent a good amount of time talking with experts inside the industry, getting a closer look at how things really work. I love being in this mode. There’s so much new out there worth exploring and building.

• Going forward, this space will carry updates from the companies I follow closely: their transcripts, the various forums, media coverage, and more. It won’t be limited to tech either. Any topic worth studying over time, I’m more than happy to share it here.

• One thing to bear with me on: a lot of what I have right now can only be estimated from public data. Whatever I’m able to disclose from my expert interviews, I’ll bring out in the paid membership.

Charts Speaks

Every day we look at a huge number of charts. Each week we’ll put a few here to walk through what we’re seeing.

On July 14, 2026, the US reported that the June 2026 CPI year over year rate fell sharply from 4.2% to 3.5%, below the 3.8% the market expected. Core CPI year over year also eased from 2.9% to 2.6%, under the 2.8% consensus. The reason for the pullback is simple: energy. June energy costs rose 15.7% year over year (down from 23.5% in May), and gasoline’s year over year gain dropped from 40.5% in May to 26.7%. The driver was the ceasefire between the US and Iran and the reopening of the Strait of Hormuz, which reversed the oil price shock from earlier in the year.

We think the Fed will hold rates steady at the July 2026 FOMC meeting, buying more time to watch how genuinely inflation is cooling. And if the US and Iran get back to talks and things go well, commodity inflation should keep coming down, which leaves the door open for the Fed to hold rates flat through the end of 2026. On top of that, the five major reforms the Fed has proposed are set to wrap up discussion by year end, and could still lay the groundwork for structural rate cuts over the longer term.

The Nasdaq and the Philadelphia Semiconductor Index have already broken below their quarterly moving averages, which for the tech market is a signal of a medium term consolidation. Look closely at the components and you’ll see Korean memory names deleveraging and getting marked down on valuation. But we’re not going to force the call that it’s time to buy the dip, because we don’t know what other horror stories might come out one after another. We’ll keep respecting the market.

And because the market has started to come down a bit, now is the time for us to spend more effort on the semiconductor space. It means some good names may be getting sold off with the bad.

When the Semis Lose the 60 Day Line

Both the Philadelphia Semiconductor Index and QQQ closed below their 60 day moving averages this week. Before deciding what that means, I went back and tested every single break in the historical record.

First, a definition. Most breaks of the 60 day line mean nothing. In our sample, roughly two thirds get reclaimed within five trading days, the classic whipsaw. So I call a break an **effective breakdown** only when the index closes below the 60 day average and stays below it for more than five trading days. That one filter changes the picture completely.

Here is what effective breakdowns have looked like:

Two things stand out. The typical effective breakdown resolves fast: the low arrives within a week or two, the average gets reclaimed in about three weeks, and a new high follows within two months. The problem is the tail. Once a breakdown stays in force for more than 20 trading days, the median further decline grows to 18% for SOX and 12% for QQQ, and only one in four break day buys in SOX turns profitable within a month. Every catastrophic case (2000, 2002, 2008) started out looking like all the others.

So an effective breakdown is not a sell everything signal. It is a respect the tape signal: median outcomes stay mild, but tail risk roughly triples the moment the line stays lost for a week.

**Data notes.** Daily closes from Yahoo Finance. PHLX Semiconductor Index (^SOX): May 1994 through July 2026, about 8,100 trading days. Invesco QQQ Trust, dividend adjusted: March 1999 through July 2026, about 6,900 trading days. Effective breakdowns cluster in 1995 to 1998, 2000 to 2002, 2008, 2011, 2015, 2018, 2020, 2022 and 2025. All horizons are in trading days. Past performance, not prophecy.

Quotes from company

Beyond the charts, what companies say and how they say it is another place we keep piecing insight together.

This week kicked off Q2 earnings season: the four big banks and AEHR (07/14), JNJ (07/15), TSMC’s call and Netflix (07/16, Netflix after the US close), ASML. The most interesting part is the timing. What the bankers said 48 hours before the correction reads, after the fact, like a preview.

Where the market stands

“It’s getting close to as good as it gets. We just don’t know how long it’s going to last.”
— Jamie Dimon, JPMorgan (JPM) CEO

About as good as it gets, and no one can tell you how much longer it holds.

“The conflict in the Middle East has weighed a bit on global growth whilst giving inflation a second wind... it’s a nuanced story because that growth is not lifting all boats.”
— Jane Fraser, Citi (C) CEO

Growth isn’t reaching everyone evenly. A K shaped split, now officially confirmed.

“...in many ways, we’re still just getting started as a company... We’re under 45% penetrated into addressable households around the world... We’re capturing, we think, just 7% of addressable revenue market.”
— Spence Neumann, Netflix (NFLX) CFO

Read it next to Dimon’s line: one says the economy is about as good as it gets, the other says the company is only getting started. Penetration under 45%, just 7% captured of a $670 billion addressable revenue market, and roughly 5% of global TV viewing share. Where you are in the cycle and how long the structural runway runs are two different things, and both can be true at once.

Demand on the AI supply side is still being revised up

“I believe from this day on all the way to probably 2029, 2030, the demand is very strong. Whether in between there’s a dip or not, I’m not very sure. But the trend is so robust that I believe we are witnessing a kind of a new industry.”
— C.C. Wei, TSMC (2330-TT／TSM-US) Chairman & CEO

Demand runs all the way out to 2029 or 2030, and we’re witnessing a new kind of industry. Whether there’s a dip along the way, he himself says he isn’t sure. That honesty carries more weight than the optimism does.

“So if you’re asking about the AI’s CAGR, let me give you not a number, but it’s stronger and stronger and stronger. So we don’t give you the number today because it continues to increase.”
— C.C. Wei, TSMC Chairman & CEO

Back in January they gave a five year CAGR outlook for AI revenue, in the mid to high 50s. Pressed on it this quarter, he said stronger three times over and refused to update the figure, because it’s still moving up.

“Last time, we said our CapEx in the next three years will be significantly higher than the CapEx in the past three years. Now the CapEx in the next three years will be even more significantly higher than the past three years.”
— Wendell Huang, TSMC CFO

From significantly higher to even more significantly higher. The wording itself is the guidance.

Leverage has been in plain sight all along

“You can see most of this on a daily basis through volumes of the New York Stock Exchange, the CME, volumes through hedge funds. It’s not a secret. Margin loans.”
— Jamie Dimon, JPMorgan (JPM) CEO

Explaining the 86% year over year jump in equities trading revenue, Dimon pointed straight at margin loans. Leverage was never a secret. It’s just that no one wants to look at it while things are going up.

“Do we see some excesses on outside? We always do. A lot of that went to a different market, not in the banking system. Some of that’s come back and now they have to do it more on bank lendable terms...”
— Brian Moynihan, Bank of America (BAC) CEO

The excess leverage moved outside the banking system (private credit), and some of it is now flowing back, forced to be redone on terms banks will accept.

“We didn’t talk about a huge deterioration in credit underwriting standards... a very mild one... I think there’ll be some outliers out there, just like there were, by the way, in the great financial crisis.”
— Jamie Dimon, JPMorgan (JPM) CEO

Credit standards have loosened, but only mildly. When the cycle turns, though, it won’t be a normal distribution. There will be outliers.

The consumer is still holding up

“Consumer spending has recently expanded and continued to outperform our expectations... the spending picked up during the second quarter and now is running at 6%+ year-over-year comparison.”
— Brian Moynihan, Bank of America (BAC) CEO

“When you look at the delinquencies and the net credit losses... both delinquency and credit losses are down year-over-year.”
— Gonzalo Luchetti, Citi (C) CFO

Card spending is accelerating (6%+), while delinquencies and credit losses are falling. No cracks on the consumer side right now.

“...they pay the least per hour of viewing compared to comparable SVOD offerings. In some cases, they would have to pay twice as much per hour for a competitive service.”
— Greg Peters, Netflix (NFLX) Co-CEO

Price increases in the US, Mexico, and Spain in the first half played out consistent with the past, with no unusual churn. Consumers can still take a price hike.

Insightology View

• The biggest scare in the market this week was the valuation reset and deleveraging trade in Korean memory. Memory is about half of the Korean market, so its swings are naturally 2 to 3 times those of other markets. On top of that, retail piled on margin and leveraged ETFs, stacking leverage to an estimated 9 to 10 times. Once forced liquidation kicks in, the selling turns into a stampede that spills over to the rest of the world. But what matters more to us is what the world looks like 18 months out. Without a correction there’s no opportunity.

• For the next 18 months, we think TSMC’s guidance on this call is a useful reference: over the next three years, both CapEx and revenue growth will be significantly higher than the past three years. The company even noted that AI demand will run through 2029 and 2030.

• Across the four quarters from the second half of last year to the first half of this year, the surge in AI demand rippled through the supply chain and set off a bullwhip effect. We’ve also mapped out the price increases across the entire supply chain.

Company Updates

These are some of the companies we stay most constructive on. The table makes its debut in the next issue: a rolling list of roughly 20 to 30 names, most of them US listed, refreshed every week. Down the road we’ll add financial models and valuation updates for these companies.

Disclaimer: This newsletter reflects personal views and information gathering only, and does not constitute investment advice. Companies mentioned are for illustration purposes only. Please do your own research and make your own investment decisions at your own risk.

Life

We took advantage of the MLK long weekend to take a trip down to Louisiana. It was my first time there! I’d heard you absolutely have to try the Gumbo, Jambalaya, Crawfish Étouffée, and Southern Classic Chicken to really get a taste of the Deep South.

While I was there, I couldn't help but wonder what it would be like to actually live and work there. I'm super curious about the local lifestyle, so I'll definitely have to explore more next time I visit.

On a different note, I’ve started shifting my focus this year toward real-world AI applications and getting involved in more hands-on projects. I’m even considering jumping into the industry to learn the ropes. Because of this, my future posts will look a little different: aside from my usual tech industry tracking, I’ll be sharing more insights on both the technical and business sides of things.

If any of you are currently building AI apps, hit me up! I’d love to chat and brainstorm together.

Market

TSMC Revises Up CAPEX

The absolute highlight of this week was TSMC's earnings call. Since the news about TSMC buying land in Arizona broke before the call, the market consensus had already anticipated a CAPEX hike. However, what really surprised everyone was that TSMC revised their capital expenditure even higher.

They also announced that CAPEX over the next three years will be “significant.” This is a straightforward, no-brainer win for fab facility engineering and equipment stocks!

Most of the facility engineering companies are based here in Taiwan, while the equipment supply chain is still dominated by US, European, and Japanese firms, with a smaller portion coming from local Taiwanese suppliers.

Highlights:

• Q1 2026 Guidance: Gross Margin estimated at 63%-65%; Operating Margin at 54%-56%.

• AI Growth: AI CAGR (2024–2029) raised to “mid-to-high 50%” (up from “mid-40s” last quarter).

• 2026 CAPEX: Estimated at USD 52B–56B. About 70%-80% will go toward advanced process technologies and advanced packaging.

• Node Updates: N3 will reach the corporate average gross margin this year, and some capacity is being converted from N5 to N3.

• N2 Ramp: N2 is expected to be a larger node at initial ramp compared to N3.

• Long-term Outlook: CAPEX for the next three years will be significantly higher than USD 100B.

• Legacy: Adjusting 8-inch capacity.

TSMC Advanced Process Roadmap & Fab Construction Update

1. N2 (2nm) Process: Demand Exceeding Expectations

• Customer Status: Apple and AMD are scheduled for initial wafer starts by the end of 2025. Qualcomm, Intel, and MediaTek are expected to join between 2026 and 2027.

• Capacity Planning:

  • Hsinchu Fab 20 P1 / Kaohsiung Fab 22 P1: Mass production officially commenced in late 2025.

  • Hsinchu Fab 20 P2 / Kaohsiung Fab 22 P2: Trial production to begin in Q1 2026.

  • Kaohsiung Fab 22 P3: Tool move-in accelerated to the end of Q1 2026.

2. A16 & A14 (Angstrom Class) Processes

• A16: The first process to introduce Backside Power Delivery Network (BSPDN) technology. Currently, only NVIDIA has been secured as a customer, resulting in a smaller initial production scale. Capacity at Hsinchu Fab 20 P3 and beyond will pivot to the A14 process.

• A14: Risk production is expected in 2027, with official mass production in 2028. Anticipated locations are Taichung Fab 25 and Hsinchu Fab 20.

3. U.S. Arizona Construction Progress

• Fab 21 P1: Mass production officially started in 2025.

• Fab 21 P2A (N3 Process): Completion scheduled for late 2026, with mass production in 2027.

• Fab 21 P2B: Construction began in Q2 2025; mass production slated for 2029 at the latest.

Semiconductor Facility Engineering Market Overview

The market is divided into four core sectors: Cleanroom & M&E Integration, Water Recycling, Gas Supply, and Chemical Supply. Driven by TSMC’s accelerated expansion, revenue recognition for these sectors is expected to peak between 2026 and 2028.

I. Cleanroom & M&E Integration

• Scope: Plant planning, design, construction, and inspection (HVAC, Power, Central Monitoring, etc.).

• Key Players: UIS (United Integrated Services / 漢唐 - Historic major contractor), Yankee Engineering (洋基工程), L&K Engineering (亞翔 - Recent entrant).

• Key Player Updates:

  • L&K Engineering (亞翔): Secured civil and M&E bids for TSMC Fab 22 P7, P8, and P9. Estimated revenue contribution is NT$40–50 billion. Also holds projects for Micron and Vanguard International (VIS) in Singapore, peaking in 2026.

  • UIS (漢唐): The only contractor capable of establishing a US base; secured exclusive bids for Fab 21 (Arizona). Revenue contribution from US construction is 3–5x higher than in Taiwan, though gross margins are lower.

II. Gas Supply Systems

• Classification:

  • Primary Piping (Main Trunk): High entry barrier; dominated by Taiwan Puritic Corp (和淞).

  • Secondary Piping (Hook-up): Handled by Rayzher Industrial Co., Ltd.(銳澤), Wholetech System Hitech Limited (漢科), Genii Ideas Co., Ltd. (聚賢), and Chi Yi Hsin Technology Co., Ltd. (騏億鑫).

• Key Player Updates:

  • Taiwan Puritic Corp(和淞): Holds a near-monopoly on primary piping for advanced processes. Backlog approx. NT$33.1 billion as of Q2 2025. Will benefit from tool installation across multiple N2 and N3 fabs in 2026. Cooperation extends to the A14 process (trial line setup in 2026).

  • Rayzher Industrial Co., Ltd. (銳澤): Focuses on secondary piping. With the main structure of US Fab 21 P2A completed, revenue recognition is expected to peak in H2 2026. Also benefits from tool uninstallation/installation (hook-up demand) during the retrofit of older TSMC fabs (Fab 2/3/5/6) into advanced packaging facilities. Actively promoting its Particle Catcher Device (PCD).

III. Chemical Supply Systems

• Scope: Storage, transport, and distribution of chemicals, emphasizing contamination prevention and material compatibility.

• Key Players: Marketech (帆宣), Nova (朋億), Cica-Huntek (矽科宏晟).

• Key Player Updates:

  • Marketech (帆宣): Order backlog approx. NT$91 billion as of Q4 2025 (NT$73 billion from semiconductor facilities).

  • Revenue Peaks: 2026 (Kaohsiung Fab 22 P3, US Fab 21 P2A, Tainan AP8 P2); 2027 (Kaohsiung Fab 22 P4/P5, Taichung Fabs).

  • Agency Business: Agent for Japanese supplier AIM’s Bonder/De-bonder. Benefits from CoWoS/FOPLP capacity expansion by TSMC, ASE, and Powertech.

IV. Water Recycling Systems

• Classification:

  • Ultra-Pure Water (UPW): Extremely high standards; primarily handled by Organo (Japan)

  • Wastewater Recycling: Primarily handled by Mega Union (兆聯實業) and Trusval (信紘科).

• Key Player Updates:

  • Mega Union (兆聯實業): Order backlog approx. NT$23.5 billion as of Q3 2025. Followed TSMC to the US; Fab 21 P2A revenue recognition will peak in 2026. US projects contribute higher revenue but yield lower gross margins.

TSMC and Samsung Scaling Down Legacy 6-inch and 8-inch Fabs to Concentrate on Advanced Processes

During the recent earnings call, TSMC mentioned plans to gradually phase out certain 6-inch and 8-inch wafer fabs. The company intends to shift its focus toward higher-margin advanced processes (such as 3nm, 5nm, 7nm, and 2nm) required for AI and High-Performance Computing (HPC) applications.

TSMC

• Strategy: Phasing out select 6-inch and 8-inch fabs to pivot resources toward high-margin advanced processes for AI and HPC.

• Product Migration: Products originally manufactured in 8-inch fabs (e.g., CIS [CMOS Image Sensors] and PMIC [Power Management ICs]) have already been migrated to 12-inch mature or advanced process nodes.

Samsung Electronics

• Closure Plan: Plans to shut down the S7 8-inch line at its Giheung campus in the second half of this year. (The campus houses lines S6, S7, and S8; S6 and S8 will continue operations).

• Capacity Impact: following the closure of S7, monthly capacity is expected to drop from roughly 250,000 wafers to below 200,000 wafers (S7 accounts for approx. 50,000 wafers). Existing orders have been transferred to other facilities.

Market Impact: Supply Reduction & Price Hikes

With major players exiting the space, a supply gap is emerging, which is expected to drive prices up.

• Capacity Forecast: Global 8-inch wafer capacity is projected to decrease by approximately 2.4% in 2026.

• Price Fluctuation: The tightening supply is expected to trigger a price hike for 8-inch foundry services, estimated between 5% and 20%.

Beneficiaries & Competitive Landscape

Tier-2 foundries are welcoming a wave of order transfers. As Samsung and TSMC release market share, orders are flowing to foundries that specialize in mature processes.

• Top Winner: DB HiTek (South Korea), Vanguard International Semiconductor (VIS).

  • Current Status: Running at full capacity with a significant order backlog; currently unable to accept new orders.

  • Competitive Advantage: Specializes in analog processes (such as BCD - Bipolar-CMOS-DMOS) and excels in “High-Mix Low-Volume” (HMLV) production (e.g., PMICs and DDIs [Display Driver ICs]).

  • Opportunity: As TSMC and Samsung continue to reduce capacity, customer orders are highly likely to shift to DB HiTek and Vanguard International Semiconductor (VIS).

What I Read

HBM & DRAM Manufacturing Economics: Process Complexity, Yield Dynamics, and Supply Constraints

TrendForce: Micron收購PSMC銅鑼廠，2027年全球DRAM供給可望上修

https://www.trendforce.com.tw/presscenter/news/20260119-12882.html?fbclid=IwY2xjawPbcYlleHRuA2FlbQIxMQBzcnRjBmFwcF9pZBAyMjIwMzkxNzg4MjAwODkyAAEeDTgpDe2h6q82uFFeCyfE2htJQbpNK7kaYtEKtMGWYuQ9UdySujNKMrVc71A_aem_pUwLpLO2pfGC7-KJIfbpjA

SiFive To Adopt NVLInk Fusion For Future Data Center RISC-V CPU Designs

https://www.servethehome.com/sifive-to-adopt-nvlink-fusion-for-future-data-center-risc-v-cpu-designs/?fbclid=IwY2xjawPbe7pleHRuA2FlbQIxMABicmlkETFNUDRlbXBqYkI3ak1RTEYwc3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHto6b0rkm0N5flHD8r9CdiklosN4uzo4tmF4cRwDw3cjXqb0E1-W6CAYDs8k_aem_7h6uBUWq4WuhPCLvqZx8jQ

Cerebras Inks Transformative $10 Billion Inference Deal With OpenAI

https://www.nextplatform.com/2026/01/15/cerebras-inks-transformative-10-billion-inference-deal-with-openai/?fbclid=IwY2xjawPbfDJleHRuA2FlbQIxMABicmlkETFNUDRlbXBqYkI3ak1RTEYwc3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHr8ZbLQo3S0eOXYyG762KtmtI1YA5Qt4ZV1Voy_RSxiWgIaq8kELKB0E_GFD_aem_4a2aWudTKAUmJbnyhcD0gg

Cerebras CS-3: the world’s fastest and most scalable AI accelerator - Cerebras

https://www.cerebras.ai/blog/cerebras-cs3?fbclid=IwY2xjawPbfDBleHRuA2FlbQIxMABicmlkETFNUDRlbXBqYkI3ak1RTEYwc3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHrtRb8_42C9CgBIo02mrtCbbW9fyCc0JVON53N-RYx3qHPwBdMRzQLiErbqe_aem_0SUywXqZB_YXVijmAgWyBQ

Here are the 55 US AI startups that raised $100M or more in 2025

https://techcrunch.com/2026/01/19/here-are-the-49-us-ai-startups-that-have-raised-100m-or-more-in-2025/

Samsung to close down one of its 8-inch foundry fabs within the year

https://www.thelec.net/news/articleView.html?idxno=5549&fbclid=IwY2xjawPbgBFleHRuA2FlbQIxMABicmlkETFhdUJYVXRybDNQczZuQk1Vc3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHlxfbh35FKdf5LZcOTGNow7atEaY8Yv75icteyn-xB-EMgqxBPc-HLdCRCJN_aem__bpZ6bsUFkAtVbCgrnXX9g

翁家翌：OpenAI，GPT，强化学习，Infra，后训练，天授，tuixue，开源，CMU，清华｜WhynotTV Podcast #4

The non-technical PM’s guide to building with Cursor | Zevi Arnovitz (Meta)

Life

Last week, my wife and I took a trip to Miami. This is arguably the most comfortable season to visit—perfect for escaping the winter chill and soaking up the sun on South Beach.

