Recently, at a major Wall Street investment conference—the 2026 Sohn Conference, tech investment mogul and CIO of Atreides Management, Gavin Baker, was interviewed.
Baker previously managed over $17 billion in assets at Fidelity and is a seasoned investor in the semiconductor field.
In the interview, he put forward several judgments that directly challenge market consensus: the most undervalued AI chip currently is Amazon's Trainium; TSMC's "conservative" capacity expansion strategy is helping the industry avoid a bubble; orbital computing power will prove its viability within two years and will begin to impact the ground-based data center ecosystem by the end of this decade.
He stated, I would never short Google, I would never short Broadcom, but I do think Trainium is very much undervalued today.
Trainium is undervalued to a far greater degree than the others. What Trainium means for 2026, especially once Trainium 3 truly ramps in the second half of this year, is analogous to what TPU meant for 2025. If someone is very bullish on TPU today, perhaps they should look at their 13F to see if they hold Lumentum or Celestica—those are the two best proxies for investing in TPU. I own one of them, so I think I have the right to say this.
He also said that TSMC refuses to expand capacity as rapidly as Jensen Huang wants. "Jensen goes to TSMC every three months, and they expand maybe 5%. Jensen wants them to double or triple capacity. If capacity actually doubled or tripled, Nvidia could probably sell $1.5 trillion worth of chips next year—I'm serious."
Regarding the memory cycle, Baker said that looking at every memory cycle over the past 25 years, now is 100% the time to sell memory.
I was actually a Micron analyst back in 2000, I remember going to their analyst day in Sun Valley, I've been through countless memory cycles, and based on historical patterns, it is indeed time to sell.
However, there is one cycle you absolutely should not have sold in—that was the mid-1990s, which I consider the last truly meaningful capacity cycle. Relative to that cycle, we might still be in the very early stages now.
Regarding AI revenue, Baker stated that the labor structure of S&P 500 companies will face a "significant adjustment," but the shift in AI pricing models from "subscription" to "pay-as-you-go" will cause revenue growth to outpace external expectations—he compared this to the mobile telecom industry's "per-minute overage" profit model back in the day.
He also said that reading is overwhelmingly important and emphasized that he almost never actively requests meetings with listed company management—these executives are extremely well-trained and will never say anything not already in the earnings call or the 10-Q filing.
And I can read those documents much faster than they can talk. The following is a summary of the key points compiled by Touzi Zuoyeben (WeChat ID: touzizuoyeben), shared with everyone:
Amazon Trainium: The AI Chip Most Undervalued by the Market
When Blackstone Senior Partner Jas Khaira interviewed Baker, he asked: Among Nvidia's competitors—Google TPU, Amazon Trainium, Intel Gaudi—which one is the most underestimated by the market? Baker answered: "Trainium, without a doubt."
He provided specific technical logic. The current mainstream cutting-edge AI models all use an architecture called "Mixture of Experts" (MoE). To infer such models, an infrastructure called a "Switched Scaleup Network" is required.
Baker said: "There are only two companies in the world today with operational switched scaleup networks—one that drives Nvidia GPUs, and the other for Amazon Trainium."
This is a technical threshold easily overlooked. Google TPU does not possess equivalent capability in this area—Baker pointed out a detail directly: "Google invented the ML Perf benchmark, but they don't submit TPU results to their own benchmark; you can tell this drives Jensen crazy."
Baker also judged that once Trainium 3 ramps massively in the second half of this year, Trainium's position in 2026 will be equivalent to TPU's position in 2025. He said he had invested in TPU supply chain companies like Celestica, "I think I'm qualified to say that."
He added: "I would never short Google, I would never short Broadcom, but I do think Trainium is very much undervalued today."
Space-Based Data Centers: Proven Within 2 Years, Grabbing Share by End of Decade
Another topic that drew attention in this conversation was "Orbital Compute"—the idea of placing data centers in space.
Khaira asked Baker: When will this truly commercialize?
Baker's answer provided a clear timeline: "I believe within the next two years, its feasibility and economics will be proven. By the end of this decade, it will begin to capture meaningful market share."
The logic lies in the two major hard constraints facing ground-based data centers: power and cooling. In space, power comes from the sun, and cooling comes from the satellite's shaded side.
