Gavin Baker's three contrarian bets: TSMC saving the market, Trainium being undervalued, and space computing power revealed within two years

Tech investment mogul Gavin Baker boldly predicts at the Sohn Conference: Amazon Trainium is the most undervalued AI chip this year; TSMC's "stubborn old men" are using physical constraints to block the bubble; Space orbital computing power will prove feasible within two years, impacting the ground data center supply chain.
(Background: Amazon increases investment in Anthropic by $25 billion: not buying AI models, but the moat of computing power)
(Additional context: Amazon reportedly invests billions in OpenAI, pushing its own chip Trainium to challenge NVIDIA's dominance)

Table of Contents

Toggle

• The Undervalued Treasure: Why Trainium Will Be the TPU of 2026
• TSMC's "Stubborn Old Men" Are Using Physical Limits to Block the Bubble
• Space Computing Power to Show Results in Two Years: Ground Device Makers Are Most at Risk
• Countercyclical Memory Cycle Operations: This Might Be Like the Mid-1990s
• AI Pricing Shifts to "Pay-as-You-Go": Revenue Will Surpass Expectations

When the market's focus is entirely on NVIDIA GPUs and Google TPUs, Gavin Baker points out a critically overlooked keyword at the Sohn Conference—Amazon Trainium. This veteran tech investor, who managed $17 billion at Fidelity, gave an exclusive interview, stating that the most undervalued AI chip company in his view is Trainium, and predicts that after Trainium 3 reaches mass production in the second half of this year, it will reenact the explosive growth of TPU in 2025. Not only that, he boldly challenges the consensus: TSMC's management is using a conservative expansion strategy to save everyone’s market, preventing AI bubbles from repeating history.

The undervaluation of Trainium far exceeds others. Its significance for 2026, especially after Trainium 3's volume ramp-up in the second half of this year, is akin to TPU's significance for 2025. If someone is very bullish on TPU today, they might want to check their 13F filings—see if they hold positions in Lumentum or Celestica—that's where the best investments in TPU are. I hold one of these, so I feel confident in saying this.

He also said that TSMC refuses to expand capacity as quickly as Jensen Huang hopes. "Jensen visits TSMC every three months, and they expand about 5%. Jensen wants their capacity to double or triple. If capacity truly doubles or triples, NVIDIA could sell about $1.5 trillion worth of chips next year—I’m serious."

Regarding the memory cycle, Baker states that based on the past 25 years of memory cycles, now is 100% the time to sell memory stocks.

I was actually an analyst at Micron in 2000, remember attending their analyst day in Silicon Valley, and have experienced numerous memory cycles. From historical patterns, now is indeed the time to sell.

However, there is one cycle that should never be sold—mid-1990s, which I consider the last true capacity cycle. Compared to that cycle, we might still be in very early stages.

Regarding AI revenue, Baker states that the labor structure of S&P 500 companies will face a "major adjustment," but the shift from "monthly subscriptions" to "pay-as-you-go" pricing will cause revenue to grow faster than expected—comparable to the mobile industry’s shift from flat-rate plans to per-minute billing.

The Undervalued Treasure: Why Trainium Will Be the TPU of 2026

He also emphasizes that reading is overwhelmingly important, and he almost no longer actively meets with management of public companies—these executives are trained to say nothing beyond their earnings calls or 10-Q filings.

Blackstone senior partner Jas Khaira asked Baker during an interview: among NVIDIA’s competitors—Google TPU, Amazon Trainium, Intel Gaudi—which is the most undervalued? Baker replied: "Trainium, without a doubt."

He provided specific technical reasoning. Currently, leading AI models all use a architecture called "Mixture of Experts" (MoE). To infer such models, a foundational infrastructure called "Switched Scaleup Network" is required.

Baker said: "Only two companies worldwide are implementing active Switched Scaleup Networks—one is powering NVIDIA GPUs, the other is Amazon Trainium."

This is a very overlooked technical barrier. Google TPU does not have comparable capability—Baker pointed out a detail: "Google invented the MLPerf benchmark, but they don’t submit TPU results to their own benchmark. You can see this makes Jensen (Jensen Huang) crazy."

Baker also predicts that after Trainium 3’s large-scale production in the second half of this year, Trainium’s position in 2026 will be comparable to TPU’s in 2025. "I’ve invested in TPU supply chain companies like Celestica before, so I feel qualified to say this."

He added: "I will never short Google, nor Broadcom, but I do believe Trainium is seriously undervalued right now."

TSMC’s "Stubborn Old Men" Are Using Physical Limits to Block the Bubble

Another topic that drew attention in this conversation is "Orbital Compute"—the idea of placing data centers in space.