You can really feel that the city is booming, particularly in terms of finance and infrastructure. It has effectively become the “Wall Street of the South,” a movement largely spearheaded by Ken Griffin moving Citadel’s headquarters here. Other major players like Thoma Bravo, Point72 Asset Management, and Elliott Management have all established a presence following suit.

The Latin American influence is massive here; it feels like the unofficial capital of Latin America. You can honestly get by without speaking English, as Spanish seems to be the primary language on the streets.

However, my top recommendation has to be Sanguich. They cure their own ham and make a special Mojo sauce. Every bite is an explosion of savory juices, yet it’s never cloyingly greasy. I highly recommend it!

Market

AEHR:

Following AEHR's financial results and earnings call this week, we view this as a company that warrants close attention.

The Case for AEHR: The primary reason to watch AEHR is the surging market demand for Burn-in equipment. This demand is driven by three core factors: chips have become too expensive, too complex, and failure is absolutely not an option (mission-critical). In the past, burn-in might have been a sampling process for certain products. However, for today’s high-end applications, it has evolved into a mandatory “100% inspection” step.

AEHR is a Leader in Burn-in: There are surprisingly few participants in this space because the required know-how is incredibly specialized. While traditional Japanese manufacturers (such as Espec and Shinwa) provide environmental stress test equipment, they tend to focus on general ICs or automotive parts. They often cannot match AEHR’s capabilities in large-scale Wafer-Level Burn-in.

Concentration Profile: AEHR is a “pure-play” company focused almost exclusively on the niche burn-in market. It does not have diversified business lines to spread risk.

Two Main Product Lines:

• Wafer-Level Burn-In (WLBI): Centered on the FOX Series (FOX-XP / FOX-NP / FOX-CP), combined with their proprietary WaferPak (full-wafer contactors) and the WaferPak Aligner for automation. These are used for burn-in, stabilization, and testing at the wafer or singulated die/module stage.

• Packaged-Part Burn-In (PPBI): Expanded through the acquisition of Incal, this line includes Sonoma (ultra-high power), Tahoe (medium power), and Echo (low power, high parallelism) systems. These target post-packaging reliability demands in AI/HPC, data centers, and mixed-signal markets.

  

  Financials & Earnings Call Highlights

  Financial Data:

  • Sluggish Revenue and Earnings: Q2 revenue was $9.9 million, a 27% decrease year-over-year. The primary cause was a decline in WaferPak (consumables) shipments. Gross margin also dropped from 45.3% in the same period last year to 29.8%, driven by an unfavorable product mix (a lower percentage of high-margin consumables).

  • Bookings are the Core Highlight: Despite soft revenue this quarter, the company secured $6.5 million in new orders within the first six weeks of Q3, bringing the effective backlog to $18.3 million. The CEO emphasized that bookings for the second half of Fiscal Year 2026 are expected to reach $60 million to $80 million. The majority of this is for AI processor burn-in at both the wafer and package levels. This figure far exceeds revenue projections and serves as a key leading indicator for future growth.

  “As a result, we’re reinstating guidance for the second half of fiscal ‘26. For the second half of fiscal ‘26, which began November 29, ‘25, and ends this May 29 of ‘26, Aehr expects revenue between $25 million and $30 million dollars. As stated earlier, although we’re not providing formal bookings guidance based on customer forecasts recently provided to Aehr, we believe our bookings in the second half of this fiscal year will be much higher than revenue, between $60 million and $80 million dollars in bookings, which would set the stage for a very strong fiscal ‘27 that begins on May 30, 2026.” -Gayn Erickson (Chief Executive Officer, Director)

  • New Customer Activity: Since the last earnings call, two additional AI processor companies have initiated plans for wafer-level benchmark evaluations.

  • Flash Memory Milestone: Just before the Christmas holiday, the company completed a wafer-level benchmark test with a leading global NAND flash memory manufacturer, preparing for future HBF (High Bandwidth Flash) production.

  ---

  Conference Call Key Takeaways

  1. AI Momentum & Market Potential Institutional investors are highly focused on the composition and validity of the $60M–$80M booking guidance for H2 FY2026.

  • Order Composition: Management confirmed that the vast majority of these substantial orders come from AI processors (covering both wafer-level and package-level testing), while the share of Silicon Carbide (SiC) is extremely low.

  • Single Client Potential: When asked if the AI market is truly larger than SiC, the company replied that a single large AI client’s system requirements could reach 20–30 units (at $4M–$5M per unit). The AI business scale could reach hundreds of millions of dollars in the coming years.

  • 2027 Outlook: These orders are expected to come in at the end of FY2026 and convert into revenue starting in Q1 FY2027 (June 2026), underpinning expectations for strong growth.

  2. Technology Path: Wafer-level vs. Package-level Competition & Synergy

  • Benchmark Delays: The company addressed why AI wafer-level testing progress has been slower than expected. Clients, accustomed to package-level testing, need to adjust their DFT (Design for Test) rules and test vectors when shifting to wafer-level. This “learning curve” caused delays of several weeks to months.

  • Coexistence: Management emphasized that package-level systems (Sonoma) and wafer-level systems (Fox) will coexist long-term. Package-level is suitable for initial qualification (HTOL), while wafer-level testing offers massive yield savings in the production phase by preventing the scrapping of expensive HBM (High Bandwidth Memory) and substrates.

  • Sonoma’s Strategic Position: The Sonoma system is currently the “preferred standard” for AI clients. Its 2,000-Watt ultra-high power testing capability is unique in the market and allows for a seamless transition to automated production.

  3. Specific Market Progress: Flash & Photonics

  • High Bandwidth Flash (HBF): Regarding NAND Flash benchmarks, clients are shifting from traditional enterprise SSDs to HBF optimized for AI workloads. This requires higher power consumption, playing directly into Aehr’s technical strengths.

  • Silicon Photonics: Although volume production schedules for leading clients have been pushed to FY2027, this aligns with the launch cadence of new AI platforms.

  4. Capacity & Operational Efficiency Given the revenue dip, attention turned to the company’s ability to handle future explosive order growth.

  • Doubling Capacity: Through facility renovations, capacity has increased 5x, with the current ability to deliver over 20 systems per month.

  • Consumables & Margins: The drop in gross margin was explained by a temporary decrease in the shipment ratio of high-margin WaferPaks. As the installed base of systems rises, consumable revenue is expected to rebound.

  5. Competitive Landscape Currently, AEHR is the only player in wafer-level burn-in testing. Most market data is based on package-level testing. As clients begin migrating toward wafer-level burn-in, AEHR occupies a critical position.

  “So as we get our arms around the market, the market data that would be out there would be packaged-part because no one’s doing wafer level except for us, and so we’re creating our own models related to, okay, for that unit capacity, if you went to wafer-level burn-in, what would that look like? Kind of similar to what we had to go through in the original silicon carbide side of things of the whole market, and we’re not sitting here, everybody included Nvidia and Google and Microsoft and Tesla and these guys all went with us.”-Gayn Erickson (Chief Executive Officer, Director)

  ---

  Insightology View

  We believe the wafer-level Burn-in market represents a new wave of demand in semiconductor testing. As power consumption continues to rise, AEHR is undeniably situated in a critical position to capitalize on this shift. Furthermore, we believe the acceleration in demand for the broader testing industry will sustain.

• Watch List:

  • AEHR (USA)

  • TER (Teradyne - USA)

  • Chroma (TW)

  • Hon. Precision (TW)

  • Advantest (JP)

  • ASE (TW)

NVIDIA Enters Storage Servers: The Inference Context Memory Storage Platform (ICMS)

NVIDIA CEO Jensen Huang unveiled the “NVIDIA Inference Context Memory Storage Platform (ICMS)” during his keynote, marking NVIDIA’s official entry into the storage server domain.

The core purpose of ICMS is to alleviate the burden of the “KV Cache” (Key-Value Cache) generated by Large Language Models (LLMs) during AI inference.

• The Problem: KV Cache acts as the AI model’s “short-term memory” and excessively consumes precious GPU High Bandwidth Memory (HBM) capacity.

• The Solution: The new architecture offloads this data to storage servers and retrieves it only when needed, thereby freeing up GPU HBM space. This is expected to resolve the recent “degradation” issues (where ChatGPT seems to get “dumber”) caused by user surges at OpenAI.

The Shift to SSDs: In the AI training era, market attention was solely on HBM crowding out DRAM. With NVIDIA’s entry into storage servers, it declares that in the AI inference era, SSDs will play a far more critical role than before. Previously, inference relied on DRAM/HBM. The future model will be “Inference using DRAM/HBM + NVMe SSD simultaneously,” and behind those SSDs lies NAND technology.

Rack Architecture & The Changing Role of the DPU

Rack Configuration: The top two trays of each NVIDIA ICMS rack will house 8 Bluefield-4 DPU boards, while all trays below will be populated with SSDs.

DPU: From Supporting Actor to Lead: Unlike the previous NVIDIA GB300 NVL 72 racks where the Bluefield-3 DPU was merely optional, in the ICMS rack, the Bluefield-4 DPU plays the critical role of managing all SSDs. This signifies a shift where the DPU moves from a peripheral component to a central, leading role.

• Note: The Bluefield-4 DPU board integrates the Grace CPU and DPU into a single package.

Supply Chain Details:

• LPDDR5X: Supplied by Micron.

• SSD Controller: Supplied by SanDisk.

• BMC: Each DPU board is equipped with an ASPEED AST2600 BMC.

Insightology View

1. NAND/SSD is the True Structural Beneficiary This is the most significant impact. We are seeing several structural changes:

• KV Cache is moving from “cold data” to “quasi-hot data.”

• Latency is no longer just a storage issue; it is now an inference KPI.

• Capacity is no longer just a cost center; it is a throughput multiplier.

2. Changing Metrics What is the result? SSD metrics are shifting from “TB per Dollar” to “KV per Rack / Latency Bound.” Demand will pivot from “capacity-oriented” to “Performance × Latency × Network Affinity” oriented. This creates a highly favorable environment for high-capacity, enterprise-grade, and high-endurance SSDs.

3. Industrial Tech Meets AI We believe industrial-grade SSD technology is well-suited for the high-value AI Server/Data Center market. However, many of these are not “new architectures” or “new technologies” born purely for AI; they are existing technologies in the industrial sector (such as pSLC—simulating SLC using MLC or TLC NAND) that have been mature and widely used for years.

4. NVIDIA’s Strategic Triad From a hardware perspective, this is a storage architecture adjustment. But from NVIDIA’s perspective, three things are happening simultaneously:

• Inference efficiency improvements increase GPU ROI.

• BlueField-4 becomes a mandatory component for AI factories, driving DPU volume.

• Storage architecture becomes locked in by NVIDIA’s software stack (via NVIDIA DOCA™, NVIDIA NIXL, NVIDIA Dynamo).

Essentially, this will compel more non-hyperscaler customers to rely on NVIDIA’s turnkey solutions rather than cobbling together their own storage, network, and software stacks.

Hi everyone, Merry Christmas! Over the past two weeks, because I was traveling with my family for the holiday, I have combined two weeks' worth of content into a single issue to fully dedicate my time to them. Christmas is my favorite holiday of the year; beyond the beautiful decorations, gift exchanges, and games, what I love most is the atmosphere of everyone coming together.

The tallest Christmas tree I've seen this year.

Market

MU

I. Micron (MU) Business Segment Outlook

1. Data Centers: Management has revised the 2025 global server shipment growth forecast upward from 10% to 17–19% YoY. This is driven by cloud service providers continuing to build data centers, which sustains strong demand for HBM, Server LPDDR5, and Gen 9 process NAND Flash. For 2026, agreements on HBM supply and pricing have already been reached with customers, and HBM4 has passed validation, with production ramp-up scheduled for Q2 2026.

2. PC/Notebooks: Due to memory supply shortages, PC and laptop brands are facing production bottlenecks. Management predicts that 2026 shipments may experience a slight decline, which would offset the benefits of increased average memory capacity per unit.

3. Mobile & Intelligent Edge: Small-scale shipments of 1-gamma process 16Gb LPDDR6 have commenced. Additionally, samples of 1-gamma process 24Gb LPDDR5 are being provided to customers to continue improving the product mix.

4. Auto & Industrial: Because these products have longer lifecycles, demand for DDR4 and LPDDR4 remains robust. To support the long-term needs of these customers, the company announced an expansion plan for its Virginia plant in June 2025.

II. HBM Market and Supply Chain Dynamics

1. Market Growth: The HBM market size is projected to reach approximately $35 billion in 2025 and grow to $100 billion by 2028, representing a compound annual growth rate (CAGR) of 40%.

2. Kinik and Technical Requirements: It is estimated that Micron uses Kinik’s carrier wafers as temporary carriers for the HBM stacking process. Kinik plans to expand capacity by 70,000 pieces per month in Q3 2026 to meet the needs of logic chips (advanced processes) and memory customers.

3. CMP and Hybrid Bonding: As HBM stacking layers increase, every layer requires CMP (Chemical Mechanical Polishing). The future introduction of Hybrid Bonding will significantly raise requirements for surface flatness, driving demand for higher-specification diamond discs and boosting Kinik’s operational performance.

III. DRAM & NAND Supply and Demand Forecast

1. Supply and Demand Balance: For both DRAM and NAND, supply growth is projected at approximately 17–20% (primarily driven by the 1-gamma process), while demand growth is expected to be around 20%. Consequently, the expansion of the supply gap is expected to be limited.

2. Short-term Risks: Cloud service providers currently hold high Server DRAM inventory (70–80 days), which is nearing the “safety level” of 90–120 days; this may lead to a slowdown in procurement. Furthermore, rising DRAM prices could potentially suppress demand for consumer electronics like PCs and smartphones.

3. Long-term Equilibrium (2027+): The market is expected to return to balance starting in 2027 as new production capacities from major manufacturers—including Samsung (P4), SK Hynix (Yongin), Micron (ID1), Nanya Technology, and Winbond—are completed and come online.

4. NAND and HDD Competition: As HDD manufacturers (Seagate, WD) expand capacity for HAMR and Ultra SMR technologies (ranging from 36TB to 44TB), the trend of using QLC NAND to replace HDDs for “cold storage” will slow down, easing NAND supply pressure.

5. Potential Supply Shock: A major variable is Samsung’s Line 12 in Hwaseong. If Samsung converts its 100,000 monthly wafers of MLC NAND capacity entirely to TLC/QLC, it could add an extra 10–15% to global supply, potentially disrupting the market balance

Nvidia’s Potential Acquisition of Groq: A Strategic Move for Talent and Software Supremacy

While social media hype often paints Groq as a revolutionary “Nvidia killer” due to its unique hardware architecture, a deeper look at the technical and business realities suggests that Nvidia’s interest in the startup is likely driven by compiler technology and top-tier talent rather than just its hardware.

1. The SRAM Debate: Innovation or Limitation?

Groq’s primary claim to fame is its use of SRAM (Static Random-Access Memory) instead of the industry-standard HBM (High Bandwidth Memory). However, this approach comes with significant trade-offs:

• Cost and Scalability: SRAM is integrated directly onto the same die as the logic chip (the GPU, TPU, or LPU) and is traditionally used for cache memory (L1/L2/L3). Because Groq does not use the most advanced manufacturing processes, and because SRAM area does not shrink effectively as logic chips advance, the cost structure is unfavorable, making it extremely difficult to expand capacity.

• Not a “Magic Bullet”: Several other startups, such as Graphcore and Cerebras, have pursued similar SRAM-heavy paths with poor market results. Graphcore, for instance, was eventually sold to Arm at a significantly depressed valuation.

• The HBM Success Story: Proponents of Groq often also praise Google’s TPU, yet the TPU utilizes large-capacity HBM—the same technology used by Nvidia and AMD—proving that removing HBM is not the only way to achieve high performance.

2. Business Realities vs. Social Media Hype

Despite the buzz on platforms like X (formerly Twitter), Groq’s commercial performance has been relatively stagnant:

• Stagnant Adoption: The LPU (Language Processing Unit) has been on the market for two to three years, yet it still lacks adoption by major mainstream customers.

• Market Reception: Actual revenue and business traction have remained cold compared to the enthusiastic reception in tech circles.

3. Why Nvidia Wants Groq: Compilers and Determinism

The true value of Groq likely lies in its software layer.

• Deterministic Execution: A historical weakness of Nvidia’s CUDA platform is its dynamic scheduling, which makes it difficult to achieve “Zero Jitter”. In contrast, Groq’s compiler achieves “Deterministic Execution,” providing predictable, ultra-low latency.

• Batch-1 Efficiency: While Nvidia dominates heavy model training, Groq’s architecture is exceptionally efficient for Batch-1 scenarios—where a single user requires an immediate, real-time response.

• The Hybrid Strategy: Nvidia could potentially develop “hybrid architecture” products, using HBM for massive training tasks while integrating Groq’s compiler and SRAM tech for edge computing or specialized high-speed inference.

4. The “Acqui-hire”: Securing the Architects of the AI Era

Perhaps the most compelling reason for an acquisition is talent. Groq was founded by Jonathan Ross, a core designer of the original Google TPU.

• TPU Legacy: Ross was instrumental in creating the hardware that powered AlphaGo’s historic victory in 2016.

• Software-Defined Hardware: Ross founded Groq to evolve the TPU concept from a rigid, internal-use tool into a more flexible, software-defined system. By moving control from the hardware circuits to the compiler (software), his LPU aims to solve the efficiency gaps found in traditional GPUs.

5. Strategic Defense

Finally, this would be a defensive acquisition. Tech giants like Google, Amazon, and Microsoft are all developing in-house chips to reduce their dependence on Nvidia. By acquiring Groq, Nvidia prevents its rivals from leveraging a unique architecture that has already proven it can outperform Nvidia in specific, niche low-latency applications. At a potential price tag of $20 billion, this is a relatively “inexpensive bet” for Nvidia, representing only a few quarters of their typical R&D budget.

What I Read

• DeepSeek’s Impact on AI (Democratizing AI Compute, Part 1)

• Groq AI Beats NVIDIA -- Cheap Inference Processors

• Laid Off After 25 Years in Tech: The Anxiety, Sacrifice, and Reality No One Talks About

• AI capabilities progress has sped up

• How Is Groq So Fast? An Overview of Groq’s TSP Architecture

• Think Fast: A Tensor Streaming Processor (TSP) for Accelerating Deep Learning Workloads https://ieeexplore.ieee.org/document/9138986

• A Software-Defined Scale-Out Computer Architecture https://ieeexplore.ieee.org/document/9808428

• Steam, Steel, and Infinite Minds

• MU Transcript

Hi everyone, how was your week?

This week was also a busy one. I attended several analyst conferences, busy compiling data, and during the process, I also reflected on some of my mistakes this year and areas for correction, and have been continuously organizing our outlook for 2026.

I am also very grateful to Andrew Lu for sharing our earlier articles. This week, the number of followers also exceeded 100 people. This column will continue to record observations and reflections on technology, life, and investment. Seeing that people are following the articles I write for the first time gives me a feeling of affirmation.

Life: Long-Termism

• Recently, I watched two interview videos with Chinese investor Duan Yongping. He rarely gave public interviews over the past few years, but in the last few months, he has appeared twice. The biggest takeaway was the reminder to return to “Long-Termism” and the “Normal Mindset” (Ping Chang Xin).

• For US investors, this investor—Duan Yongping—might be unfamiliar. He was the first Chinese person to win the bid for lunch with Warren Buffett and is the founder of China’s Bubugao and Xiaobawang (the parent company of the mobile phone brand Vivo). After 2000, Duan Yongping chose to retire, stepping away from day-to-day management (which has been over 20 years now), living a “normal” life primarily focused on playing golf and exercising. Although he let go of the businesses, the corporate culture he established (rationality, integrity/doing the right thing, consumer orientation) continues.

• In the US, one of his funds, H&H International Investment, currently manages US stock holdings worth approximately $14.7 billion.