Baker described a potential orbital compute provider's satellite design he saw: radiators stretching three to four hundred feet, the satellite body itself being a rack—8 feet high, 2.5 feet wide, 4 feet deep—with multiple racks connected by lasers forming a virtual data center. The radiators are placed behind the racks' shadows.
He pointed out that once this path becomes viable, the biggest impact will be on suppliers of power and cooling equipment for ground-based data centers: "Those industrial companies massively expanding capacity to support data center construction might face a (demand) cliff."
He also emphasized that existing ground-based data centers will still have value; training and reinforcement learning will still happen on the ground. "I can't imagine that in the next seven years we will never build another ground-based data center," but the direction of incremental demand is being redefined.
TSMC's 'Stubborn Old Men': Helping the Global Market Avoid a Bubble
A common question in the market: Could AI investment become a repeat of the internet bubble?
Baker's answer is: This time might be different, and the reason is surprising—the conservatism of TSMC's management.
He said that throughout history, with every major new technology, from railroads, canals, PCs, the internet, to AI, bubbles almost invariably occurred. Investors get excited about the new tech, market consensus forms, the bubble inflates, and eventually bubble money funds the infrastructure build-out—that's how the internet progressed.
"We don't want a bubble. Bubbles are awful, going through them is painful, and what happens after they burst is even more painful."
But this time he is "optimistic that" we might avoid a bubble, precisely because of the physical constraints existing in the real world—shortages of watts (power) and wafers.
The key to the wafer shortage lies in TSMC's attitude. Baker said: "TSMC is run by stubborn old men in their 70s." (He then joked that 70 is the new 50, and he himself is 50.)
These individuals have experienced the journey of Taiwan's semiconductor industry from chasing Intel, which was considered "the impossible dream of a lifetime," to achieving it over their lifetimes. They deeply understand what a bubble and crash would mean for TSMC.
Therefore, they simply refuse to expand capacity as rapidly as Jensen Huang wants.
"Jensen goes to TSMC every three months, and they expand maybe 5%. Jensen wants them to double or triple capacity. If capacity actually doubled or tripled, Nvidia could probably sell $1.5 trillion worth of chips next year—I'm serious. But the flip side of that, for everyone else, would likely be very painful."
Baker's conclusion is: These "stubborn old men," by enforcing a real-world physical constraint, are objectively helping everyone avoid a bubble—and this type of constraint has never existed in any previous technological revolution.
Memory Cycle and AI Revenue Explosion
In the conversation, Baker also mentioned two noteworthy judgments.
Regarding the memory cycle: Memory prices have risen 60% to 70% this year, and Micron's gross margin could reach over 60%, far exceeding historical averages (around 16%).
Baker admitted that according to the patterns of memory cycles over the past 25 years, "now is 100% the time to sell memory stocks." But he believes this time might be similar to the true capacity cycle of the mid-1990s, "we might still be early," and one should not simply apply the historical template.
Regarding AI revenue scale: Baker judges that the point where OpenAI and Anthropic combined revenue reaches $200 billion is not far away.
He cited Jensen Huang's remark: Huang wants his best engineers to spend at least half their salary on AI tokens.
Baker's judgment is that this trend means the labor structure of S&P 500 companies will face a "significant adjustment," but the shift in AI pricing models from "subscription" to "pay-as-you-go" will cause revenue growth to outpace external expectations—he compared this to the mobile telecom industry's "per-minute overage" profit model back in the day.
Investment Philosophy: Reading, Pattern Recognition, and a Misdirected Letter
In the interview, Khaira also asked Baker where his investment edge comes from.
Baker's answer was succinct: "Reading, overwhelmingly." He said he almost never actively requests meetings with listed company management—"They are very well-trained, they never say anything that's not in the earnings call or the 10-Q, and I read much faster than they can talk."
He admitted one of the most painful lessons of his career was once writing a letter to a company's board demanding a stock buyback, only to have that company go bankrupt 18 months later. "That is a permanent lesson about high leverage—sometimes not everything goes according to plan."
Baker mentioned that a lifelong struggle throughout his career has been overcoming Peter Lynch's adage—pull the weeds and water the flowers, meaning sell the losers and hold the winners. But for some reason, this has been incredibly difficult for him.
Baker is extremely valuation-sensitive, a contrarian at heart, and the 52-week low list is his comfort zone. He admitted he has been stubbornly holding onto memory stocks. But this is a lifelong practice; he tries to improve a little bit each year in this regard.
