Recently, at the prestigious Wall Street investment conference—the 2026 Sohn Conference—Gavin Baker, Chief Investment Officer of Atreides Management and a leading tech investor, was featured in an exclusive interview.

Baker previously managed over $17 billion in assets at Fidelity and is an experienced investor in the semiconductor sector.

In the interview, he made several direct challenges to market consensus: the most undervalued AI chip today is Amazon’s Trainium; TSMC’s conservative expansion strategy is helping the industry avoid a bubble; and space-based computing power will be proven feasible within two years and begin disrupting ground-based data center infrastructure by the end of this decade.

He said he would never short Google or Broadcom, but he does believe Trainium is currently significantly undervalued.

Trainium is far more undervalued than any other. The significance of Trainium to 2026—especially after Trainium 3 truly scales up in the second half of this year—is analogous to the significance of TPU to 2025. If someone is very bullish on TPU today, they should review their 13F filings to see if they hold Lumentum or Celestica—those are the two best ways to invest in TPU. I hold one of them, so I feel confident saying this.

He also said that TSMC is not expanding production as quickly as Jensen Huang hopes. “Jensen Huang visits TSMC every three months, and they expand capacity by about 5%. Huang wants them to double or even triple capacity. If capacity were truly doubled or tripled, NVIDIA could sell around $1.5 trillion worth of chips next year—I’m serious.”

Regarding memory cycles, Baker stated that based on every memory cycle over the past 25 years, now is 100% the right time to sell memory.

I was actually an analyst at Micron back in 2000, and I remember attending their analyst day in Sun Valley; I’ve been through countless memory cycles, and historically, now is indeed the time to sell.

However, there is one cycle you should never have sold—mid-1990s, which I believe was the last true capacity cycle. Compared to that cycle, we may still be in a very early stage.

Regarding AI revenue, Baker said that the workforce structure of S&P 500 companies will face a "significant adjustment," but the shift in AI pricing models from "subscription-based" to "pay-per-use" will drive revenue growth faster than expected—he compared this to the mobile calling industry's profit model of charging "overage fees per minute" beyond plan limits.

He also said that reading is overwhelmingly important and emphasized that he almost no longer proactively meets with management teams of public companies—managers who are rigorously trained to say nothing beyond what is covered in earnings calls or 10-Q filings.

And I read these documents much faster than they speak. Below are the key highlights compiled by the class representative of the Investment Workbook (WeChat ID: touzizuoyeben) for your sharing:

Amazon Trainium: The Most Undervalued AI Chip in the Market

During his interview with Baker, Blackstone senior partner Jas Khaira asked which of NVIDIA’s competitors—Google’s TPU, Amazon’s Trainium, or Intel’s Gaudi—is the most undervalued by the market. Baker replied: “Trainium, without a doubt.”

He outlined the specific technical logic. Current mainstream cutting-edge AI models all adopt an architecture called "Mixture of Experts" (MoE). To infer these models, an infrastructure known as a "Switched Scale-up Network" is required.

Baker said: “Globally, there are currently only two companies with operational switch-based scaling networks—one powering NVIDIA GPUs, and the other Amazon Trainium.”

This is a technical barrier that’s easily overlooked. Google’s TPUs don’t have the same capability in this regard—Baker directly pointed out a detail: “Google invented the MLPerf benchmark, yet they don’t submit TPU results to their own benchmark; you can tell how frustrated Jensen (Jensen Huang) is about this.”

Baker also judged that after the large-scale mass production of Trainium 3 in the second half of this year, Trainium’s position in 2026 will be equivalent to that of TPU in 2025. He said he has previously invested in TPU supply chain companies such as Celestica, “I think I’m qualified to say that.”

He added, "I would never short Google or Broadcom, but I do believe Trainium is currently severely undervalued."

Space data center: Results will be clear within two years; aiming to capture market share by the end of this decade.

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

Khaira asked Baker: When will this be truly commercialized?

Baker's response provided clear timelines: "I believe its feasibility and economic viability will be validated within the next two years. By the end of this decade, it will begin to capture a meaningful market share."

The logic is that ground-based data centers face two hard constraints: power and cooling. In space, power comes from the sun, and cooling comes from the satellite's shaded side.