• He mentioned that the core essence is the “Normal Mindset” (Ping Chang Xin) and “integrity/doing the right thing” (Ben Fen). The “Normal Mindset” is truly not easy because the premise of having a Normal Mindset is “Long-Termism,” thinking rationally about the long run. It is difficult for truly ordinary people to maintain a Normal Mindset because people are easily swayed by various external temptations and short-term returns.

• Although in the interview he admitted to being a “lazy person,” always having “no great ambitions,” and just wanting to live each day well, he emphasized that finding what you love to do is extremely important. If you don’t like what you are doing, you must try hard to like it to find enjoyment and do the job well, which means doing the right things (Dao) plus doing things right (Shu). I believe this is his Normal Mindset.

• For me, writing, investing, business, and building products are all things I enjoy. Even if there were no AI today, I would still continue to do these things. However, it is inevitable that due to drastic short-term changes, new “stock market prodigies” emerging every year (young or middle-aged), new products, or news about who raised how much money, one might forget the Normal Mindset.

• “The Great Dao is simple” (Dà Dào Zhì Jiǎn) is also his core concept, which is quite similar to Charlie Munger’s core concepts.

• In business, “subtraction” means focusing on core businesses and eliminating unhealthy, unsustainable temptations, concentrating on core value (building a healthier, longer-lasting enterprise or life).

• When making decisions, prioritizing whether something is the right thing to do comes before “whether there is money to be made”.

• In investing, if you don’t understand it, resolutely don’t touch it. His margin of safety in investing is not about a low price, but rather how deeply one understands the company. If you cannot deeply understand the operational reality of an enterprise, you must restrain from investing.

• Focus energy on the few things that can be deeply understood and done well, and firmly eliminate all tempting or incorrect options that do not conform to the principle of “integrity/doing the right thing” (Ben Fen). As he said: “The main reason we are who we are is because of the things we don’t do.”

• Insightology also hopes to continue building a deeper understanding in technology and business and become a healthier, longer-lasting enterprise, not just through articles, but by implementing it in our lives.

AI and Horses, Bubble?

• Technological progress is linear (or exponential but steady), but the moment “replacement” occurs is sudden.

• This week, I read an article by Andy Jones, which mentioned analogies for the replacement of tools through the emergence of the steam engine, computer chess, and artificial intelligence.

• The engine was invented in 1700. Over the next 200 years, engine technology steadily improved, increasing performance by 20% every decade. During the first 120 years of steady improvement, horses hardly noticed any change. Subsequently, between 1930 and 1950, 90% of horses in the US disappeared. The development of the engine proceeded steadily, but the moment it rivaled the performance of horses happened suddenly.

• Now let’s look at recent chess. People started paying attention to computer chess in 1985. Over the next 40 years, the rating of computer chess improved by 50 points annually. This meant that in 2000, human International Chess Grandmasters had a 90% chance of winning against a computer. But ten years later, when the same human Grandmaster played against a computer, the probability of losing was as high as 90%. The development of chess was steady, but reaching parity with human capability was sudden.

• Capital expenditure in the field of artificial intelligence has been quite stable. Currently, global annual spending on AI data centers is equivalent to 2% of US GDP. This figure seems to have been steadily doubling over the past few years. Based on signed agreements, this number is highly likely to continue doubling in the coming years.

• The author was one of the earliest researchers hired by Anthropic. A significant part of the author’s work in 2024 was answering technical questions from new hires. At that time, the author and other senior employees answered about 4,000 questions from new hires every month. By December, Claude’s level was finally high enough to answer some of our questions. In December, such questions indeed existed. Six months later, 80% of the questions the author was asked had disappeared.

• The summary of the article is that the author believes that the progress of technology is accelerating, and the act of “replacement” is becoming increasingly sudden, whether for horses, International Chess Grandmasters, or AI researchers. It took decades for horses to be surpassed, years for Chess Grandmasters, but researchers were surpassed in just six months.

Insightology View

• AI drives the cost of intelligence towards zero, which is not good for employees, but for entrepreneurs, it may provide significant profit margins. However, for already established companies, we believe they must embrace AI, and the sooner the better, just as the author’s article suggests: technological progress seems steady, but replacement is sudden.

• But for investors, there is no need to debate whether the technology is strong enough; instead, they should focus more on “whether the price is reasonable” and the “cyclicality of human behavior,” because history does not repeat itself, but it rhymes.

• After reading Andy Jones’ article, the chart is terrifying, showing a future that is ‘linear and inevitable,’ but it ignores the power of capital cycles. Looking at Howard Marks’ article, I personally believe Andy Jones’ piece is full of the fear of “missing out on the future”. However, this collective FOMO (Fear of Missing Out) is precisely the psychological basis for bubble formation. When everyone believes that “humanity is doomed” and sells off human assets (such as traditional companies, skills education) to chase AI, human assets might paradoxically become valuable due to being excessively undervalued.

• Certain bubbles (like AI) will leave behind infrastructure that changes the world, but even if the outcome is that AI changes the world, the process will be accompanied by massive capital destruction. After the 2000 dot-com bubble burst, Amazon fell 90% before re-emerging.

• Railways changed the world, but most early railway investors went bankrupt; the internet changed the world, but those who bought in 1999 lost their life savings. Although we might truly be the ‘horses,’ if the market has already driven the price of ‘horse meat’ down to zero and inflated the price of the ‘engine’ to sky-high levels, then as an investor, I would rather buy undervalued human assets during panic than chase overheated AI during euphoria. Trees don’t grow to the sky; cycles always return.

• The “physical reality” of technology and the “financial reality” of the market are two different things. Even if humanity is destined to be replaced, this process will definitely not be a straight line, and it must involve massive bubble bursts and value re-assessment.

• https://andyljones.com/posts/horses.html?fbclid=IwdGRleAOlVC9leHRuA2FlbQIxMQBzcnRjBmFwcF9pZAo2NjI4NTY4Mzc5AAEe4tTCZ4ERJ07qq9d-6gcJJsQf9jumNbbrtiRuD_GXgw64sY2mkT4A6aSn1xM_aem_dLA_fT6Q0111Is0UwZ_aXQ

AVGO Conference Call

• We believe the market over-interprets short-term stock price fluctuations. Setting aside the stock price, from an industry perspective, Broadcom is transforming from a pure chip supplier to directly providing complete cabinet-level solutions. This solution, besides the XPU, also includes back-end components like switches and optical communication, and may be priced as an entire cabinet, thus affecting the gross margin.

• Some market commentators discussed that the gross margin figures were lower than expected. We believe this is an unreasonable assertion because anyone with a normal mind knows that switching from selling chips to selling cabinet solutions will inevitably lower the gross margin, but the company can earn more money. We do not view this as negative news.

• This conference call re-confirmed that the biggest current bottleneck for AI chips is packaging. This verifies previous supply chain observations, where many orders flow to external OSAT factories (back-end assembly and testing plants) when TSMC cannot fully meet demand. The following two paragraphs are some conclusions from our research this week.

• “ That’s an interesting challenge that we have been addressing constantly and continue to, and with the strength of the demand and the need for more innovative packaging, advanced packaging, because you’re talking about multi-chips, multi-chips in creating every custom accelerator now, the packaging becomes a very interesting and technical challenge. Building our Singapore fab is to really talk about partially insourcing those advanced packaging. We believe that we have enough demand. We can literally insource not from the viewpoint of not just cost, but in the viewpoint of supply chain security and delivery. We’re building up a fairly substantial facility for packaging, advanced packaging in Singapore, as you indicated, purely for that purpose to address the advanced packaging side. Silicon-wise, no, we go back to the same process source in TSMC, and so we keep going for more and more capacity in 2 nm, 3 nm.”

TSMC CoWoS capacity

• This week, we have some research and updates regarding TSMC’s CoWoS capacity. TSMC’s Tainan Science Park (Nanke) AP8 plant and Chiayi AP7 plant are scheduled to begin equipment installation in the fourth quarter of 2025.

• Estimated CoWoS advanced packaging capacity is expected to be approximately 75,000–80,000 wafers/month by the end of 2025. The capacity target for the end of 2026 has been revised upwards from the original 110,000–120,000 wafers/month to 120,000–130,000 wafers/month, and 140,000–150,000 wafers/month by the end of 2027.

• The CoWoS capacity forecast was recently raised in early December. TSMC continued to place orders with CoWoS equipment suppliers from the end of October to the end of December. Due to the long delivery lead time of Japanese equipment (7 to 8 months), plus 1-2 months for order adjustments, this batch of orders is expected to commence production in the fourth quarter of next year. This wave of CoWoS capacity increase is expected to be the last significant growth in 2026.

• Despite the capacity revision, it still cannot fully meet customer demand. To address this issue, some products with simpler RDL structure designs will gradually be outsourced to professional outsourced testing and assembly (OSAT) firms. Furthermore, TSMC is internally studying the transfer and outsourcing of the Flip Chip Bump process to free up factory space for further expansion of Cu Pillar Bump process stations. Future CoWoS advanced packaging capacity might potentially increase by another 10,000–20,000 wafers/month.

• Main Corresponding Downstream Customers

• The capacity increase is mostly allocated to NVIDIA. The expected production capacity for Rubin plus Blackwell in 2026 has been revised upward to 650K (originally 580K). Broadcom’s monthly capacity continues to be revised upward, currently viewed at 240K (about 180K allocated to Google), with subsequent revisions of 10K every 1-2 months. MediaTek (MTK) capacity for 2026 is expected to be revised upward to 100K. AWS is requesting CoWoS-R capacity from ASE.

Second Supply Chain

• Clients are actively establishing a second supply chain besides TSMC, prompting professional outsourced testing and assembly (OSAT) firms to significantly increase advanced packaging capacity and begin entering the front-end CoW process.

• ASE Technology Holding (including SPIL): It is expected to undertake orders for TSMC-like production lines and utilize the clients’ self-built second production line. It is anticipated that by the end of 2026, advanced packaging capacity will expand from 7,000–8,000 wafers/month to 20,000–25,000 wafers/month.

ASE Technology Holding Total CoWoS Advanced Packaging Capacity

• 2024: Approximately 4 kwpm (primarily CoWoS-S)

• 2025F: Approximately 7 kwpm (Slight increase in CoWoS-S, small addition of CoWoS-R)

• 2026F: Approximately 25 kwpm (Significant growth, CoWoS-L ratio increases notably, followed by CoWoS-S and CoWoS-R)

• Google: The next-generation TPU v8p (Zebrafish) will also continue to adopt CoWoS-S advanced packaging, expected to be launched in the second half of 2026. Google will jointly design the silicon interposer with Powerchip Semiconductor Manufacturing Corp. (PSMC) as the second supplier, and since the patent of AP Memory (AIP) is used, licensing fees must be paid to AIP. Subsequently, ASE and SPIL will be responsible for the front-end CoW process.

• AMD: The Epyc Venice CPU adopts CoWoS-L advanced packaging. Because the RDL structure is relatively simple and does not involve the TSV process, part of the orders will use ASE and SPIL as the second supplier.

• Amkor: NVIDIA Vera CPU adopts CoWoS-R advanced packaging. Because the RDL structure is relatively simple and does not involve the TSV process, Amkor will serve as the second supplier for the front-end CoW process, while part of the subsequent oS process orders will be outsourced to ASE and SPIL

Optical Communications: 1.6T and CPO Architecture

• The projected market size for 1.6T optical modules in 2026 is 30 million units (from the supply chain perspective, this can currently be raised to 36-37 million, but we believe it still needs to align with CoWoS capacity, so we will gradually revise it upward in the future).

• The current focus is on 2027. To our knowledge, Google and NVIDIA continue to increase their figures for 2027, expected to be stronger than 2026 (projected to be more than double). 2026 and 2027 are very strong and concentrated on the highest specification 1.6T.

• Industry-wise, China A-share Innolight and US-listed Lumentum will both benefit from this trend. We currently see LITE receiving very large 1.6T optical module orders from Google.

• Regarding CPO (2027 NVL576): The design solution for NVIDIA’s 2027 NVL576 architecture (interconnection of four sub-cabinets) adopts CPO (Co-packaged Optics) interconnection, primarily promoted by Mellanox. Because optical fibers have no distance limitations, the current setup connects 4 sub-cabinets and can simultaneously interconnect 2-4 large Kyber cabinets. (2 cabinets total over 1,000 dies, 4 cabinets total over 2,000 dies). CPO connects the sub-cabinets and does not affect the interconnection of the PCB backplanes within the cabinet.

• 2 Large Kyber Cabinets = 1,152 GPU Dies

• 4 Large Kyber Cabinets = 2,304 GPU Dies

• CPO Supply Chain:

• FAU (Fiber Array Unit): Approximately $350 USD value per Rubin. China A-share Everfu Communication leads significantly in precision and will be the main beneficiary.

• CW Laser: CPO also requires CW light sources, which will exacerbate the shortage of optical chips.

• Shuffle: Fiber optic hub box. Suppliers are Corning + T&S Communications, BizLink.

Source: GF Securities

Have a great weekend! We’ll be back in your inbox next week. Thanks for reading!

Insightology Research

What I Read

Article

• https://www.trendforce.com.tw/presscenter/news/20251208-12821.html?fbclid=IwY2xjawOpVAJleHRuA2FlbQIxMQBzcnRjBmFwcF9pZBAyMjIwMzkxNzg4MjAwODkyAAEeI8EUTkMIko-w_6kiq2tMdlMDCNeD4eSPfYrb0VX1Fb9VnAQyKRE10mTRVpw_aem_bmrzZbv04go4saTPwhPseg

• https://andyljones.com/posts/horses.html?fbclid=IwdGRleAOlVC9leHRuA2FlbQIxMQBzcnRjBmFwcF9pZAo2NjI4NTY4Mzc5AAEe4tTCZ4ERJ07qq9d-6gcJJsQf9jumNbbrtiRuD_GXgw64sY2mkT4A6aSn1xM_aem_dLA_fT6Q0111Is0UwZ_aXQ

• https://www.oaktreecapital.com/insights/memo/is-it-a-bubble?fbclid=IwY2xjawOpe3lleHRuA2FlbQIxMABicmlkETE5SUI4OElhbXdweW5YOFNSc3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHtmE8o2dQT7suPzwPGffnhUhZy62BM2RfbqdjN4CfmAX7SdJqTxfP2_TXj9t_aem_Yri51Pm9U6Q0DD3HYemQ3g

• https://www.trendforce.com/news/2025/12/08/news-tsmcs-cowos-l-s-reportedly-fully-booked-osat-partners-step-up-with-ases-cowop-in-focus/

• https://www.techpowerup.com/339963/tsmc-prepares-cowos-to-copos-shift-with-750-x-620-mm-panels

• https://www.trendforce.com.tw/presscenter/news/20251210-12828.html?fbclid=IwY2xjawOpqJZleHRuA2FlbQIxMABicmlkETFNcnhGNTNhdFdPRTFLV1g0c3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHva294bjeJiHXLCmczw0t-3lD6KhUjf9JDQUI-ZzUTQBZ4iyefVu5iwI2uHn_aem_QTF2aHEc9OEQ152ppf9_1g

• https://www.hpcwire.com/2025/12/09/nvidia-says-its-not-abandoning-64-bit-computing/?fbclid=IwY2xjawOpqNRleHRuA2FlbQIxMABicmlkETFNcnhGNTNhdFdPRTFLV1g0c3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHsssxtVjmwNBMoMFcfTEoYdjEfSnQqbEX4QZSB5Sb1KeG9U93E5SuGb858Yr_aem_S4tDRKPuuzESIiN2CD_fzA

Podcast

Book

• The Big Money: Seven Steps to Picking Great Stocks and Finding Financial Security

Transcript

• AVGO Earnings Call Transcript

• ORCL Earnings Call Transcript

• ADBE Earnings Call Transcript

• SNPS Earnings Call Transcript

• COST Earnings Call Transcript

• LULU Earnings Call Transcript

• RH Earnings Call Transcript

Conclusion and Thesis

1. Starting from 2026, UMC is expected to gradually move away from the highly competitive environment of low-cost mature processes and shift toward high-value mature processes and specialty technologies. This strategic shift, particularly benefiting from its expansion into US manufacturing, is advantageous for raising the company’s Price-to-Earnings (PE) ratio.

2. The Macro View: TSMC Holds the Line, UMC Steps Up：With TSMC (2330.TW) signaling clearly that it has no interest in acquiring an equity stake in Intel, the spotlight has unexpectedly shifted to UMC (2303.TW). The company is seizing this vacuum to execute a crucial strategic pivot, positioning itself as a stabilizer in the global supply chain.

3. The Core Catalyst: The Intel & Polar Alliance：UMC is moving away from the “build-at-all-costs” model toward smart collaboration.

   • The Intel Partnership: UMC will co-develop a 12nm platform with Intel.

     • The Deal Structure: This is a classic “Asset-Light” play. Intel provides the manufacturing capacity (fabs) and capital expenditure, while UMC leverages its IP and extensive client rolodex to fill that capacity.

     • Why it Matters: UMC gains 12nm capacity without the heavy depreciation burden of building a new fab, while Intel solves its utilization issues.

   • The Polar Semiconductor Deal: Simultaneously, UMC is partnering with Polar to supply high-voltage (HV) power components, specifically targeting Tier-1 US automotive giants.

4. The “Hidden Gem”: Silicon Photonics (SiPh)：Beyond the headline deals, UMC is aggressively planting flags in next-gen connectivity.

   • Strategic Partner: Collaborating with IMEC (Belgium), the world’s leading semiconductor R&D hub.

   • The Tech: Entering the Silicon Photonics (SiPh) front-end wafer foundry market using mature 28/22nm processes.

   • The Timeline:

     • 2026: Pilot Production.

     • 2027: Volume Ramp-up.

   • Significance: As data centers demand higher bandwidth, SiPh is the next bottleneck solver. UMC is positioning itself early in the supply chain using cost-effective mature nodes.

5. The Risk Factor

   • Intel’s Execution: The primary overhang is execution risk on Intel’s part. If Intel struggles with yield rates or mass production timelines for the 12nm platform, UMC’s capacity expansion plans could face delays.

The Analyst Take: A Long-Term Re-rating

This is not an Nvidia-style momentum trade. Do not expect UMC to replicate the parabolic moves of AI component suppliers.

• The Narrative: UMC is transforming from a commodity foundry trapped in price wars to a strategic partner with specialized, high-margin offerings.

• The Playbook: This is a value turnaround play. We are looking for a gradual re-rating of the stock as it decouples from the brutal price competition of standard mature nodes.

• Verdict: Patience is required. Accumulate for the long haul as the “asset-light” strategy begins to improve return on equity (ROE) and margins.

Company Overview and Core Business Transformation

UMC, established in 1980, is currently the world’s fourth-largest wafer foundry manufacturer. UMC’s strategy differs from TSMC; it supports several semiconductor chip design companies (such as Novatek, Faraday Technology, PixArt Imaging, and Holtek Semiconductor) through its own capacity and technology to ensure its foundry capacity utilization rate is maintained. Major customers include international companies such as MediaTek, Qualcomm, Samsung, NXP, and Texas Instruments.

Source : Company

Thesis

I. The Core Business: “Boring is the New Stable”

The Short-Term Reality Check Let’s be blunt: if you are looking for explosive quarterly growth right now, look elsewhere. The core business is currently about stabilization, not acceleration. “Flat” is the new “Up.”

• Utilization Rate (UTR) Recovery: We saw UTR climb back to 78% in Q3 2025. We project a slight moderation to 74%–76% in Q4 due to year-end seasonality.

• Pricing Power (ASP): Despite headwinds, U.S. dollar-denominated ASP (Average Selling Price) is holding steady for Q4 2025.

• The China Factor (The Elephant in the Room): The competitive landscape remains brutal. Chinese foundries (SMIC, Hua Hong) have aggressively expanded capacity across 28nm–90nm.

  • The Threat: In 2025, the bulk of global mature capacity growth is coming from China. Their quotes are often 30–40% lower than Taiwanese peers.

  • The Impact: We anticipate a general 5–10% pricing erosion across the mature node market in 2025.

The Strategic Shift: From Volume to Value UMC is smart enough not to fight a price war they can’t win. They are actively pivoting the revenue mix:

• Old UMC: Low-end, commoditized legacy nodes (highly sensitive to Chinese price dumping).

• New UMC: High-Value Legacy + Specialty Processes.

  • Focus Areas: 22/28nm, Automotive, PMIC (Power Management), High-Voltage BCD, eNVM, and RFSOI.

  • Why it works: These are “sticky” products with higher technical barriers. Clients here care more about reliability and performance than saving a few cents on the wafer.

• Growth Spots: Management expects double-digit growth in 22/28nm revenue and high-single-digit growth in 8-inch PMIC.