Khaira asked Baker: when can this truly become commercial?

Baker’s answer gave a clear timeline: "I believe within the next two years, its feasibility and economics will be validated. By the end of this decade, it will start capturing meaningful market share."

The logic is that terrestrial data centers face two major physical constraints: power and cooling. In space, power comes from the sun, cooling from the satellite’s shadow side.

Baker described a satellite design he saw from a potential orbital compute provider: radiators three to four hundred feet long, with the satellite body itself being a rack—8 feet high, 2.5 feet wide, 4 feet deep—multiple racks connected via lasers, forming a virtual data center. The radiators are placed behind the racks’ shadow.

He pointed out that once this route is feasible, the biggest impact will be on ground data center power and cooling suppliers: "Those large industrial companies expanding capacity to support data center construction might face demand suddenly stopping."

He also emphasized that existing ground data centers still have value; training and reinforcement learning will still happen on the ground. "I can’t imagine we won’t build a ground data center in the next seven years," but the trajectory of incremental demand is being redefined.

Space Orbital Computing to Show Results in Two Years: Ground Device Makers Are Most at Risk

A common question in the market: will AI investment turn into a repeat of the internet bubble?

Baker’s answer: this time might be different, and the reason is unexpectedly—TSMC’s management is conservative.

He said that every major new technology in history—from railroads, canals, PCs, the internet to AI—almost without exception, experienced a bubble. Investors get excited about new tech, market consensus forms, bubbles inflate, and ultimately, infrastructure is built with bubble money—the internet was built this way.

"We don’t want a bubble. Bubbles are terrible; going through a bubble is painful, and bursting it is even worse."

But this time, he "optimistically believes" we might avoid a bubble, thanks to physical constraints in the real world—shortages of watts (power) and wafers.

The key to wafer shortages is TSMC’s attitude. Baker said: "TSMC is run by stubborn old men over 70." (He then joked, 70 is the new 50, and he himself is 50.)

This group has experienced Taiwan’s semiconductor industry going from chasing Intel—considered an impossible dream—to achieving it with a lifetime of effort. They deeply understand what a bubble and crash would mean for TSMC.

Countercyclical Memory Cycle Operations: This Might Be Like the Mid-1990s

So, they refuse to expand capacity as quickly as Jensen Huang hopes.

"Jensen visits TSMC every three months, and they expand about 5%. Jensen wants their capacity to double or triple. If capacity truly doubles or triples, NVIDIA could sell about $1.5 trillion worth of chips next year—I’m serious. But on the other hand, this could be very painful for everyone."

Baker’s conclusion: these "stubborn old men," by enforcing real physical constraints, objectively help everyone avoid bubbles—and such constraints have never appeared in any previous technological revolution.

In the conversation, Baker also mentioned two other important judgments.

Regarding the memory cycle: memory prices have already risen 60% to 70% this year, Micron’s gross margin could reach over 60%, far above the historical average (~16%).

Baker candidly said that based on the memory cycle law of the past quarter-century, "now is 100% the time to sell memory stocks." But he believes this cycle might be similar to the true capacity cycle of the mid-1990s, "we might still be in the early stages," and shouldn’t simply apply historical templates.

Regarding AI revenue scale: Baker judges that the point when OpenAI and Anthropic together reach $200 billion in revenue is not far off.

AI Pricing Shifts to "Pay-as-You-Go": Revenue Will Surpass Expectations

He cites Jensen Huang’s statement: Jensen wants his top engineers to spend at least half their AI token compensation.

Baker believes this trend means the labor structure of S&P 500 companies will face a "major adjustment," but the shift from "monthly subscriptions" to "pay-as-you-go" pricing will cause revenue to grow faster than expected—comparable to the mobile industry’s shift from flat-rate plans to per-minute billing.

In the interview, Khaira also asked Baker where his investment edge comes from.

Baker’s simple answer: "Reading, overwhelmingly." He said he almost no longer actively meets with management of public companies—"They are very well trained, and never say anything outside their earnings calls or 10-Qs, and I read much faster than they speak."

He admitted one of his most painful lessons was writing a letter to a company’s board requesting a buyback, only for the company to go bankrupt 18 months later. "That’s a permanent lesson about leverage—sometimes things don’t go as planned."

Baker mentioned that throughout his career, he has been trying to overcome Peter Lynch’s adage—"weed out the losers, hold the winners"—but for some reason, it’s extremely difficult for him.

He is highly sensitive to valuation, fundamentally a contrarian investor, most comfortable at the 52-week lows. He admits he’s been holding onto memory stocks stubbornly. But this is a lifelong journey, and he tries to improve a little each year.