Baker described the satellite design he saw from a potential orbital computing supplier: heat radiators stretching several hundred feet in length, with the satellite body itself being a rack—8 feet tall, 2.5 feet wide, and 4 feet deep—multiple racks connected via lasers to form a virtual data center. The radiators were positioned behind the shadows of the racks.

He noted that once this approach becomes viable, the biggest impact will be on suppliers of power and cooling equipment for ground-based data centers: “Companies in the industrial sector that have significantly expanded production to support data center construction may face an abrupt end to demand.”

He also emphasized that existing ground-based data centers still hold value, and training and reinforcement learning will continue to take place on the ground: "I cannot imagine that within the next seven years we will never build another ground-based data center." However, the direction of incremental demand is being redefined.

TSMC’s “Stubborn Elders”: Helping the Global Market Avoid a Bubble

A common question in the market is: Will AI investing become a repeat of the internet bubble?

Baker's response was: This time might be different, and the reason is surprising—the conservative stance of TSMC's management.

He said that throughout history, every major new technology—from railways and canals to PCs, the internet, and AI—has almost invariably been accompanied by a bubble. Investors become excited about the new technology, a market consensus forms, the bubble expands, and ultimately, the bubble’s capital funds the infrastructure development—the internet followed this exact path.

We don’t want a bubble. Bubbles are bad; going through a bubble is painful, and it’s even more painful when it bursts.

But this time, he "optimistically believes" we may avoid a bubble, precisely due to physical constraints in the real world—the shortage of watts (electricity) 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 immediately joked that 70 is the new 50, and he himself is 50.)

These individuals witnessed Taiwan’s semiconductor industry go from being seen as chasing Intel—an impossible dream in their lifetime—to achieving it over the course of their careers. They understand better than anyone what a bubble and its collapse would mean for TSMC.

Thus, they refused to ramp up production as quickly as Jensen Huang hoped.

Jensen Huang visits TSMC every three months, and they typically increase production by about 5%. Huang wants them to double or even triple their capacity. If capacity were truly doubled or tripled, NVIDIA could sell around $1.5 trillion worth of chips next year—I’m serious. But the other side of this story could be extremely painful for everyone.

Baker's conclusion is that these "stubborn old men," by enforcing a real-world physical constraint, inadvertently helped everyone avoid a bubble—a constraint that has never existed in any previous technological revolution.

Memory cycle and AI revenue surge

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

Regarding memory cycles: Memory prices have risen 60% to 70% this year, and Micron’s gross margin could exceed 60%, far above the historical average of approximately 16%.

Baker admitted that, based on the memory cycle patterns of the past 25 years, “you should 100% sell memory stocks right now.” But he believes this cycle may resemble the true capacity cycle of the mid-1990s, and “we may still be in the early stages,” so historical patterns shouldn’t be applied simplistically.

Regarding AI revenue scale: Baker estimates that the combined revenue of OpenAI and Anthropic reaching $200 billion is not far off.

He cited Jensen Huang’s statement: Huang wants his top engineers to spend at least half of their salary on AI tokens.

Baker believes this trend will lead to a "significant adjustment" in the workforce structure of S&P 500 companies, but the shift in AI pricing models from "subscription-based" to "pay-per-use" will drive revenue growth faster than expected—he compares this to the profitable model of mobile calling plans that charged extra per minute beyond the included allowance.

Investment Wisdom: Reading, Pattern Recognition, and a Misaddressed Letter

During the interview, Khaira also asked Baker where his investment advantages come from.

Baker’s response was concise: “Reading is overwhelmingly the most important.” He said he rarely initiates meetings with management teams of public companies anymore—“They’re extremely well-trained and never say anything beyond what’s in earnings calls or 10-Qs, and I read faster than they speak.”

He admitted that one of the most painful lessons of his career was writing a letter to a company’s board requesting a stock buyback, only for the company to go bankrupt 18 months later. “It’s a permanent lesson about high leverage—sometimes not everything goes according to plan.”

Baker mentioned that throughout his career, he has struggled to overcome Peter Lynch’s adage: pull the weeds and water the flowers—sell losers and hold winners. Yet, for some reason, this has been extremely difficult for him.

Baker is extremely sensitive to valuation and is, at heart, a contrarian investor—the 52-week low list is where he feels most at home. He admits he has stubbornly held onto memory stocks. But it’s a lifelong journey, and each year he strives to make small improvements in this area.