Gavin Baker's three contrarian bets: TSMC is bailing out the market, Trainium is undervalued, and space computing will be unveiled within two years

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯3658 words

Tech investment veteran Gavin Baker made a bold prediction at the Sohn Conference: Amazon Trainium is the most underrated AI chip this year; TSMC’s "stubborn old men" are using physical constraints to block a bubble; and orbital compute will prove its feasibility within two years, disrupting the ground-based data center supply chain. (Context: Amazon increases investment in Anthropic by $25 billion: It’s not buying an AI model, but a moat of computing power.) (Background: Amazon reportedly plans to invest $10 billion in OpenAI, pushing its self-developed Trainium chip to challenge NVIDIA’s hegemony.) In an exclusive interview, Baker identified what he considers the most underrated AI chip company, predicting that after Trainium 3 enters mass production in the second half of this year, it will repeat the explosive trajectory of TPU in 2025. Furthermore, he challenged the consensus: TSMC’s management is using a conservative expansion strategy to save the market for everyone, preventing the AI bubble from repeating historical mistakes. While market attention is focused on NVIDIA GPU and Google TPU, Gavin Baker pointed out a severely overlooked keyword at the Sohn Conference—Amazon Trainium. The veteran tech investor, who once managed $17 billion at Fidelity, stated that Trainium is far more underrated than others. The significance of Trainium for 2026, especially after Trainium 3 ramps up in the second half of this year, is akin to the significance of TPU for 2025. "If anyone is very bullish on TPU today, they might want to check their 13F to see if they hold Lumentum or Celestica—those are the two best targets for investing in TPU. I hold one of them, so I feel I have the standing to say this." He also noted that TSMC refuses to expand capacity as quickly as Jensen Huang hopes. "Jensen Huang visits TSMC every three months, and they expand capacity by about 5%. Jensen wants them to double or triple their capacity. If capacity were to double or triple, NVIDIA could probably sell $1.5 trillion worth of chips next year—I’m serious." Regarding the memory cycle, Baker stated that looking at every memory cycle over the past 25 years, now is 100% the time to sell memory stocks. "I was actually a Micron analyst in 2000, and I remember going to Sun Valley for their analyst day. I’ve been through countless memory cycles, and historically, it is indeed time to sell." However, there is one cycle where one should absolutely not sell—the mid-1990s, which he considers the last true capacity cycle. Benchmarked against that cycle, we might still be in a very early stage. Regarding AI revenue, Baker noted that the labor structure of S&P 500 companies will face "major adjustments," but the shift in AI pricing models from "subscription-based" to "usage-based" will cause revenue growth to exceed external expectations—he compared this to the "overage charges per minute" profit model of the mobile telecommunications industry back in the day. He also emphasized that reading is overwhelmingly important, noting that he almost never proactively meets with listed company management—these executives are extremely well-trained and will never say anything beyond what is in their earnings calls or 10-Q filings. When Blackstone senior partner Jas Khaira asked Baker which of NVIDIA’s competitors—Google TPU, Amazon Trainium, or Intel Gaudi—is the most underrated by the market, Baker replied: "Trainium, without a doubt." He provided specific technical logic. Current mainstream frontier AI models use an architecture called "Mixture of Experts" (MoE). To perform inference on such models, an infrastructure called a "Switched Scaleup Network" is required. Baker said: "There are only two companies in the world currently running a Switched Scaleup Network—one is what powers NVIDIA GPU, and the other is Amazon Trainium." This is an easily overlooked technical barrier. Google TPU does not have equivalent capabilities in this regard—Baker pointed out a detail: "Google invented the ML Perf benchmark, but they don't submit TPU results to their own benchmark. You can tell that this drives Jensen crazy." Baker also judged that after Trainium 3 enters mass production in the second half of this year, Trainium’s status in 2026 will be equivalent to TPU’s status in 2025. He said he had invested in TPU supply chain companies like Celestica, "I feel I have the qualifications to say this." He added: "I would never short Google, nor would I short Broadcom, but I do think Trainium is severely underrated right now." Another topic that drew attention in the conversation was "Orbital Compute"—the idea of placing data centers in space. Kh

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching1 items

2026-05-31

Nvidia early investor Gavin Baker's barbell strategy: long AI infrastructure, short the broader market

Similarity 170%關鍵字 gavin/baker同分類 zh

💡 Currently matching via keywords + symbols (MVP) · Will be upgraded to embedding semantic search later

Raw Information

ID:264a801b05

Source:動區 BlockTempo

Published:2026-05-25 06:41:20

Category:zh_news · Export Category zh

Symbols:Unspecified

Community Votes:+0 / −0 · ⭐ 0 Important · 💬 0 Comments