II. The Alpha Play: The “Asset-Light” American Expansion

The Headline: While TSMC politely declined to acquire equity in Intel, UMC stepped up to fill the vacuum—not with cash, but with competence.

The Strategy: “Fabless” Foundry Expansion UMC is expanding its US footprint without the heavy anchor of capital expenditure (Capex). This is Asset-Light Expansion 101: Don’t build the factory; just run it.

1. The Intel Alliance (Arizona)

• The Deal: Utilizing Intel’s depreciated Fab 12, 22, and 32 to develop a 12nm FinFET platform.

• The Logic: Intel provides the facility; UMC brings the clients (Realtek, MediaTek, Broadcom).

• Launch: Production scheduled for 2027.

• Synergy: Intel fills empty fabs; UMC gets 12nm capacity without spending billions on concrete and steel.

2. The Polar Semiconductor Partnership (Minnesota)

• The Context: Polar, originally a Control Data factory and later owned by Japan’s Sanken Electric, recently brought in US Private Equity (Niobrara & Prysm Capital) to qualify for the CHIPS Act.

• The Problem: Polar has the funding and the “US Entity” status, but lacks broad technical depth.

• The Solution: UMC injects its 8-inch manufacturing expertise and client base.

• The Niches: Focusing on High-Voltage Power and Sensors (Automotive). Clients include Allegro MicroSystems and Sanken Electric.

Why This is a “Triple-Win”

1. UMC: Gains “Made in USA” capacity at zero construction cost.

2. Intel / Polar: Monetizes idle assets and gains access to the lucrative “mid-range” mobile/comm market (MediaTek, Qualcomm, etc.).

3. Clients: Secures a non-China supply chain for mature nodes, hedging geopolitical risk.

The Financial Impact (The Long Game)

• Short Term (1-2 Years): Negligible financial contribution. UMC is not charging licensing fees during the setup phase.

• Long Term (2027+): This is where it gets interesting.

  • Revenue: Potential for NT$25B–30B (approx. US$800M–1B) annually.

  • Significance: This represents ~10–12% of current revenue.

  • The Kicker: This revenue is “High Purity” Profit. Since UMC didn’t pay to build the fabs, there is no depreciation drag. It flows straight to the bottom line.

III. The Bonus: Silicon Photonics (SiPh)

The “Hidden Call Option” While the market obsesses over GPU computing, data transmission is the bottleneck. Enter Silicon Photonics.

• The Tech: UMC is focusing on PICs (Photonic Integrated Circuits). Unlike logic chips, PICs do not require 3nm processes.

• The Edge: Mature nodes (130nm/180nm) are actually better for this—lower cost, lower transmission loss, and higher CP value.

• The Partner: Collaboration with IMEC (Belgium) to license the 12-inch iSiPP300 process.

• The Timeline:

  • 2026: Pilot Production.

  • 2027: Volume Shipments.

Financial Impact Estimate

• The “Pure Profit” Factor: While a 12% revenue boost (approx. NT$27B) is significant, the real story is the quality of these earnings. Because UMC is utilizing existing factories (Intel/Polar) rather than building new ones, there is minimal depreciation drag. This revenue should flow through to the bottom line at a much higher margin than UMC’s standalone fabs.

• ASP Premium: Note the estimated ASP for the Intel 12nm node ($4,000+). This reflects the “Geopolitical Premium”—clients are willing to pay more for chips guaranteed to be manufactured on U.S. soil to avoid tariff or supply chain risks.

Risk：What Could Break the Thesis?

1. The “China Drag” on Core Earnings (The Involution Risk)

The threat of commoditization in the mature nodes remains the elephant in the room.

• The Scenario: If Chinese foundries (SMIC, Hua Hong) launch an aggressive price war (”involution”) to capture 28nm market share before 2027, UMC’s pricing power will erode.

• The Impact: We estimate this could compress UMC’s standalone EPS from ~NT$4.0 to ~NT$3.0.

• The Net Result: Such a decline in the core business would effectively neutralize the financial accretion gained from the new U.S. operations.

2. Intel’s Execution Risk (The “Wildcard”)

In my view, this is the single largest variable in the equation.

• The Scenario: Intel’s financial health is volatile. Risks include a further deterioration of their balance sheet forcing a divestiture of fab assets, or technical friction during the 12nm technology transfer.

• The Impact: If the deal crumbles due to counterparty risk, the projected ~NT$0.60 EPS upside from the U.S. expansion effectively evaporates to zero.

Conclusion

The Analyst Take: UMC is a long-term structural transformation play, not a short-term trade.

• Manage Expectations: Do not expect the parabolic stock price action seen in AI hardware component suppliers. This is not an Nvidia proxy.

• The Narrative: We are watching a slow-burn decoupling. UMC is gradually exiting the “low-tide” cycle of commoditized price wars and moving toward value-added partnerships.

• Conclusion: This thesis requires patience. The “Asset-Light” strategy will take time to reflect in the ROE, but the direction of travel is correct. Wait for the strategy to bear fruit.

Disclaimer：My crystal ball is currently in the shop for repairs. While I believe in this thesis, I cannot predict the future (or what Intel will do next week). I am sharing my notebook, not managing your portfolio. If you make millions, you’re welcome. If you don’t, remember: I’m just a writer on the internet.

Elon Musk recently invoked the Kardashev Scale, sparking a conversation about an undeniable reality: AI Data Centers are devouring energy at an exponential rate.

Historically, we viewed this problem through the lens of conservation. We worried about efficiency. But the math is becoming unforgiving. Analysis from the SRC’s Decadal Plan for Semiconductors suggests a grim timeline: by approximately 2040, global computing demand will outstrip the entire world’s energy supply.

We are sprinting toward an invisible but solid “Energy Wall.” Without radical breakthroughs, the exponential growth of AI could hit a hard ceiling as early as 2030.

Source:The Wall Street Journal, WSJ

But what if the solution isn’t to use less, but to access more? What if the solution is to move the infrastructure off-planet entirely?

It sounds like science fiction, but it is actually a conclusion rooted in hard physics that dates back to 1960.

1. Not Sci-Fi, But Science (Princeton, 1960)

To understand why we are now talking about launching servers into orbit, we must travel back to June 1960.

While the world was obsessed with the Space Race, Freeman Dyson, a legendary physicist at the Institute for Advanced Study in Princeton, published a short but profound paper in the prestigious journal Science.

The paper was titled “Search for Artificial Stellar Sources of Infrared Radiation.”

At the time, astronomers were struggling with how to find extraterrestrial civilizations. Most were focused on “listening” for radio waves (like in the movie Contact). Dyson proposed a completely different perspective: Don’t listen for what they say; look for what they consume.

Dyson’s logic boiled down to three core concepts:

• A. The “Exponential Growth” is Irreversible: Dyson noted that any industrial civilization (humans included) experiences “Malthusian growth.” Simply put, a civilization’s appetite for energy grows exponentially. Eventually, planetary resources (fossil fuels, nuclear, or surface solar) hit a hard limit.

• B. The Only Exit is the Star: The Earth intercepts only about one two-billionth of the Sun’s total energy. The rest flows uselessly into deep space. Dyson deduced that an advanced civilization must eventually build a structure to intercept and collect that lost energy.

• C. The Infrared Signature (Waste Heat): This was his brilliant insight into thermodynamics. Energy used is not destroyed; it is converted into heat. If a civilization wraps a star to generate electricity, that heat must be radiated away to prevent the structure from melting. Therefore, an advanced civilization wouldn’t look like a bright star—it would look like a dark object glowing intensely in Infrared.

2. Debunking the Myth: It’s Not a “Sphere”

Pop culture (like Star Trek) often depicts a Dyson Sphere as a solid, hollow metal shell completely enclosing a star.

This is a misunderstanding. Dyson himself later clarified that a solid shell is mechanically impossible; gravity would tear it apart.

The scientifically accurate concept is a “Dyson Swarm.” Imagine a swarm of mosquitoes surrounding a streetlamp. It involves billions of independent satellites, solar mirrors, and space stations orbiting the sun in a dense formation. They form a loose but massive network to harvest energy.

3. The Convergence: Why We Are Building the “First Lego Brick” Today

Back to 2025. Why are we dusting off a 60-year-old physics paper? Because AI has driven us right into the “Energy Wall” Dyson predicted.

The concept of Space-Based Data Centers proposed by today’s tech industry is essentially the micro-implementation of a Dyson Swarm. Here is why the physics and economics finally make sense:

• Unlimited Solar Power (Concept B): Instead of waiting for that tiny fraction of sunlight to filter through Earth’s atmosphere, we can place data centers in Sun-Synchronous Orbits (SSO) or on the Moon’s “Peaks of Eternal Light.” There, solar panels receive pure, unfiltered energy 24/7.

• Radiative Cooling (Concept C): The “waste heat” problem Dyson identified has a natural solution in space. The background temperature of the universe is 2.7 Kelvin (-270°C). Projects like the EU’s ASCEND and startups like Lumen Orbit are validating how to dump heat directly into this deep freeze, eliminating the need for the massive water-cooling systems used on Earth.

• In-Situ Consumption: There is no need to beam electricity back to Earth (which is inefficient). We use the power right where it is gathered to run the compute, and then beam only the processed data back.

• Plummeting Launch Costs: Thanks to SpaceX (Falcon 9 and Starship) and Blue Origin, the cost to put heavy hardware into orbit is crashing down, making this economically viable.

• Satellites as Servers: Future constellations like Starlink or Kuiper won’t just be communication relays; they will be orbital compute nodes.

Further Reading & Deep Dives

To prove that sending data centers into orbit is not just a pipe dream, here are three key references from government, academia, and the private sector that form the theoretical bedrock of “Space-Based Computing.”

1. The Definitive Official Study: The EU ASCEND Project

Funded by the European Commission, this massive study aimed to verify whether orbital data centers could be a solution for achieving net-zero emissions.

• Project Title: ASCEND (Advanced Space Cloud for European Net zero emission and Data sovereignty)

• Lead Agency: Thales Alenia Space (on behalf of the European Commission)

• Key Conclusions:

  • Technical Feasibility: The report confirms that utilizing space’s “unlimited solar energy” and “extreme cold environment” effectively solves the two biggest pain points of terrestrial data centers: power consumption and heat dissipation.

  • Environmental Threshold: The report honestly notes that to be environmentally viable, we need to develop new launch systems that are 10x greener than current rockets (aligning with the goals of Starship regarding full reusability and fuel efficiency).

2. Academic Architecture Analysis: Cost Models & Cooling

This paper provides a detailed examination of the economic models for Space-Based Data Centers (SBDC), specifically quantifying the trade-offs between latency and cost.

• Paper Title: Space-Based Data Centres: A Paradigm for Data Processing and Scientific Investigations

• Publication: A. A. Periola & M. O. Kolawole (2019, IEEE Access)

• Core Insights:

  • The Cost Trade-off: While launch costs remain high, SBDCs offer massive savings on expensive terrestrial land and increasingly scarce cooling water resources.

  • Best Use Case: The architecture is described as perfect for “Non-Real-Time AI Training”, where high bandwidth is needed but millisecond latency is not critical.

3. Industry Whitepaper: The Business Case for AI Training

This is the most relevant document regarding the current startup ecosystem and Elon Musk’s perspective, offering a concrete path to commercialization.

• Whitepaper: Why we should train AI in space

• Publisher: Lumen Orbit (now Starcloud) — A Y Combinator-backed startup.

• Core Arguments:

  • Energy Arbitrage: Solar panels in space are 5-8x more efficient than on Earth due to the lack of atmospheric interference and the absence of night (in specific orbits).

  • Ultimate Passive Cooling: Leveraging the Stefan-Boltzmann Law, heat can be radiated into the 2.7K (-270°C) background of deep space with near-zero energy consumption.

  • Strategic Positioning: The paper proposes a specific business model: focusing exclusively on AI Model Training rather than Inference, cleverly sidestepping the physical limitations of communication latency.

Conclusion

By moving AI compute to space to harvest solar energy directly, we are effectively placing the first “Lego brick” of a Dyson Sphere.

AI, once thought to be limited by Earth’s energy constraints, might be the very catalyst that pushes humanity out of the cradle. We are watching science fiction become engineering reality, one orbital server rack at a time.

Note: The Kardashev Scale

The Kardashev Scale was introduced in 1964 by Soviet astronomer Nikolai Kardashev. He posited that as a technological civilization advances, its energy consumption grows exponentially. He classified civilizations into three levels:

• Type I (Planetary): A civilization that can utilize all the energy available on its home planet (Earth).

• Type II (Stellar): A civilization capable of harnessing the total energy output of its star (the Sun)—i.e., by building a Dyson Swarm.

• Type III (Galactic): A civilization that can control the energy of its entire galaxy (the Milky Way).

At our current technological level, humanity has not yet reached Type I status. We are currently estimated to be at approximately Type 0.7.

Source：

https://all-fiction-battles.fandom.com/wiki/The_Kardashev_Scale

https://fermatslibrary.com/s/search-for-artificial-stellar-sources-of-infrared-radiation

——-

Life： Efficacy, Slow Living, and the Fear That Drives Us

The Distinction Between Efficiency and Efficacy

This week, while using AI tools, I had a major realization: being efficient is different from being effective. Both are important, but from the perspective of achieving goals, the primary focus should be on efficacy. Getting things done quickly isn’t necessarily useful; sometimes it’s just fast.

In life, the more important the matter, the less you need to chase efficiency. For things you genuinely enjoy, you should pursue slowness, prolonging the enjoyment of the process. Sometimes, high speed becomes a habit and creates a shield, ignoring the truly important parts, which might ultimately result in a lot of useless work, rendering the efficiency pointless.

AI will significantly change work efficiency, but whether it is effective primarily depends on the human user. I recently found that using AI to summarize long articles or podcasts often yields only dispensable/non-essential information. Moreover, many people online already help shorten and highlight the main points, which you repeatedly encounter. For example, the main impressions I had after listening to Jensen Huang on Joe Rogan today were not the ones provided by the AI summaries. The feelings and insights gained from listening to the whole thing without an AI summary were completely different from the text later summarized by the AI.

Fear: The Common Driver Behind Success

Especially Jensen’s interview reminded me of a previous interview I saw with Mike Tyson and Sadhguru. Jensen mentioned that he is driven by fear, not greed. Both of them came from challenging environments and transformed these experiences into later achievements.

• Jensen Huang‘s parents sent him and his brother to the United States, where he spent time in a school in one of America’s poorest counties (Clay County, Kentucky).

• Mike Tyson grew up in an environment filled with bad things (such as witnessing death and prostitutes).

Both admitted that fear, anxiety, and struggle were constant states throughout their careers. Jensen Huang confessed that his drive stems more from the motivation not to fail than the motivation to succeed. He wakes up every morning feeling anxious, worried that the company might soon collapse. This “sense of vulnerability, the sense of uncertainty, the sense of insecurity“ has never left him. Mike Tyson also said that despite being considered one of the “Toughest guys in the world,” he confessed to Sadhguru that he has been afraid his entire life. Even now that he is successful, he cannot pinpoint exactly what he is afraid of.

From the two of them, we can see that the journey of life is like surfing; you cannot predict the waves, but you must possess the skills to handle the existing ones. Mike Tyson’s life journey also reflects this, as Sadhguru once commented that without maintaining balance, talent and ability can eventually turn against oneself.

The most important lesson for me, as someone who recently arrived in the U.S., is cultivating balance and the sense of life’s pleasantness. The pleasantness of the body is health; the pleasantness of the mind is peace; the pleasantness of the emotion is love; and the pleasantness of life energy is ecstasy. If we do not make our minds and emotions pleasant and peaceful, then we cannot feel happiness even when enjoying dinner or taking a walk down the street

Market: AI Industry Strategy, Acquisitions, and the Competitive Landscape

1. Credo Launches ALC to Disrupt Mid-to-Short Distance Market; BizLink Strategically Acquires New Fusheng

Credo launched ALC using MicroLED technology to disrupt the mid-to-short distance optical transmission market within 30 meters, while BizLink (貿聯) supplemented key optical fiber manufacturing processes through the acquisition of New Fusheng, positioning itself to be a major beneficiary of this potentially huge business opportunity.

During this conference call, Credo announced the new product line ALC (Active LED Cables). I personally believe this represents potential upside for BizLink (which the market has not priced in). This area could potentially erode the market share of optical module manufacturers in the mid-to-short distance range (30m).

ALC Product and Strategy:

• Business Model: The business model for ALC is exactly the same as for AEC (Active Electrical Cables): Credo provides the core chip (MicroLED optical engine), which is then handed over to the cable manufacturer for packaging, testing, and assembly, and finally sold under the Credo brand to Amazon or Microsoft. The probability of switching contract manufacturers is low because this requires validation by the CSPs (Cloud Service Providers).

• Technology & Scope: ALCs combine the MicroLED technology from the previously acquired company, Hyperlume. ALCs are expected to provide the same reliability and power consumption characteristics as AECs, but with a thinner wire gauge and lengths up to 30 meters.

• Market Growth: The network scale will expand from inter-rack to inter-row. The number of connections is expected to reach 10 times that of AEC, with room for increased ASP. This is highly suitable for row-level (scale-out) expanded networks.

• Timeline and TAM: Sampling to major customers is projected for FY2027, with preliminary volume production starting in FY2028. The TAM (Total Addressable Market) for ALC is eventually expected to be more than double the TAM for AEC.

Analysis of the Hyperlume Acquisition: Hyperlume, a startup specializing in MicroLED optical interconnect technology, completed its transaction at the end of September 2025. Hyperlume holds the core technology and intellectual property for ALC. Therefore, Credo will directly offer ALC solutions to its customers (primarily Hyperscalers such as Amazon and Microsoft).

Notably, Hyperlume founders Nadeem Mirza was Credo’s former VP of Product, and Parker Robinson was Credo’s former Director of System Engineering. The strategy was to allow former employees familiar with the company’s strategy to leave and establish a startup (Hyperlume), using external venture capital (such as Intel Capital, LG, EIC) for funding and experimentation. Once the technology was successfully validated (that MicroLED was truly viable), Credo then funded the buyback. If developed internally at Credo, the project might have failed due to financial reporting pressures or the prioritization of other R&D resources.

BizLink’s Strategic Move: This presents competition for major optical communication module manufacturers (Coherent, Innolight) and major cable manufacturers (Amphenol, Molex). These companies currently supply large volumes of 400G/800G optical modules and AOCs to Hyperscalers. If ALC succeeds, it will directly erode their market share in the “mid-to-short distance (within 30 meters)“ segment.

The risk for BizLink to watch out for is whether a second connecting cable manufacturer capable of providing fiber optic capabilities—such as Amphenol or Molex—will enter the market.

BizLink announced this week the acquisition of New Fusheng Optoelectronics (New Fusheng Optoelectronics), specifically to strategically position itself for ALC. New Fusheng is an enterprise with a “Taiwanese capital background“ (wholly owned by Taiwan, Hong Kong, or Macau corporate entities). New Fusheng provides OEM/ODM manufacturing for Corning, CommScope, and other optical communication brands.

• Technology Requirement: ALC (Active LED Cable) combines a MicroLED chip with fiber optic bundles. ALC needs to use high-density fiber optic bundles to transmit the optical signal emitted by the MicroLED.

• New Fusheng’s Strength: New Fusheng’s strength lies in specializing in the manufacturing of MPO/MTP high-density fiber optic connectors and multi-fiber patch cords. New Fusheng possesses the mass production capability to precisely align multiple optical fibers and package them into a connector. (The most technically difficult aspect is polishing the connector end face to a specific angle to prevent light reflection back, which would interfere with the laser).

• Strategic Significance: While BizLink is very strong in AEC, the precision polishing and alignment of fiber optics is a completely different process. Therefore, by acquiring New Fusheng, BizLink avoids having to build a fiber optic factory from scratch and directly obtains the necessary fiber assembly capacity for ALC.

2. AWS Trainium 4 Adopts Nvidia NVLink Fusion/NVSwitch

AWS re:Invent announced this week that Trainium 4 will adopt Nvidia’s NVLink Fusion technology for chip-to-chip interconnection. This move not only allows Nvidia to successfully penetrate the AWS ecosystem with NVSwitch hardware, earning substantial revenue, but also severely squeezes the living space for third-party Switch suppliers like Broadcom.

The Shift to XPU-to-XPU Interconnect: Previously, when Nvidia licensed NVLink C2C to CPU manufacturers like Intel or Arm, the focus was embedding the IP inside the customer’s CPU to connect the CPU with the GPU/XPU. In these scenarios, a standalone Switch chip was usually not required, nor was it related to “Scale Up” (vertical scaling). However, AWS’s focus this time is completely different: AWS chose NVLink Fusion to handle high-speed interconnection between “XPU-to-XPU,” which is the core “Scale Up“ requirement.

In this solution, Nvidia will provide a small NVLink Fusion Chiplet, allowing the customer’s Trainium 4 chip to be integrated into the package. The main function of this Chiplet is to enable the customer’s XPU to communicate with the standard NVSwitch chip.

Nvidia’s Business Logic: Nvidia’s business logic is very clear: selling IP licensing fees is a small matter. For Nvidia, the largest contribution and source of revenue is the sale of expensive NVSwitch chips or Trays/Boxes.

We can divide the Switch functions in AI systems into three categories:

1. CPU-to-XPU Switch: Responsible for communication between the CPU and the AI processor.

2. XPU-to-XPU Switch (Scale-Up Switch): Responsible for high-speed interconnection between AI chips (this is the core market targeted by NVSwitch).

3. Networking Switch (Scale-Out Switch): Responsible for network communication between racks or clusters.

AWS’s Trainium 4 chose to use NVSwitch for category (2), which is the Scale-Up of XPU-to-XPU. In this segment, Nvidia’s NVSwitch product constitutes the “all green“ portion, meaning the customer has no room for customization or choosing third-party products.

Market Impact and Timeline: This constitutes a huge impact on third-party chip suppliers. Whether they provide PCIe Switches for scale-up or Broadcom’s Ethernet Scale Up Switch, they will all be affected or even replaced by NVSwitch in the XPU interconnection market for Trainium 4.

According to Nvidia’s own GB200/300 NVL72 design ratio, the ratio of GPUs to NVSwitches is approximately 4:1. This means that if AWS Trainium 4 shipments are substantial, Nvidia could sell hundreds of thousands of NVSwitch chips.

However, Trainium 4 will not be available until 2027 at the earliest. In 2026, the market will still be dominated by Trainium 3, and the XPU interconnection technology used then might still be AWS’s proprietary NeuronSwitch or third-party PCIe Switches.

What I Read

• https://www.bizlinktech.com/zh-tw/news/bizlink-strengthens-its-optical-interconnect-capabilities-with-acquisition-of-xfs-communications-inc

• https://aws.amazon.com/blogs/aws/top-announcements-of-aws-reinvent-2025/

• https://huggingface.co/blog/hf-skills-training?fbclid=IwY2xjawOgBWhleHRuA2FlbQIxMQBzcnRjBmFwcF9pZBAyMjIwMzkxNzg4MjAwODkyAAEehynrP3Gj64BagKHDx6cG_nF8R4tuVFFwGn6fgpHeHnmsk81jEHqXS2OAjyY_aem_VtRruhu7l8Sin5xCjVs0tw

• https://developer.nvidia.com/blog/aws-integrates-ai-infrastructure-with-nvidia-nvlink-fusion-for-trainium4-deployment/?fbclid=IwY2xjawOgBq5leHRuA2FlbQIxMABicmlkETFXTUZFb2N5NGdJNTRmSWlBc3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHjZG2Mc9tX7VfzU9xq3-FdnbMnDqCwCW9n3jZM4LEPXWF4ExbcgwHD8zV1W2_aem_3wOyetes38VOeYNWBZlDJQ

• https://developer.nvidia.com/blog/advancing-performance-with-nvidia-sharp-in-network-computing/

• https://credosemi.com/blogs/active-electrical-cables-aec-becoming-an-important-part-of-data-center-architectures/

Transcript

• Credo Earnings Call Transcript

• Marvell Earnings Call Transcript

I’ve tried to break down next year’s ASIC trends and supply chain outlook in this post. Just wanted to share my findings with you all! If I got anything wrong, please let me know. I’d love to correct it so we can all grow and learn together

The trajectory for high-end AI ASIC chips suggests they will become the next critical growth engine for the AI server supply chain, surpassing high-end GPU growth in terms of percentage increase from 2025 through 2026.

Key Market Momentum and Forecasts

High-end ASIC chip shipments are projected to experience significant year-over-year (YoY) growth, exceeding 40% in both 2025 and 2026.

• ASIC Growth Superiority: The growth momentum of high-end ASIC chips is notably superior to high-end GPUs. High-end ASIC YoY growth is projected at 40.6% in 2025 and 41-43% in 2026, compared to high-end GPU YoY growth of 30% and 23% respectively.

• 2026 Shipment Volume: Total high-end ASIC shipments are forecasted to reach 7.234 million units in 2026.

• Primary Driver: Google’s TPU chips are identified as the strongest growth engine for ASIC shipments, continuing to lead the volume share among ASIC chips. Google’s TPU shipments are projected to reach 3.325 million units in 2026, showing an increase of nearly 1 million units compared to 2025.

--------------------------------------------------------------------------------

Google TPU (The Strongest Driver)

Google’s TPU supply chain is a key area of focus for 2026.

• TPU v7p (Ironwood): This mainstream version is scheduled for mass production in 3Q25. It will use the 3nm process and 192GB HBM3E memory.

◦ Architecture: It maintains the classic 64-accelerator design and utilizes the HLC architecture, with four accelerators per single motherboard.

◦ Scale-up: The scale-up range will be expanded to 9,216 accelerators across 144 cabinets. This expansion relies on the OCS (Optical Interconnect) architecture, which provides a unidirectional bandwidth of 600GB/s.

• Future Planning: The TPU v7e (A5921) is scheduled for mass production in 2H26. MediaTek (聯發科) is expected to mass-produce the TPU v7e by the end of 2026.

• Supply Chain: Amkor is anticipated to enter the TPU advanced packaging supply chain in 2H26.

AWS Trainium

AWS Trainium shipments are expected to decline in 2026 compared to 2025, as the platform faces a product transition period.

• 2026 Shipment: Shipments are estimated at 1.476 million units in 2026.

• Trainium3: This version is expected to enter mass production in 2Q26, utilizing the 3nm process and 144GB HBM3E memory.

• Trainium3 MAX: This server system will debut a liquid cooling architecture (water-to-air cooling solution). The architecture is similar to NVL72, featuring four ASICs per computing board.

◦ Interconnect: AWS is collaborating with Astera Labs on the Scale-up solution to achieve Any-to-Any interconnectivity.

◦ Future Outlook: Given the close partnership between AWS and Astera Labs (a major promoter of UALink), the next-generation Trainium4 server is speculated to adopt the UALink architecture, which is projected to launch its first solution in 2027.

• Supply Chain: Alchip is the main chip design partner for Trainium3. Wiwynn and QCT are responsible for the front and back-end server assembly.

Meta ASIC (Athena, Iris)

Meta’s high-end AI ASIC products are expected to achieve volume shipments in 2026.

• 2026 Shipment: Shipments are projected to reach 600,000 units in 2026, a significant increase from the 50,000 units shipped in 2025.

• Athena: Scheduled for mass production in 4Q25 (5nm, 216GB HBM3E). The chip will utilize a blade server cabinet design.

• Iris: Scheduled for mass production in 2H26 (3nm, 288GB HBM3E).

• Roadmap: Meta has a rapid product roadmap, including Arke (2027f, 2nm, HBM4) and Olympus (2028f, 2nm, HBM4E(e)).

• Supply Chain: Broadcom (博通) is the sole design partner for Meta ASIC, supplying the complete accelerator interconnect and networking solutions (including Tomahawk 5/6 and Tomahawk Ultra). Quanta (廣達) is actively vying for the Iris server back-end assembly orders.

Microsoft Maia

Microsoft’s next-generation Maia 200 is set for mass production in the second half of 2026.

• Maia 100: This was an experimental product, with combined shipments of 23,000 units across 2024 and 2025.

• Maia 200: Expected to enter mass production in 2H26, with an estimated full-year shipment of 330,000 units. It uses the 3nm process and 128GB HBM3 memory.

• Future Plans: The subsequent Maia 300 may potentially adopt Intel 18A process and advanced packaging technology.

Open AI

Open AI has announced an ambitious long-term ASIC deployment plan.

• 10GW Plan: Open AI declared on October 13, 2025, a plan to deploy 10GW of ASIC by the end of 2029, estimated to total at least 5 million accelerators.

• Titan1: The first-generation AI ASIC (Titan1) is projected for mass production in 4Q26 (3nm, 144GB HBM4).

• Titan2: Projected for mass production in 4Q27 (2nm, 288GB HBM4)

Table 1: High-End AI Accelerator Shipment Volume and Forecast (2024–2026) (Units: Thousands)

Table 2: Key ASIC Accelerator Specifications and Planning (2024–2028)

Source：DIGITIMES

Before we dive into the this week’s shots, I want to share some highlights from our Thanksgiving trip to Rocky Mountain National Park and the Great Sand Dunes.

As we wrap up the year, I have three things I am deeply grateful for:

Health & Clarity: For having a healthy body and a clear mind that allows me to keep creating and outputting content.

New Beginnings: For the courage to make the big decision to move to the US this year, restarting both my career and my life.

Support System: For my family and my wife, who have supported my decisions and stood by my side every step of the way.

Market

1. Who actually gets hurt by Google selling TPUs?

TL;DR : It’s not NVIDIA. It’s the underperforming internal ASICs.

The narrative that “Google renting out TPUs kills NVIDIA” is lazy analysis. Here is the reality of the Hyperscaler/CSP landscape: Everyone wants to break NVIDIA’s stranglehold. The goal is to reduce NVIDIA’s share of their compute spend to a custom ratio (e.g., 60% or 50%). To do this, they build custom ASICs.

The Problem: Building a competitive ASIC is incredibly hard. It’s not just about the chip; it’s about vertical and horizontal scaling, network clustering, and the full software stack.

• The investment is massive.

• If a Hyperscaler forces an inferior internal ASIC onto their teams just to save money, they risk degrading their core product performance (search, recommendations, AI models). This is “tripping over dollars to pick up pennies.”

The Solution: Google offering TPUs to other companies provides a high-quality “Plan B.” If a Hyperscaler’s internal ASIC roadmap is unstable or underperforming, they can now rent Google TPUs to diversify away from NVIDIA. In this scenario, Google TPU acts as a “Second Source” competitor to AMD and internal ASICs, not necessarily a replacement for NVIDIA’s flagship performance.

2. The Real Story: NVIDIA has been playing catch-up to TPU

The market has the timeline backward. People think Google TPU is finally catching up to NVIDIA. Reality: NVIDIA has spent the last two years frantically trying to catch up to Google’s TPU Pod/Cluster architecture.

• The TPU Lead: Google has trained Gemini on its own TPUs from day one. The previous generation TPU v4 (with 4,096 chip scale-up) was industry-leading.

  • Note: Even when Google had the best clusters, OpenAI (using “weaker” NVIDIA GPUs) beat them initially. Gemini 3’s recent success isn’t just because the TPU got better; it’s because Google’s AI science, algorithms, and data engineering improved. Hardware is the enabler, but the model architecture is the differentiator.

• The NVIDIA Panic (2023-2024): If Google had opened up TPU sales in 2023, NVIDIA would have been in trouble. Back then, NVIDIA’s scale-up capabilities were limited (NVL8).

  • Remember the chaos? After the GH200 failed to gain traction, NVIDIA aggressively pivoted to the GB200 NVL72 architecture.

  • They changed specs constantly. The Taiwanese supply chain (ODMs and component makers) had thousands of R&D engineers doing “volunteer work,” redesigning boards repeatedly because NVIDIA was rushing.

  • Why the rush? Jensen Huang wasn’t worried about AMD; he was terrified of the gap between his clusters and Google’s TPU Pods.

The Current Landscape (Late 2025): Now, the gap has closed.

• NVIDIA: The GB200/GB300 NVL72 is stabilizing. While the scale-up domain (72 GPUs) is smaller than TPU’s (4,096+), the NVL72 rack offers massive advantages: NVLink C2C coherency, shared memory, and the ability to link CPUs and 480GB of LPDDR5X into a massive VRAM pool.

• Google: Relies on OCS (Optical Circuit Switches) for interconnects.

• Verdict: It is now a balanced field. Each has its strengths.

By the time Google opens TPUs to other Hyperscalers in late 2025, NVIDIA will have solidified its position with Blackwell (GB300), FP4 precision, and the Rubin architecture.

3. Agentic AI: Why the “Genius” Model is Overrated

New Research: NVIDIA’s “ToolOrchestra” & The Butler Philosophy

NVIDIA recently published a paper that shatters the “Bigger is Better” myth. They trained a small 8B parameter model (ToolOrchestra) that scored 37.1% on the “Humanity’s Last Exam” (HLE) benchmark—beating OpenAI’s GPT-5 (35.1%).

The Insight: The “Self-Enhancement Bias” Big models like GPT-5 suffer from “Affluenza.” When given a task, GPT-5 tries to do everything itself (98% of the time) or delegate to a mini-version of itself.

• Analogy: This is like hiring a $2,000/hour top-tier lawyer. You ask them to buy you a coffee, and they insist on doing it personally, charging you their hourly rate, instead of using an app or an intern. It’s a commercial disaster.

The “General Contractor” Approach NVIDIA’s 8B model acts like a smart General Contractor (or Jarvis from Iron Man). It isn’t a genius; it is a Steward.

• It has a “Rolodex” of tools.

• Need math? It calls a Python interpreter.

• Need facts? It runs a Google Search.

• Need deep reasoning? Only then does it pay to call the expensive GPT-5 API.

The Business Case:

• Cost: 30-40% of a monolithic model.

• Efficiency: 2.5x better at solving complex tasks.

Strategic Implications: This defines the second half of the AI game. The competition isn’t just about who has the biggest brain (Model); it’s about who builds the best Compound System. NVIDIA is signaling to developers: We don’t need more arrogant geniuses. We need budget-conscious, highly effective managers. This validates NVIDIA’s “Scale Across” strategy—controlling the flow of compute is more important than just owning the raw model weights.

https://arxiv.org/pdf/2511.21689

What I Read

• https://www.trendforce.com.tw/presscenter/news/20251125-12794.html?fbclid=IwY2xjawOZj3xleHRuA2FlbQIxMQBzcnRjBmFwcF9pZBAyMjIwMzkxNzg4MjAwODkyAAEeZkmmME9_jYj27D3WsiJabbD-RI1mscNpVQsJvRuPa4SiG3uczaMXueo0Jng_aem_J2k3-uqn1D8kkyGb-qO-WA

• https://finance.technews.tw/2025/11/28/sk-hynix-is-​​actively-expanding-its-standard-dram-production-capacity/?fbclid=IwY2xjawOZj7JleHRuA2FlbQIxMABicmlkETFCT2dxM0hGZWJ2VmhXSWQxc3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHr4oGY-HA_xiILe7x5mV7bIFIGlHhviJH3FHHZa0A_CbC4Ck5ygxQoYiGHVk_aem_wx8mrF3PBsdfypW3qqFAdw

SemiAnalysis

TPUv7: Google Takes a Swing at the King

The two best models in the world, Anthropic’s Claude 4.5 Opus and Google’s Gemini 3 have the majority of their training and inference infrastructure on Google’s TPUs and Amazon’s Trainium. Now Google is selling TPUs physically to multiple firms. Is this the end of Nvidia’s dominance…

Read more

9 months ago · 255 likes · 8 comments · Dylan Patel, Myron Xie, Daniel Nishball, Wei Zhou, Jeremie Eliahou Ontiveros, Ivan Chiam, Cheang Kang Wen, Clara Ee, Wega Chu, Kimbo Chen, AJ Kourabi, and Michael Chen

• https://x.com/ArtificialAnlys/status/1993878037226557519?s=20&fbclid=IwY2xjawOZkZ9leHRuA2FlbQIxMABicmlkETFCT2dxM0hGZWJ2VmhXSWQxc3J0YwZhcHBfaWQQMjIyMDM5MTc4ODIwMDg5MgABHmVHq97RsBnAxyZ5F2iyRSM17lrXAn_w7MaxtpKVYiHsumQLspAQbSTRrMbN_aem_TLC1nXqXrH0qY7ejBnVj4A

Nomad Semi

Deep Dive on HBM

This deep dive on HBM is written jointly with Ray Wang…

Read more

a year ago · 175 likes · 10 comments · Moore Morris and Ray Wang

This thanksgiving trip took us from Austin to Denver. Usually, we plan our stops around Costco, but on this route through the Texas panhandle, there was almost nothing—except for Love’s Travel Stops.

Seeing so many of them sparked my curiosity. As investors and founders, we often look at high-tech moats, but Love’s offers a masterclass in business model evolution. It’s not just a “gas station”; it is a comprehensive service ecosystem for the logistics industry.

Here is the breakdown of an empire built on diesel and smart pivots.

1. The Origin Story: Spotting Opportunity in Failure

Love’s is fully owned by the Love family (founded by Tom Love). It started in Watonga, Oklahoma, when Tom leased an abandoned service station to create “Musket Station.”

• The First Pivot (1970s): Tom boldly introduced Self-Service Gasoline. This reduced labor costs and allowed him to sell fuel at competitive prices, quickly capturing market share in price-sensitive rural areas.

• The C-Store Model: He opened a site in Guymon, Oklahoma, that combined self-service gas with a “Mini Stop.” This was the prototype for the modern Convenience Store model: use low-margin fuel as a Loss Leader (traffic driver) to convert customers into high-margin retail buyers (cigarettes, drinks, food).

2. Riding the Macro Tailwind

Two major events in 1980 supercharged their growth:

1. The completion of the Interstate Highway System.

2. The Motor Carrier Act of 1980, which deregulated the trucking industry.

This created an explosion in long-haul trucking. In 1981, Love’s opened its first true “Travel Stop” in Amarillo, Texas. They shifted their target persona from “general drivers” to “professional truckers.”

• Service Upgrade: 24/7 operations, showers, and massive truck parking lots.

• The Result: The company evolved from a “street-side convenience store” to a critical “transportation infrastructure provider.”

Today, they operate 640+ locations across 42 states and are a fixture on the Forbes list of America’s Largest Private Companies (2023 Revenue est. $45B - $60B, depending on oil prices).

3. The Moat: Vertical Integration & Closed-Loop Ecosystem

Love’s isn’t just selling gas; they are capturing value at every link in the chain. This is a classic “Closed-Loop Ecosystem”:

• Musket Corporation (Trading): Their commodities trading and logistics arm. They handle fuel procurement, hedging, and wholesale, maximizing the Fuel Margin.

• Gemini Motor Transport (Logistics): Their own fleet. They transport fuel from terminals to stations, ensuring supply chain stability and lowering logistics costs.

• Speedco (Maintenance): On-site truck repair and tire services.

• Factoring (Fintech): Financial services for trucking companies.

Furthermore, they have a Real Estate Moat. Acquiring large plots of land right off major highway exits (necessary for truck turning radiuses) involves high regulatory hurdles. It is incredibly difficult for new entrants to replicate their footprint.

The industry has three main players: Pilot Company (majority-owned by Berkshire Hathaway), Love’s, and TravelCenters of America (TA).

What fascinates me most is that Tom Love repeatedly rejected Wall Street’s IPO offers. He understood a fundamental truth that plagues many public tech companies today: Quarterly earnings pressure kills long-term vision.

By staying private, Love’s avoided the pressure to cut CapEx (Capital Expenditures) to please analysts. This allowed them to invest counter-cyclically—buying land and expanding when oil prices crashed or the economy dipped.

Takeaway

Tom Love’s journey—from an abandoned station to a multi-billion dollar empire—is a lesson in cash flow and customer retention. Every step he took addressed a core pain point:

“How do I get the people passing by to stop longer and spend more on my property?”

Whether you are building a SaaS platform or a hardware startup, the question remains the same.

How to Apply Tom Love’s Strategy, Spirit, and Opportunity Discovery in the Modern Era

Tom Love’s success was not built on “inventing new technologies,” but on “restructuring existing needs” and “extreme operational discipline.” In an era dominated by AI and digitization, studying his strategy is arguably even more valuable. He demonstrated how to mine immense profits from “unsexy” traditional industries by solving fundamental inefficiencies.

1. Opportunity Discovery: Identifying “Fragmentation & Friction”

Tom Love’s core insight was identifying a disconnect: fueling and shopping were separate activities (inefficiency), and the basic needs of truck drivers were being ignored.

How to apply this today:

• Seek out “Forgotten Infrastructure”:

  • Tom Love’s Era: Gas stations were dirty, and no one cared about the driver’s experience.

  • Modern Era: Observe industries that are “essential but offer a terrible experience.” Examples include elderly care, blue-collar recruiting, waste management, and traditional logistics warehousing. These sectors are often low on digitization and high on fragmentation—ripe for consolidation and optimization.

• Practice the “Bundling Strategy”:

  • The Love’s Model: Bundling “Fuel + Snacks + Showers + Parking.”

  • Modern Application: Consider what else your target audience needs immediately before or after completing a task.

    • Example: Modern gyms are no longer just about equipment; they bundle physical therapy, nutrition subscriptions, and social co-working spaces.

• Focus on “Humanity in B2B”:

  • The Insight: Most saw truckers merely as “capacity”; Tom Love treated them as “people” (who need hot showers and hot food).

  • Modern Application: In SaaS or Enterprise Services, look beyond features to “emotional value.” Making a user’s workflow less stressful and more dignified is a powerful competitive moat.

2. Strategic Execution: From MVP to Ecosystem

Love’s was not built in a day. It started with leasing a single abandoned station.

How to apply this today:

• Low-Cost Trial ( The MVP Mindset):

  • The Precedent: With only $5,000, Tom Love didn’t buy land; he leased an abandoned site.

  • Modern Application: Do not start with heavy CapEx or massive app development. Use existing tools (No-code, existing platforms) to test the business model. Only deploy significant capital once you have validated that “someone is willing to pay.”

• The Courage for Vertical Integration:

  • The Precedent: Once scaled, Love’s integrated upstream (Musket for trading, Gemini for transport).

  • Modern Application: Many startups fail by relying too heavily on third-party platforms (e.g., Facebook Ads or Amazon Logistics). Learn from Love’s: once your core business is stable, gradually control your supply chain or own your data. That is where the margin lies.

• Asset Duality:

  • The Precedent: Love’s appears to be retail, but fundamentally, it is a real estate business (controlling highway exits).

  • Modern Application: Consider if your business can accumulate “Digital Assets” or “Network Effects.” Build a community (digital real estate), not just a transactional product.

3. Entrepreneurial Spirit: Servant Leadership & “On-Site” Culture

Tom Love famously said: “We sell what everyone else sells (fuel, cola). Why should they choose us? Because of Clean Places and Friendly Faces.”

How to apply this today:

• The “Clean Toilet” Philosophy:

  • The Habit: Even as a billionaire, Tom Love would inspect the restrooms during site visits. This represents an obsession with detail.

  • Modern Application: Whether in software or hardware, your “toilet” is your customer support response time, your UI/UX usability, and your refund process. These unglamorous details determine retention.

• Contrarian Investing:

  • The Habit: The Love family historically purchased land during economic recessions when it was cheapest.

  • Modern Application: When the market is chasing a bubble (like a specific AI hype cycle), remain calm and preserve cash. When the market panics and crashes, that is your moment to acquire premium assets (talent, technology, user base).

Life

For Thanksgiving, we’re spending the week in Denver with friends. We chose to road trip from Austin to Denver, driving from the vast, almost barren plains of Texas to the hilly terrain of Colorado. What I’m most curious about is the people who live here—I wonder how they spend a typical, ordinary day.

Calling all locals! 📣 If you were raised here, please share your top recommendations for the most interesting and must-do things in Colorado.

Market:

GOOG Releases Gemini 3.0 Pro – Our Thoughts

From a technical standpoint, we believe the core success of Gemini 3.0 Pro lies in the extreme execution of Scaling Laws and significant innovations in the underlying model architecture.

Gemini 3 implements Model-Native Multimodality, mixing data across different modalities (text, image, video) during pre-training. Its reasoning mechanism adopts a Tree of Thoughts (ToT) architecture, allowing the model to process multiple lines of reasoning in parallel internally. It validates and selects these paths through a self-rewarding mechanism, significantly improving accuracy. Furthermore, the model utilizes an engineering wrapper for Environment Engineering to provide feedback validation, continuously breaking through performance bottlenecks. This is where we see a distinct technical divergence from previous models.

Additionally, for developers, Google launched the Antigravity IDE. Leveraging Gemini 3 Pro’s Browser Use capabilities—which scored highest in Screenshot Pro benchmarks—it achieves fully automated frontend development and testing.

However, testing with friends revealed that Gemini 3’s performance in Real World Visual Understanding (such as security camera scene analysis) has actually declined. Simultaneously, feedback suggests that for complex tasks involving repeated multi-hop external searches and information synthesis, its performance may still lag behind ChatGPT Pro.

We believe GOOG retains an economic advantage in hardware. The key to Google’s successful execution of Scaling Laws is the integration of its TPU software and hardware. Compared to companies lacking proprietary hardware capabilities—who must absorb NVIDIA’s high profit margins—Google possesses a distinct economic advantage in executing large-scale Scaling Laws.

This release has strengthened my confidence in GOOG and its supply chain, including CLS, LITE, and Taiwan-based companies like Guang Sheng (FOIT).

The Netscape Déjà Vu

The launch of Gemini 3 triggered a sudden realization. We’ve seen this movie before. The precedent has been staring us in the face the whole time: Netscape.

For those who might be too young to remember, Netscape was the OpenAI of the 90s. They built the first browser that truly took off, single-handedly ushering the world into the internet age. They were powerful, adored, and looked invincible.

But we know how that story ended.

Netscape proved that humans had a massive demand to “go online.” But then Microsoft—the incumbent giant—woke up. They bundled Internet Explorer into Windows for free. It was the kiss of death. Despite Netscape’s IPO war chest and desperate attempts to build a moat, they couldn’t compete with “free” and “default.”

Today, the board is set up exactly the same way.

• OpenAI is Netscape: The pioneer who opened the door.

• Google is Microsoft: The giant with the unfair distribution advantage.

Google isn’t just a software company; they possess an economic advantage that OpenAI cannot match. They have self-developed TPUs (meaning rock-bottom inference costs compared to paying the “NVIDIA tax”), billions of users locked into Gmail and Maps, and a massive Cloud infrastructure for enterprise.

Just as Microsoft used Windows to win the browser war, Google is now bundling Gemini into its ubiquitous ecosystem to squeeze OpenAI out.

The Momentum Shift Crucially, the one thing keeping OpenAI alive—ChatGPT’s superior performance—is slipping. With Gemini 3 now pulling ahead in benchmarks, investors have little reason to keep buying OpenAI’s pitch that they just need “more money to reach AGI first.”

So, is AI a bubble? If you look at history, the answer isn’t Yes or No—it’s “Yes, but...”

The Dotcom bubble popped, and companies vanished (RIP Netscape), but the internet infrastructure remained. It paved the way for today’s Big Tech.

We are likely heading toward a similar endgame:

1. The Pioneer Bubble: OpenAI, the frontrunner, risks popping or fading away like Netscape.

2. The Infrastructure Boom: The massive capex spending is real. We are building the rails for the next generation of intelligence.

This means hardware and infrastructure companies still have a long runway of good times ahead. That part isn’t a bubble.

OpenAI kicked off a war they might not finish. They cracked open the door to a new world, but it looks like Google is the one rushing in to conquer the territory.

Whatever happens, it’s been a Good Game. But right now, those two G’s stand for Google.

Insightology View : Rethinking the ROI Panic

I recently watched Nana Banana Pro successfully generate Traditional Chinese text for the first time—instantly creating finished product ads. Between this and the release of Gemini 3 Pro, we aren’t just seeing an update; we are witnessing a massive generational leap.

This progress has forced me to re-evaluate the skepticism that has dominated the market over the last two weeks.

Investors have been stressing over ROI, depreciation, and the path to monetization. The fear was that revenue wasn’t taking off. But let’s look at the root cause:

• Revenue was flat because applications were stagnant.

• Applications were stagnant because the base models weren’t good enough.

There was a capability bottleneck.

Now that we see base models smashing through those ceilings, the bottleneck is clearing. Better foundational models will unlock new software capabilities and use cases, which in turn will drive the revenue growth everyone has been waiting for. The anxiety surrounding ROI and monetization is likely to alleviate as these new capabilities hit the market.

The takeaway: We are still in the rapid-growth phase of foundational models. It is a mistake to let today’s restrictive financial models limit our imagination for what’s coming next.

Book : Rewire: Break the Cycle, Alter Your Thoughts and Create Lasting Change

In the process of researching AI, I feel like I am researching how the human brain works. This week, I finished reading Rewire, and it gave me a lot of inspiration.

In the AI era, software tools have become more convenient and the barriers lower. Building a website, writing an article, creating a service, or building a brand is incredibly fast today. However, I want to reverse the focus and understand more deeply how our brains function, and how to use AI to achieve a better balance in life and work.

Especially as new AI tools are released daily, I feel that while absorbing this information, my brain is slightly overworked. This fueled my curiosity about how to effectively manage emotions, brainpower, and the body. Starting from within, once the core is stable, I believe we can handle external information changes with more ease.

The book mentions several exercises I’ve been trying recently:

• The Physiological Sigh: This is a breathing technique to quickly regulate the nervous system, known as “cyclic sighing.” It artificially accelerates the process of returning to calmness. It has been proven to be the most effective way to immediately return you to a parasympathetic state (rest and digest). The Method: Inhale twice consecutively, then exhale slowly. The double inhale is crucial because it forces collapsed alveoli (air sacs) to pop open and re-inflate. This effectively removes excess carbon dioxide from the system, as high CO2 levels cause feelings of stress and agitation. Essentially, it sends a safety signal to the brain, telling the body, “I am no longer facing any threat.” I have physically felt my heart rate drop and return to a relaxed state.

• Deliberate Disconnection: Avoiding the internet specifically at the beginning and end of the day. The book cites this as a key strategy to protect “brain hardware” and avoid cognitive overload. Now, for the first 30 minutes after waking up and the last hour before bed, I don’t use my phone. I’ve found that not replying to messages instantly or checking social media hasn’t impacted my life negatively; in fact, I feel better, and I work more efficiently. When people are tired and mentally drained, they rely on mental shortcuts, regressing to automatic and rigid states. Strategic Rest (like staying away from screens) allows the brain to recharge, avoiding the old patterns of rumination, binge-eating, or doom-scrolling.

• Journaling Before Bed: I used to think journaling was a waste of time. But since swapping scrolling time for journaling and reading, my anxiety regarding information has actually decreased. The book notes that journaling helps reduce emotional burdens and the frequency of ruminating thoughts. When journaling, we must use complete sentences, which differs from the disorganized state of rumination. Writing guides thoughts toward something more cohesive, organizing them so they don’t run wild without a beginning or end. Activating the Logical Brain: Journaling activates the areas of the brain related to speech generation, requiring the use of the frontal cortex (the logical thinking part). This mobilizes the rational part of the brain, preventing the emotional limbic brain from taking charge. Externalizing and Managing Thoughts: Through writing, we externalize ruminating thoughts, making them manageable. This allows you to face your thoughts and feelings in a controlled environment, lightening the emotional load.

These are simple things, but repeating them consistently every day is not easy—breathing, sleeping, writing. Although difficult, they are far more important than researching new tech or following KOLs every day.-Insightology Research

What I Read

https://simonwillison.net/2025/Nov/18/gemini-3/

One Useful Thing

Three Years from GPT-3 to Gemini 3

I’ve been testing Google’s new Gemini 3 model. It is very good. But rather than give you benchmark numbers, I’m going to show you how far AI has come in three years by asking the AI to demonstrate it…

Read more

9 months ago · 576 likes · 39 comments · Ethan Mollick

Transcript

• Nvidia Earnings Call Transcript

• KLIC Earnings Call Transcript

Books

• Rewire: Break the Cycle, Alter Your Thoughts and Create Lasting Change

What I learn to build

• https://github.com/omnivore-app

• https://github.com/karpathy/reader3?tab=readme-ov-file

Life

In Buffett’s last shareholder letter, the sentence that touched me the most was: “Keep in mind that the cleaning lady is as much a human being as the Chairman.” Having been in the workplace for a long time, I often need to remind myself to have empathy. When you have empathy, kindness, respect, and gratitude, your perspective on things will gradually change, and these are things you can’t buy with all the money, power, or fame.

In the past, I sometimes focused too much on profit and loss at work, often overlooking the things around me. I was also often reminded by my wife, but frequently caring for the needs of those around me and helping others is more important than performance.

I was late in becoming old – its onset materially varies – but once it appears, it is not to be denied. To my surprise, I generally feel good. Though I move slowly and read with increasing difficulty, I am at the office five days a week where I work with wonderful people. Occasionally, I get a useful idea or am approached with an offer we might not otherwise have received. Because of Berkshire’s size and because of market levels, ideas are few – but not zero.

Yes, even the jerks; it’s never too late to change. Remember to thank America for maximizing your opportunities. But it is – inevitably – capricious and sometimes venal in distributing its rewards. Choose your heroes very carefully and then emulate them. You will never be perfect, but you can always be better. ”

Buffett is really a workaholic, but despite being 95 years old, he can still work happily and enjoy it, which is truly amazing. Thank you, Buffett, you are my hero in investing, business, and how to treat people. I wish I could be like him too.

I’m happy to say I feel better about the second half of my life than the first. My advice: Don’t beat yourself up over past mistakes – learn at least a little from them and move on. It is never too late to improve. Get the right heroes and copy them.

Warren Buffett's final letter to shareholders

Warren Buffett’s 1999 speech at Sun Valley

Looking at history, I think it’s worth revisiting Warren Buffett’s 1999 speech at Sun Valley, which was compiled by Fortune into the article “Mr. Buffett on the Stock Market.” It mentioned several points that we are paying special attention to in the current industry cycle:

• New technology doesn’t automatically equal high investment returns. Long-term returns must revert to first principles: Long-term return = Dividend Yield + Earnings Growth ± Change in Valuation.

• He used the auto and airline industries as examples: The automobile was one of the most important inventions of the first half of the 20th century, but of the hundreds of car manufacturers at the time, only a few survived. Investors as a whole lost heavily. Aviation changed the world, but airlines also had poor long-term returns.

• Excellent industry prospects do not guarantee good shareholder returns; competition, capital expenditures, excessive stock issuance, bankruptcies, etc., can erode economic value. (This is somewhat similar to the present, though today’s situation is a bit more complex.)

• The cash flows of the tech giants investing in AI infrastructure today are different from those in 1999. From a cash flow valuation perspective, we feel they are relatively safer. However, we believe the “debt” aspect might be a long-term risk to watch.

• AI infrastructure will continue to change the world. Besides investing, I think we are more interested in using these tools to experiment with some things. (We will announce this part later.)

• Summary: We agree that AI, like the previous internet revolution and great inventions such as the automobile, airplane, and radio, will change the world.

  • Even a profound industrial revolution does not guarantee huge returns for early investors.

  • Using the auto and aircraft manufacturing industries as examples, he pointed out that although these industries had enormous growth potential early on, only a few companies ultimately survived.

  • In 1999, despite the huge impact of the internet, investors still needed to carefully pick winners and couldn’t rely solely on macro trends. Many companies that led early in the revolution eventually failed.

Market

There were also several company earnings calls and technical articles this week.

PCB Upstream

Insightology View: Upstream fiberglass cloth continues to benefit from strong AI demand. However, we are closely monitoring capacity expansion timelines. Historically, such expansions tend to slow the price growth of upstream materials. For investors, slowing price growth implies that the potential upside for P/E ratios may be limited.

In the current AI materials market, there are shortages not only in memory but also in substrates. The substrate shortage is driven by upstream materials, including T-glass fiberglass cloth and copper foil. The primary producer of T-glass is Japan’s Nittobo (TYO: 3110), and the main supplier of copper foil is Mitsui Kinzoku.

Major Japanese tech-related companies Nittobo (3110.T) and Mitsui Kinzoku (5706.T) recently released their latest financial reports (for the period ending September 2025). Their results were contrasting: one company’s profit forecast surged, while the other’s net profit appeared to be halved. But the devil is in the details. Let’s break it down:

Nittobo (3110.T)

• Operating profit for the first half increased by 29%. The company significantly raised its “full-year net profit” forecast, now expecting an increase of +192.1%! Its products can be broken down into two main growth drivers:

• T-Glass

  • Introduction: T-Glass is a type of “foundation” that carries the chip (a key material for IC substrates). It ensures the chip does not deform under extreme heat or cold, maintaining absolute stability.

  • As AI chips become more powerful and are stacked in multiple layers (e.g., advanced packaging like CoWoS), they require a more robust and stable “foundation.” T-Glass has consequently shifted from being an “optional component” to a “standard requirement,” with demand 100% tied to the advanced packaging needs of AI chips.

• NE-Glass

  • Introduction: NE-Glass is a glass cloth that allows electronic signals to “run faster with less interference.” AI servers transmit massive amounts of data internally. If standard materials are used, signals lose significant energy and speed, and are prone to “crosstalk” (interference).

  • The job of NE-Glass is to serve as the material for AI server motherboards (PCBs) and switches, ensuring the vast data required for AI computing can be transmitted at high speed and losslessly.

  • Transmission Speed Upgrades: Upgrading from PCIe 4.0 to 5.0 and 6.0, and from 400G switches to 800G and 1.6T, requires the use of lower dielectric (Low-Dk) NE-Glass materials at every step.

  • Increase in AI Server Penetration: Traditional servers can use lower-grade materials, but AI servers must use high-speed materials of NE-Glass’s caliber. The greater the shipment volume of AI servers, the stronger the demand for NE-Glass.

  • Oligopolistic Market: Likewise, this is an oligopolistic sector where Nittobo commands key technology.

Nittobo’s Strategy: “Invest heavily in capacity expansion.”

• Specific Actions: The company has announced an investment of tens of billions of yen (recent reports suggest ~15 billion yen) for a large-scale expansion, primarily focused on its factory in Taiwan.

• Timeline: New capacity is expected to begin ramping up at the end of 2026 and be fully released in 2027-2028. The goal is to triple production capacity to meet the explosive demand “strongly forecasted by customers” over the next few years.

Mitsui Kinzoku (5706.T)

• Mitsui Kinzoku’s latest capacity expansion plans for its two major AI copper foil product lines:

• VSP™ (HVLP Copper Foil) - Used for AI Server Motherboards

  • Announced an “additional capacity expansion” for VSP™ (HVLP Copper Foil) used in AI servers, indicating that end-user demand has far exceeded the company’s previous expectations.

  • Additional Investment: Announced an additional investment of approximately 6 billion yen.

  • Expansion Locations: Taiwan factory (Taiwan Copper Foil Co., Ltd.) and Malaysia factory.

  • New Target: To further significantly increase the total monthly production capacity of VSP™ from 840 tons in September 2026 to 1,200 tons, expected to be completed by the end of September 2028.

• Microthin™ (Copper Foil with Carrier) - Used for IC Substrates

  • This is Mitsui Kinzoku’s flagship product with the highest technical barrier and profitability, used in advanced packaging such as the ABF substrates required for AI chips (GPUs).

  • Current Plan (Announced January 2025):

    • Expansion Locations: Ageo Plant (Saitama Prefecture), Japan, and the Malaysia factory.

    • Target: Plans to gradually increase production capacity, aiming to raise monthly capacity from 4.9 million square meters at the beginning of 2025 to 5.6 million square meters.

    • Timeline: According to the previous plan, the target is to be achieved between 2027 and 2030.

    • Insightology View: Although the expansion plan for Microthin™ isn’t being “topped up” as frequently as VSP™, it remains a stable focus of the company’s capital expenditure. In the latest earnings call, the company confirmed that demand for Microthin™ remains “solid” and is another major pillar of AI demand.

AMD Analyst Day

• Insightology View

  • Actually, most of the technology roadmaps discussed at the analyst meeting were things we were somewhat aware of from our research. What surprised me more was the increase in the total addressable market (TAM) to $1 trillion.

  • And the idea that “compute equals intelligence.” If intelligence is infinite, then the demand for compute is infinite. I agree with this view, but growth will still have cycles, though the wavelength of this cycle might be longer than previous ones.

  • The data center compound annual growth rate (CAGR) will reach 60% over the next 3 to 5 years. (This estimate seems a bit high, which makes us slightly apprehensive, as it implies some years will be exceptionally good, while others might be slightly weaker.)

• CEO Lisa Su pointed out at the meeting that the growth in AI infrastructure demand will not level off but will remain strong. Customer investment in AI is not “stabilizing.” AMD expects its sales growth to accelerate over the next five years, especially in: AI inference.

• A Few Numbers

  • Data center CAGR of 60% over the next 3 to 5 years.

  • Core business CAGR will exceed 10%.

  • Overall revenue annual growth rate will exceed 35%.

  • Operating profit margin will exceed 35%.

  • Earnings per share (EPS) will exceed $20, higher than market analyst expectations.

  • The total data center market size will reach $1 trillion by 2030.

• Partnerships

  • OpenAI: Announced a five-year, 6 GW-scale strategic partnership. AMD has optimized the MI450 series for many of OpenAI’s use cases, including optimizing inference.

  • Oracle: Expanded their partnership. Oracle will be one of the first companies to offer public instances of the MI450, expected to scale starting in Q3 2026.

  • Meta: The partnership has expanded to include the MI300 and MI350, with deep collaboration on open standards for the Helios rack-scale solution.

  • U.S. Department of Energy (DOE): Will adopt features from the MI400 series, including the new MI430 product, in new supercomputers.

GB300 NVL72 MLCommons MLPerf Test Report

• It’s no surprise and widely expected that it won across all model benchmarks. We also noted a few other points:

• Low data precision can accelerate model computation and reduce memory requirements. Roughly speaking, FP8 is twice as fast as BF16, and FP4 is twice as fast as FP8. However, this comes with the drawback of reduced accuracy. Therefore, model pre-training is currently quantized to FP8 at most, and model inference to FP4 at most.

  • First, an explanation of data precision:

    • Simply put, FP4 and FP8 refer to “Floating-Point” (FP) data formats. The numbers 4 and 8 mean the format uses only 4 or 8 “bits” to store a single number.

    • They are “low-precision” formats, primarily aimed at achieving extreme performance, reducing memory footprint, and lowering power consumption in AI computing.

  • Why use FP4 and FP8?

    • As AI models (especially Large Language Models, LLMs) become larger, using FP32 or FP16 encounters two main bottlenecks:

    • Memory Wall: The massive model parameters (weights) cannot all fit into the GPU’s memory (VRAM), leading to the need for expensive hardware or extremely slow data swapping.

    • Compute Speed: Processing 32-bit or 16-bit operations ultimately has its physical limits.

  • NVIDIA’s improved variant 4-bit precision quantization format, NVFP4, performs better than the industry-standard FP4 and the variant MXFP4. Previously used for inference, this has now been confirmed for “pre-training” as well.

  • NVIDIA used NVFP4 for model “pre-training” for the first time, which was validated by MLPerf (previously announced on the company’s official blog). Accuracy only drops significantly after 0.8T tokens. This is likely the first large-scale model demonstration of FP4 pre-training, which is a breakthrough. Using 5,000 GB200s, pre-training for Llama 3.1 405B takes only 10 minutes.

  • The above refers to pre-training, not inference. As for inference, there is no doubt FP4 will become mainstream, as its effectiveness is on par with FP8.

• We agree that Google’s massive-scale TPU pods (over 9,000 TPUs, though internal communication details are unclear) are a significant advantage. However, I am personally skeptical of the argument that TPUs offer a better TCO (Total Cost of Ownership) for AI inference. This is because, with the advancement of FP4 inference technology and its widespread adoption this year and next, TPU hardware still only supports up to FP8, with no native FP4 hardware support. This means it must compete with NVIDIA Blackwell, which can cut costs by half using FP4.

• The Blackwell Ultra (B300) isn’t just about the NVFP4 innovation. In terms of hardware, compared to the B200, its circuit design sacrifices FP64 to boost FP4 capabilities by 50%. Therefore, it has:

  • (1) NVFP4 data precision + (2) 50% more hardware FP4 circuitry than the Blackwell (B200).

• We are currently skeptical as to whether Google TPU can achieve a cost advantage in the AI inference market (vs. the GB300 NVL72).

  

What I Read

• Nvidia Showcases Blackwell Ultra Performance on MLPerf Benchmark

• NVIDIA Wins Every MLPerf Training v5.1 Benchmark

• https://x.com/SemiAnalysis_/status/1989369981625098262?s=20

• A16z：Jevons or Bust

Transcript

• CoreWeave Earnings Call Transcript

• Nebius Earnings Call Transcript

• AMAT Earnings Call Transcript

Books

• Rewire: Break the Cycle, Alter Your Thoughts and Create Lasting Change (2/3)

What I Enjoy

First, congratulations to the Dodgers on winning the World Series, and congratulations to Yamamoto Yoshinobu for winning the MVP this time. Because of this game, I went back and watched Yamamoto’s story. He can be described as a late-bloomer type of player. In the past, in high school and in the pros, he was not as formidable as Shohei Ohtani. He did not make it into Koshien in high school, nor was he the ace pitcher, and he only transitioned to pitching in high school.

After joining professional baseball, he was troubled by an elbow injury in his first year. Even after resting for ten days, the tightness in his elbow could not be eliminated. This made him realize that if his physical condition did not improve, he would not be able to consistently perform in professional games.

To break through the predicament, he began seeking new training methods and met trainer Osamu Yata. He thoroughly changed his pitching form and training philosophy (https://www.dailyshincho.jp/article/2025/10300602/?all=1&page=2), and even temporarily “sealed off” his signature slider to reduce the burden on his elbow. After the change, Yamamoto Yoshinobu was completely reborn, winning the Sawamura Award, MVP, and the pitching quadruple crown for three consecutive years during his time with the Orix team, becoming the strongest ace in Japanese baseball.

Quoting Mr. Osamu Yata: A tree has a trunk, branches, and roots. The movements of your hands and feet are like fiddling with the branches. The trunk is certainly important, but it cannot just be stiff. If the trunk is hollow, it will easily break. Therefore, what is truly important is building a sturdy trunk that can withstand strong winds and storms.

At the end of the article, I will also share my reflections on this.

Market: More Focus on the Optical Communication Industry

Optical Communication Industry Tracking: AI Needs Light! Optical Modules Benefit from Strong 1.6T Demand; 2026 is a Year of Sustained Explosion

First, this week I mainly focused on the latest quarterly conference calls of these two companies, Lumentum and Coherent.

Why I am researching them:

• If NVIDIA’s GPUs are the brain of AI, then what Lumentum and Coherent provide is the ultra-high-speed nervous system. When studying AI, a major trend is discovered: AI computing has shifted from solo efforts (one GPU calculating slowly) to group operations (thousands of GPUs calculating together). And Lumentum and Coherent are the most crucial arms dealers for this group operation.

• In terms of transmission specifications, the current development is reaching 800G and 1.6T specifications. What does this mean? It means transmission speeds of 800 Gigabits and 1.6 Terabits per second. Furthermore, they focus more on transmission between racks, responsible for long-distance transmission (over 7 meters). In contrast, Credo (also a company with excellent technology), which we often hear about, mainly focuses on connection cables within the rack.

• Long-distance transmission must use Coherent/Lumentum’s “optical modules,” which are expensive (lasers are expensive), and consume power (lasers generate heat and require cooling), but they have almost no loss. Therefore, this optical module is what converts the “electrical signal” inside the rack into the “optical signal” between the racks.

A brief introduction to the businesses of these two companies and their positions in the supply chain:

• Lumentum: “Optical communication components + module supplier”. Primarily produces cloud and telecom optical communication components + modules (including data center optical modules after the Cloud Light merger), Optical Circuit Switches (OCS), ROADM, tunable lasers, EML/DFB/VCSEL, etc.. Also has industrial and 3D sensing lasers (Industrial Tech).

• Coherent: “Full-line manufacturer from materials to optical modules”. Mainly plays the role of “upstream” and “midstream”. Revenue sources include Networking (data center/telecom optical transceiver modules and components, such as 400G/800G/1.6T, 100/400/800G ZR/ZR+, etc.), Materials (SiC, GaAs/ InP and other compound semiconductors and engineered materials), and Lasers (industrial, display, and scientific lasers).

Source：Yole

LITE not only has outstanding performance, but 1.6T is a huge explosion, starting a furious ramp-up in 2026

• Lumentum achieved the highest revenue in the company’s ten-year history this quarter, with total revenue reaching $533.8 million. Revenue growth exceeded 58%. The company estimates that over 60% of total revenue comes from cloud and AI infrastructure. Lumentum expects to reach the quarterly revenue milestone of $600 million two quarters earlier, in the second quarter (Q2) of fiscal year 2026.

• The company has changed its financial reporting to a single reporting segment and categorized product line revenue into two types: Components and Systems.

• Viewed by business segment:

  • Components Revenue: $379 million, QoQ +18%, YoY +64%, driven by data center and long-haul applications.

  • Systems Revenue: $155 million, QoQ -4%, YoY +47%, mainly from cloud transceiver modules.

  • EML laser shipments hit a new high, predominantly 100G, with 200G starting to scale up, expected to account for 10% by early 2026.

  • DCI narrow linewidth laser components grew for seven consecutive quarters, YoY +70%.

  • 3D sensing products account for less than 5%, showing seasonal growth.

• Major Growth Engines (AI Three Arrows): The company explicitly points out that growth over the next few years will be driven by three core product lines that together form the backbone of AI infrastructure:

  • Cloud Transceivers: This business segment is expected to accelerate growth over the next 4 to 5 quarters. The company aims to push this market opportunity to a quarterly revenue of $250 million.

  • Optical Circuit Switches (OCS): This product is used to reconfigure data center network paths. Lumentum’s confidence in this product has increased significantly, expecting OCS quarterly revenue to reach $100 million by the December quarter of 2026.

  • Co-Packaged Optics (CPO): Demand for CPO is stronger than the company initially anticipated, with substantial revenue contribution expected to begin in the second half of 2026.

  • ML Laser Shortage (Indium Phosphide): Market demand for its core Indium Phosphide (InP) laser chip (the critical light source for high-speed fiber networks) far exceeds its supply capacity. This forces the company to make difficult capacity allocation decisions, prioritizing resources for the highest-margin components, such as 200 gig MLs.

  • Capacity Expansion Strategy: Although demand is strong, the company commits to increasing unit capacity by approximately 40% in the coming quarters by improving yield and output. The focus is mainly on four-inch wafers, rather than large-scale capital expenditure expansion. Also, due to the severe imbalance between EML laser supply and demand, Lumentum is leveraging this advantage to implement targeted price increases to boost gross margin.

• 1.6T Transceivers have better margins: Management expects gross margin to continue improving in 2026 with the increasing proportion of 200 gig MLs, the launch of 1.6T transceivers, and the maturation of high-margin new product lines like OCS and CPO. Specifically, the profit margin for 1.6T transceivers will be “significantly better than 800G products”.

• OCS Mega Storyline: Targeting $100 million/quarter in 2026. The official timeline is clear: several million dollars in 2025/12, tens of millions (M) in 2026/3, $50–60M in 2026/6, $100M in 2026/12, and long-term $250M/quarter. Three customers are running qualifications, and software is still being supplemented.

• CW vs EML is not a forced choice; both will rise. Capacity can be interchanged, and no matter how the market moves, everything will sell out. This is called true supply chain flexibility + a seller’s market. Currently, LITE is also the exclusive supplier of Nvidia CW Laser.

• CPO Interest Surge: Customer and market interest and participation in CPO have increased significantly. The company ensures its Indium Phosphide capacity can be flexibly switched so that Lumentum can provide services regardless of which laser technology CPO ultimately adopts (ML or CW).

• Personal View: Lumentum is essentially saying: demand is too strong, supply is too little, everything is out of stock, everything has long-term contracts, and prices will still rise in 2026. AI is truly exploding the entire optical communication supply chain.

• What is impressive is the company’s OCS product, which allows data to be dynamically switched directly in the form of light, without converting to electrical signals. It can support tens of thousands of GPU clusters and save up to 65% of power compared to traditional switching methods.

COHR: 1.6T starting to ramp up next year, Upstream InP is key

First, a review of the company’s performance:

• Q1 2026 revenue was $1.58 billion, QoQ +3%, YoY +17%. Excluding the divested aerospace and defense business, revenue grew QoQ 6%, YoY 19%. Non-GAAP EPS was $1.16, QoQ +16%, YoY +73%.

• Q2 2026 Guidance: Revenue estimated at $1.56-$1.7 billion, Non-GAAP gross margin 38-40%, Non-GAAP EPS $1.10-$1.30.

Viewed by business segment:

• Datacenter & Communications: Q1 revenue QoQ +7%, YoY +26%; Data center business QoQ +4%, YoY +23%; Q2 guidance QoQ growth of about 10%.

• Communications (including DCI and traditional telecom): Q1 revenue QoQ +11%, YoY +55%, continuous growth for five quarters, Q2 expected to continue increasing.

• Industrial: Q1 revenue QoQ +2%, YoY +4% (excluding aerospace and defense), Q2 expected to be flat or slightly increased.

• Q1 Non-GAAP gross margin 38.7%, QoQ +70bps, YoY +200bps.

• Q1 Non-GAAP operating expenses $304 million, accounting for 19.2% of revenue, down from the same period last year.

• The key point for Coherent this quarter is: AI data center and communication demand is so strong that the supply chain is still catching up. The overall strategy—from mass production, technology, capacity, and gross margin—is accelerating.

• In terms of growth momentum, the Data Center and Communications segments are the largest and fastest-growing areas. Q1 data center revenue YoY +23%, Communications YoY +55%. The company explicitly pointed out that orders for 800 G and 1.6 T transceivers are “record-breaking,” and the adoption of 1.6 T is accelerating.

• Capacity and Deployment: With Indium Phosphide (InP) lasers being the capacity bottleneck, the company initiated a 6-inch InP production line, while simultaneously increasing capacity for EML, CW lasers, and photodiodes. The 6-inch InP production line in Sherman, Texas, USA, has started mass production, and the yield is already better than the old 3-inch line. A second 6-inch InP production line in Järfalla, Sweden, has been simultaneously started, and internal capacity is expected to double within the next 12 months. For module assembly, in addition to the existing Ipoh, Malaysia factory, the new Penang factory has started operation, and the existing component factory in Vietnam will also be upgraded to join module production.

• Compared to Lumentum, Coherent is actually a consolidation of three giants (II-VI, Finisar, and Coherent itself). For the 6-inch InP technology, the real contributor is Finisar:

  • Jumping from 3-inch to 6-inch wafers presents an exponential increase in difficulty. It is not as simple as making a pizza bigger.

  • Material Properties: Indium Phosphide (InP) is a compound semiconductor that is extremely fragile (unlike Silicon, which is sturdy).

  • Manufacturing Difficulties: The larger the wafer size, the harder it is to maintain “uniformity” and a “low defect rate” during crystal growth.

  • Finisar’s Big Bet: Years ago, when the rest of the industry was content with 3-inch or 4-inch lines, Finisar invested huge R&D funds and resources to overcome the manufacturing challenges of 6-inch InP.

• Through the acquisition of Finisar, the current Coherent inherited two “ace” factories: Järfälla, Sweden, and Sherman, Texas, USA. These two factories were the main bases for Finisar’s past R&D into 6-inch InP.

• The company mentioned that many orders are for “more than a year later,” indicating that customers are not just pulling goods in the short term but are planning ahead. High technological barrier and difficulty in switching: InP lasers are capacity-limited items, but the company gains a differentiated advantage through the 6-inch upgrade.

• Gross margin improvement is not solely reliant on volume but combines capacity upgrade and price optimization. The company stated that margin improvement comes from pricing optimization + yield improvement + cost reduction.

  

  

AXTI: InP Successfully Obtains Export License, Expected to Usher in Tsunami-like Growth

• Indium Phosphide (InP) hits a new high: InP revenue reached $13.1 million, the highest level since 2022, primarily attributed to the successful acquisition of export licenses.

• Record InP Order Backlog: As of now, the InP order backlog has exceeded $49 million and continues to grow, representing the largest backlog in the company’s history.

• Customers realize that the “just in time” model is no longer applicable, and they are willing to give longer lead times for orders so that AXT has time to submit license applications and ship on time.

• Capacity Planning:

  • AXT’s current InP capacity is about $20 million per quarter.

  • Increasing capacity by 25% only takes about three months, simply by activating more existing furnaces.

  • Doubling capacity (reaching about $40M/quarter) requires about nine months, with an estimated capital expenditure of $10 million to $15 million, utilizing existing facilities and land.

• Future 5-10 fold “tsunami” growth is mainly expected from the large-scale application of Co-Packaged Optics (CPO).

• In terms of customer structure, customer concentration and the number of large customers increased this quarter. The top five customers accounted for 45.2%, and two customers exceeded 10% (in 25Q2, the top five customers accounted for 30.9%; one customer exceeded 10%).

  

Insightology View (AXTI):

• The company estimates they hold at least a 40% share of the Indium Phosphide substrate supply in the global market. Sumitomo, the largest supplier, announced two capacity increases in the past three to four months.

• In the short term, the entire industry is expanding. We believe we must be more cautious about Chinese suppliers. After all, companies like Tianke Heda, Shanghai Simgui, and Crystal Run Semiconductor are continuously improving process yields. Looking back at SiC’s Wolfspeed, we know that being the leader doesn’t prevent you from being bankrupted by competition. Therefore, continuous attention to the future market is necessary.

Insightology View for Optical Communication Industry

• Based on the latest quarterly statements from the entire supply chain, 1.6T will begin to ramp up next year. According to our internal research, two major demand customers come from Nvidia and Google. Nvidia benefits because the yield rate of the GB300 in the supply chain next year is better than expected, and demand from CSP customers is stronger. Google benefits because TPU V7 is expected to start mass production next year, and potential customer Anthropic is expected to purchase TPUs from Google by the end of next year or the year after.

• Overall, we remain optimistic about the development of AI optical communication next year.

Memory Industry Follow Up

SanDisk Significantly Upgrades Exabyte Demand Forecast

• Significantly upgraded the Exabyte demand forecast, from an expected growth of 20% to 40%-50%. (230 EB in 2024, expected high 300EB in 2026).

• It is predicted that calendar year 2026 will be the first time that data center NAND demand surpasses mobile devices, becoming the largest sub-market in the NAND market (structural shift).

• The market is expected to be in a state of short supply. Previously, the shortage was mentioned until the end of 2026, but discussions are now actively initiated with SanDisk regarding supply issues for 2027.

• Customers are starting to sign long-term contracts with the company, whereas contracts were typically only for one quarter in the past.

• The company also provided guidance for overall NAND demand: The company predicts that the current supply capacity of various NAND manufacturers in 2026 (about 17% growth) can only satisfy about 14% of constrained demand growth. However, the actual potential market demand is about 25% growth, leaving a gap of close to half of the potential growth unmet, about 8 percentage points or higher.

Asian Memory Supply Chain: Transcend, Winbond, Nanya Tech, Phison

• Transcend (創見) - November 7 Media Briefing

  • Market Status: Chairman Tsung-Wan Su described this wave of memory shortage driven by AI as the “most severe in 30 years”.

  • Price Outlook: NAND Flash suppliers failed to deliver in October, which is rare, leading to a 50% surge in contract prices in November.

  • DRAM Supply/Demand: DDR4 capacity is difficult to increase further, and supply remains tight; DDR5 has a better chance of opening up new capacity.

  • Q4 Outlook:

    • Memory prices are expected to continue rising due to sustained supply tightness.

    • The company will flexibly adjust inventory configuration to ensure stable supply.

    • Operations will focus on high value-added industrial control and enterprise-level products (DDR4/DDR5).

• Winbond (華邦電) - November 5 Investor Conference

  • Operational Overview: Q3 DRAM segment revenue hit a three-year high, with a gross margin of up to 51%.

  • Q4 Outlook:

    • Revenue, shipments, and average selling price (ASP) are expected to grow simultaneously.

    • Capacity utilization is near full, expected to maintain good profitability.

  • 2026 Outlook: General Manager Pei-Ming Chen expects next year to be “a year of healthy growth”. There are even reports that the GM believes the memory shortage could extend into 2027.

  • DRAM Supply/Demand (Niche):

    • DDR4 demand remains strong (as many customer SoCs still do not support DDR5).

    • After the three major original manufacturers shift to high-end processes, it is difficult for them to return to producing DDR4, causing a “structural shortage,” which benefits Winbond.

  • Flash Supply/Demand: NOR and SLC NAND inventory has been cleared. Demand is robust, and prices and demand are expected to rise simultaneously.

  • Future Planning: The board approved capital expenditure of approximately NT$35.5 billion for expanding production capacity at the Kaohsiung plant (DRAM) and the Taichung plant (Flash).

• Phison (群聯) - November 7 Investor Conference

  • Operational Overview: Q3 financial report was outstanding, with quarterly earnings (EPS) exceeding NT$10 (one capital unit).

  • Long-Term Outlook (NAND): The company is optimistic about the application of AI inference, which will lead to the NAND supply tightness continuing for several years.

  • Q4 Outlook (Market Expectations):

    • Analysts estimate that due to the QLC (Quad-Level Cell Flash Memory) shortage, demand will spill over to TLC NAND.

    • Q4 NAND contract prices are expected to rise by 15% to 20%.

  • 2026 Outlook (Market Expectations): Domestic and foreign institutional investors are optimistic that Phison’s profit might double next year.

• Conclusion

  • Transcend directly pointed out the “most severe shortage in 30 years” and the surge in NAND prices.

  • Winbond emphasized the niche of DDR4’s “structural shortage” and is optimistic that this boom will continue for several years.

  • Phison benefits from rising NAND prices and pointed out that AI inference will be the key driver of NAND demand tightness for the next few years.

Insightology View: HDD demand is more concentrated in AI-benefiting data centers, allowing its growth rate to surpass NAND.

• To estimate the pure AI contribution, we first need to assume a “non-AI” baseline.

• Personal Estimate (Non-AI): Assume traditional/non-AI Server NAND bit demand naturally grows by approximately +20% ~ +25% annually.

• Calculation (AI Contribution): * Total growth in 2026 (+45%) minus Non-AI growth (+20~25%).

• = AI Inference contributes 20~25 percentage points of growth.

• Based on the 2025 Server NAND total volume of approximately 270EB as the base period, this means AI inference will bring an additional 54 ~ 68 EB of NAND bit demand in 2026.

• AI inference does not use 100% SSD (NAND). In the process of LLM inference, the storage configuration is likely hybrid.

• Key Assumption: In the storage configuration for AI inference, approximately 33% uses SSD, while 67% still uses HDD.

HDD Demand Calculation:

• If 54~68 EB of NAND demand accounts for 33%,

• Then the additional demand for HDD (accounting for 67%) will be as high as 108 ~ 136 EB.

• In summary, AI inference will potentially contribute an additional 162 ~ 204 EB of total storage capacity increment in 2026. This will drive the overall growth of two industries, respectively:

  • NAND (SSD) Industry:

    • The AI contribution of 54~68EB demand will boost the NAND overall bit demand growth rate:

    • From the recent +14~16%

    • Up to +20% ~ +23.5%.

  • HDD (Hard Disk Drive) Industry:

    • The AI contribution of 108~136EB demand will boost the HDD overall bit demand growth rate:

    • From the recent +18~22%

    • Up to +24% ~ +29%.

• The short-term growth rate of HDD (+24~29%) is expected to be higher than NAND (+20~23.5%). This seems counterintuitive but is actually very reasonable. The answer lies in the “denominator”: Although NAND is critical, its total demand still comes 70% from the short-term slowing consumer market (PC, mobile phones). This severely drags down its overall growth rate.

• In contrast, HDD demand is more concentrated in AI-benefiting data centers, allowing its overall growth rate to surpass NAND.

• This calculation is consistent with the recent optimistic views of Seagate and Western Digital. We have every reason to believe that these two companies will again raise their demand outlook for 2026 in the coming quarters.

Epilogue

Finally, returning to the beginning of the article, what intrigued me the most was the method used by the physical therapist. After assessment, Mr. Osamu Yata believed that Yamamoto’s elbow pain was not an issue with the elbow itself, but stemmed from his “incorrect use of the body” and “inconsistent pitching mechanics”. His pitching style at the time (perhaps including traditional weight training methods) placed excessive stress on his elbow.

Harmony is More Effective Than Brute Force

After some research, Mr. Osamu Yata developed a system whose goal is “not to train muscles, but to awaken the body’s dormant senses and range of motion”. This shifted the training focus from strengthening (muscular power) to sensation and connectivity (coordination). This is called the “BC Exercise” (BC エクササイズ), a very unique and “anti-traditional” training system. The meaning of BC includes “Body Coordinate” or “Bio” (Life) and “Cell”, meaning to fundamentally improve the body’s operation.

He believes that excessive muscle strengthening restricts the body’s flexibility and coordination, becoming an obstacle to peak performance. Therefore, his training menu is highly original, such as using light javelins to correct body coordination during pitching, or using inverted walking and bridge poses to strengthen full-body coordination. Behind this lies the profound concept of Datsuryoku (脫力)—which is not simply relaxation, but “not using superfluous power,” allowing necessary muscles to function at the necessary moment. This pursuit of efficient, harmonious power output is the foundation for building Yamamoto’s “unbreakable body” (壊れない身体)—a body possessing both flexibility and explosiveness. Its durability comes from the flexible connection of joints and fascia, rather than stiff muscles.

“Better Than Yesterday” — Mental Strength Proves the Limits of the Flesh

This unique physical foundation created the conditions for seemingly impossible endurance performance. However, Yata believes the true catalyst did not originate from the physical body, but from the mind. His perfect relief performance on “zero days rest” during the World Series is the best proof.

In fact, at five o’clock that morning, Yata received a message from Yamamoto. This detail made Yata realize that Yamamoto was under extreme tension and highly focused. When Yamamoto Yoshinobu’s intense will to “want to win” reached its peak, the body operated smoothly in response, as if mental strength eliminated physical fatigue and limitations. Yata’s interpretation perfectly embodies his philosophy: “It is because of that Qi—that spirit and will—that this was achieved”. This is not merely a display of willpower, but the ultimate proof that mental intensity directly acts upon the physical body, enabling it to surpass physiological limits.

“The Answer Is Within You” — Guiding “Awareness” Instead of Directly “Teaching”

The core of Coach Yata’s guiding philosophy is “not teaching from the outside, but sensing from within”. He never directly gives players the “correct answer”. This was evident from their first meeting: Facing the then 18-year-old Yamamoto Yoshinobu, Yata did not make any formal corrections. Instead, he guided him to “verbalize the current dynamics of his body,” making him describe the pressure on his soles and the shift of his center of gravity.

Through questioning and prompting, he guides the player to turn their attention to internal body sensations. Among these, “breathing” plays a conductor-like role, connecting the rhythm of the mind and body. Coach Yata uses a vivid analogy to explain this concept: “People always focus on the hands and feet, but those are just the branches and leaves. The trunk is what’s important. If the trunk is hollow, the branches will break”. This teaching method cultivated Yamamoto’s strong “subjectivity”. He stopped passively receiving training and proactively became the “internal supervisor” of his own body. This also explains why Yamamoto always refers to his training content as “trade secrets”. This is not only for confidentiality but also to “allow the acquired sensation to continue maturing within himself,” deepening his self-sovereignty over his body.

“Posture is the Root, Form is the Result” — Starting with “Standing Correctly”

In Yata’s theory, the starting point of everything is extremely simple: “Standing correctly is the beginning of everything”. He believes that if the body’s alignment (head, ribcage, pelvis, feet) is not in a straight line, the force transmission chain will deviate. To achieve this, he had Yamamoto spend a great deal of time on the most basic standing posture and breathing sensation, emphasizing that “if the root is stable, the branches and leaves will naturally be tidy”.

This highlights the fundamental difference between Yata’s theory and traditional posture correction. Traditional methods correct “form” externally, while Yata adjusts the “root” internally. He firmly believes that “Form is the result of sensation, not the cause” (形は感覚の結果であって原因ではない). Yamamoto Yoshinobu’s stable and elegant pitching form was not a deliberate imitation of some “ideal form”, but the “natural result” when his inner sensation, body balance, and breathing rhythm were all adjusted to their optimal state.

Conclusion: Returning to the Body’s Voice

The mentorship between Osamu Yata and Yamamoto Yoshinobu not only created miracles on the field but also offered a profound reflection for modern sports science. The core spirit of Yata’s theory is that true strength does not come from fighting the body or training muscles, but from engaging in a deep dialogue with the body and fully trusting internal sensations.

By “awakening” the body instead of merely “training” it, and “guiding” awareness instead of “commanding” form, and grounding everything in a stable “posture”, Yata ultimately empowered Yamamoto to become his own “inner coach”—a true master of his physical and mental state. Yamamoto Yoshinobu is not just a technically superb pitcher; he is also a “philosopher of the body”.

This made me rethink my current working philosophy. In the process of investing, researching, and living, it is more important to turn back and observe my inner self, body, and emotions. Before making every decision, I need to return to the core: only invest in businesses you truly understand. I believe Warren Buffett’s investment philosophy is very similar to Mr. Osamu Yata’s philosophy.

And to understand:

• A company’s “moat“ is the “power chain” of its business model.

• A mediocre company: might have “good products” (strong arms), but a messy “supply chain” (tight thoracic spine), and poor “cash flow” (weak core). Its strength cannot be transmitted, and it is prone to “injury” at any time.

The most important primary goal is “injury prevention”.

• This is synonymous with Buffett’s “First Rule: Never lose money,” which is the equivalent of the BC philosophy’s “Never get injured”.

• Finding the “margin of safety” of a business is the cornerstone of investment.

And Awareness (Anti-Traditional): Acting when the “market is imbalanced”.

• Similar to Buffett’s Philosophy: “Be fearful when others are greedy, and greedy when others are fearful”.

Buffett is the strongest BC Philosophy investor. Throughout his life, he has “rejected brute force” (not chasing hot spots). He only operates on his “central axis,” waiting for businesses with a “perfect power chain” (a moat). He only takes “smooth action” when provided with “absolute safety” (margin of safety). He is not “investing”; he is performing “the BC exercise of investing“.

—

In an era pursuing efficiency and data, and full of rapidly updating AI information and technology, we might need to listen more to the truest voice of our own body. The real answer has always been within us. Gradually building a sturdy trunk makes us more capable of facing the fast market winds and turmoil.-Insightology Research

What I Read

Transcript

• COHR Earnings Call Transcript

• LITE Earnings Call Transcript

• AXTI Earnings Call Transcript

• SanDisk Earnings Call Transcript

• Asia Supply Chains Earnings Call Transcript

Youtube

Book & Article

• Rewire: Break the Cycle: Alter Your Thoughts and Create Lasting Change by Nicole Vignola (Author)

• https://stratechery.com/2025/the-benefits-of-bubbles/

• https://news.yahoo.co.jp/articles/52bc23f9c45e1e309cd46109adacb535af6a1ae1

Life & Insight

The weather in Austin has turned cold these past few days, with daytime temperatures around 50°F. It’s been a while since I’ve felt this cool air, and my spirits feel refreshed as well. This week, I watched some interviews from the past six months with Roelof Botha, the Managing Partner of Sequoia Capital, and I had a few takeaways that deeply resonated with me. (I’ll put the links at the end).

Regarding how Sequoia identifies founders, manages capital, and makes decisions in the AI era, I believe it returns to first principles: desire, discipline, and continuous correction. Among these, there are a few points that, as a public market investor and industry researcher, I am also reflecting on:

• Founders should be like pirates: Sequoia looks for those “who want to be pirates, not join the navy.” The navy is disciplined and follows orders; but pirates are unconventional, unpredictable, and ignore the rules. Those who change the world often look more like pirates.

• Excess capital is poison: After selecting the right founders, the next challenge is providing the right resources. However, Botha presented a most counter-intuitive argument: excess capital is not a resource, but a poison.

• Long-term thinking: One must have long-term thinking from the very beginning (which I find extremely difficult). In the early stages, many companies face the risk of burning through all their cash. But at the same time, they cannot be so short-sighted as to only chase short-term opportunities. They need a long-term vision for product and business development. This balancing act is a challenge every founder must confront.

Working in the investment research industry, the problem I need to solve for clients is how to maximize capital efficiency at the lowest possible risk. Besides identifying good prices, recognizing good corporate value, and taking action, after watching these interviews, I am also reflecting on whether I, in this role, possess a “pirate spirit,” long-term thinking, and whether I have become too comfortable.

In the AI era, the speed at which capital markets are expanding rivals the speed of technological iteration. There are even more temptations for people to leverage or go “All in” in the stock market, while lacking real productivity. After all, something produced today might be replicated by someone else in a few days or weeks.

My own investment research is the same. I track industry news and technological changes daily, and AI has accelerated the speed and breadth of information flow. But AI is in the process of commoditizing the “technology” aspect of products (making it less valuable). The Sequoia interview mentioned a very important point: return to the fundamentals. Value is shifting from “code” back to the fundamentals of business:

• Deep understanding of customer needs: You understand the customer’s workflow and pain points better than anyone.

• Distribution capability: How do you reach the first 100 paying users in 30 days?

• Product stickiness: How do you create switching costs for users?

• Speed of brand building: How do you quickly become synonymous with a certain product category?

When technology itself becomes cheap and ubiquitous, those “old-fashioned” business fundamentals—sales, marketing, customer service, and brand—conversely become more important than ever.

Finally, there’s “passion” and “discipline,” which I also care deeply about. These viewpoints from Sequoia all seem to point to the same core: companies that can internalize this discipline into their own operating system, and deliver their passion to the world, will be the survivors that create lasting value. While doing research, I often remind myself to keep reading daily, exchanging ideas with talented people, learning new technologies, and reading history—to do this for a lifetime and enjoy the process.

I hope to continue building a deep foundation, one that can withstand the fierce winds of market cycles, and to continuously grow upwards, seeking new light.

Market

High probability of a December rate cut; the Fed is currently just managing market expectations.

First, from a macroeconomic perspective, Powell sent a very important signal at this press conference: AI capital expenditure is almost unaffected by interest rate policy in the short term. The growth of the AI industry no longer depends on the cost of capital but is driven by technological innovation and structural demand. This also implies that the U.S. economy is entering a new productivity cycle centered on AI.

• The Fed cut interest rates and, starting in December, will no longer reduce its balance sheet (QT). This is good news for market liquidity.

• Chairman Powell came out to “cool things down,” hinting that another cut in December is not certain: Although rates were cut, Chairman Powell specifically emphasized at the press conference that the labor market is only cooling “gradually” and is not yet in a bad state. He hinted that a rate cut in December is “not inevitable” and will depend on subsequent data, with the goal of preventing the market from getting over-excited.

• Insightology View: The probability of a December rate cut is still high. Powell is currently just “managing market expectations” to prevent the market from becoming too euphoric. We believe that as falling oil prices help cool inflation, it is still very likely the Fed will cut rates again in December. Short-to-medium-term US Treasuries may also have an opportunity for good performance in the future.

Major CSPs’ Operating Cash Flow Can Still Support Upward Revisions to 2026 Data Center Capex YoY

This week, all four major CSPs held their third-quarter earnings calls. Several key points include: all four raised their capital expenditure forecasts for 2025 and continued to increase investment in AI.

If you look at their operating cash flow, it can still support current and future large-scale investments in AI/data centers:

• Meta’s average quarterly operating cash flow is about $25 billion; the TTM (Trailing Twelve Months) announced in Q3 2025 was approximately $102 billion.

• Google’s (Alphabet) TTM operating cash flow announced in Q3 2025 was approximately $133.7 billion.

• Amazon’s TTM operating cash flow announced in Q3 2025 was approximately $130.7 billion.

• Microsoft’s latest TTM figure is approximately $122 billion, with the most recent four quarters each maintaining roughly $30 billion.

  

In fact, because their operating cash flow remains sufficient, there is still a possibility they will raise (or upwardly revise) next year’s Capex. Although the market may worry about free cash flow being squeezed, these major CSPs must sacrifice short-term free cash flow right now in order to win the AI race.

Insightology View: Although Meta’s and MSFT’s short-term results did not meet expectations, I believe these major companies all have the opportunity to catch up quickly. Remember last quarter, everyone was saying Amazon’s AWS growth was lagging behind its peers, but after just one quarter, everyone has become optimistic again.

Among these, electricity is also a key focus this time. Amazon stated in its earnings call that it added 3.8 GW of power capacity TTM (Trailing Twelve Months), which is double the 2022 level, and it expects to double this again by 2027. It also expects to add another 1 GW of power in the last quarter of this year. We also published an article on 800V HVDC last week and will continue to provide updates on the industry’s status.

Andy Jassy - President, Chief Executive Officer, Director:
Yeah. On the capacity side, we brought in quite a bit of capacity, as I mentioned in my opening comments, 3.8 gigawatts of capacity in the last year with another gigawatt plus coming in the fourth quarter and we expect to double our overall capacity by the end of 2027. So we’re bringing in quite a bit of capacity today, overall in the industry, maybe the bottleneck is power. I think at some point, it may move to chips, but we’re bringing in quite a bit of capacity. And as fast as we’re bringing in right now, we are monetizing it.

Memory Industry: AI Video Demand is Poised for an Outbreak; 2026 Demand Growth Rate Will Be Stronger Than Expected

• HDD Demand Outlook Sharply Raised: Major HDD players (such as STX, WDC) have significantly raised their demand outlook for the next 4 to 6 quarters.

• Main Driver: The primary reason for the surge in demand is driven by AI inference needs, particularly the expansion of AI-generated video. This is causing large cloud service providers (Hyperscalers) to accelerate procurement. This differs from past demand models that relied on non-AI data growth or cyclical server refreshes.

• Upward Revisions from Players:

  • Western Digital (WDC): Raised its 2026 bit demand growth forecast to over +23% (up from +20%), with order visibility extending into early 2027.

  • Seagate (STX): Expects 2026 bit demand growth to reach the “mid-20s” (i.e., mid-20% range), with order visibility extending from late 2026 to early 2027.

Dave Mosley - Seagate CEO:
AI-generated videos promise to further fuel demand growth. There are already numerous text-to-video tools that democratize creativity by letting anyone generate professional quality videos from text, images or sketches. We see this trend already taking hold. For example, Google reports over 275 million videos were generated on its Veo platform within the first 5 months. With a 1-minute AI video being up to 20,000 times larger than a 1,000 word text file, the data storage implications are clear. The rapid adoption and growing capability of these tools are already having a positive impact on the demand for storage.

NAND Flash: Benefiting from High-Capacity HDD Shortages, Enterprise SSD Substitution Demand Continues into H1 2026

• Short-term Positive (Continuing into H1 2026): Due to shortages of high-capacity HDDs (traditional hard drives), data centers are turning to procurement of high-capacity QLC enterprise SSDs (solid-state drives) as a substitute. This has led to a significant increase in NAND Flash demand.

• HDD supply is not expected to improve until H2 2026 at the earliest. Therefore, the substitution demand for SSDs will continue until at least H1 2026.

• This strong substitution demand is expected to offset the decline in consumer demand, allowing NAND Flash to show better-than-seasonal performance in the typically weak H1 2026 (i.e., “the off-season won’t be slow”).

• Long-term Concern (Starting H2 2026): This current wave of strong demand for enterprise SSDs is primarily triggered by the HDD shortage.

• Once high-capacity HDDs using HAMR (Heat-Assisted Magnetic Recording) technology begin mass production in H2 2026 to H1 2027, the HDD shortage issue may be alleviated.

• At that point, it will be necessary to continuously monitor whether the restored balance in HDD supply will, in turn, weaken the substitution demand for enterprise SSDs.

View from Major Memory Equipment Makers: Upward Revisions for Memory Equipment Procurement Demand Will Be Stronger Than for Logic Equipment

• AI Demand Surge Offsets Negative Impacts: Benefiting from the surge in end-market AI demand in Q3 2025, it is sufficient to offset the impact of declining semiconductor equipment procurement demand from China.

• Major Equipment Makers Raise Outlook: ASML raised its 2026 outlook from “flat” to “no lower than 2025.” Lam Research (LRCX) expects revenue growth to continue into H2 2026.

• Data Centers Driving Growth: Lam Research estimates that every $100 billion in data center investment drives approximately $8 billion in additional semiconductor equipment spending. Considering that data center investment next year (2026) may increase by $80-$100 billion compared to this year (2025), this could contribute an additional $6.5-$8 billion in semiconductor equipment spending. This is equivalent to contributing an extra 6-7 percentage points of growth momentum, making it possible for the 2026 semiconductor equipment industry value to maintain over 10% growth.

• VEU Revocation Impacts Memory: The US revocation of VEU (Validated End-User) authorization for Samsung’s and Hynix’s fabs in China restricts their ability to expand or upgrade advanced processes there. This has a huge impact on the supply of advanced NAND and DRAM processes.

• Memory Equipment Demand is Strongest: The VEU restrictions, combined with the AI demand surge, will force Samsung and Hynix to accelerate the expansion of advanced memory capacity outside of China. It is expected that the upward revision for 2026 memory equipment procurement demand will be stronger than that for logic equipment.

• Market Signal Corroboration: In ASML’s new orders last quarter, memory-related orders grew +186% quarter-over-quarter (QoQ), significantly higher than logic-related orders.

Most of this week’s information was heavily focused on memory. However, I believe that AI Server PCBs (Printed Circuit Boards) are the next key area to watch, and I will be analyzing this topic in the coming one or two weeks.

My own personality isn’t suited for being an aggressive pirate, plundering other people’s territories or resources. I much prefer to observe and predict the ocean currents and, taking it a step further, to help others follow these currents together as we all advance toward the great voyage. - Insightology Research

What I Read

Transcript

• Seagate_STX 25Q3 Earnings Call Transcript

• WDC 25Q3 Earnings Call Transcript

• SK Hynix 25Q3 Earnings Call Transcript

• Lam Research_LRCX 25Q3 Earnings Call Transcript

• KLAC 25Q3 Earnings Call Transcript

• Meta 25Q3 Earnings Call Transcript

• MSFT 25Q3 Earnings Call Transcript

• AMZN 25Q3 Earnings Call Transcript

Youtube

• Lessons from 20 Years of Venture Capital: Roelof Botha (Managing Partner at Sequoia Capital)

• Sequoia’s Roelof Botha on Decision Making, AI, and the Next Trillion Dollar Markets | Ep. 28

• Roelof Botha on PayPal Panic & Sequoia’s Long-Term Strategy | The Library of Minds

• Sequoia’s Roelof Botha: Why Venture Capital is Broken & How Great Companies Are Built

Book

• On Writing Well: The Classic Guide to Writing Nonfiction (Continue Reading)