Inside the Mind of a Tech Investor with Gavin Baker

Summary

• The memory call: maybe don’t sell. Prices are up 60% and Micron’s margins are high-60s vs a historical average closer to 16% — and Baker, Micron’s analyst in 2000, admits “based on every memory cycle we have had for the last 25 years, this is the time to be selling memory. 100%.” Except one: the mid-90s, “the last true capacity cycle,” and on that template “we may still be very early.” On the numbers other Sohn panels gave: “I take the over on every number. Every single number.”
• Maybe no bubble this time — TSMC limits supply. Every profound technology bubbles, but AI has physical constraints past manias lacked: TSMC’s “flinty old men and women in their 70s,” guardians of Morris Chang’s legacy, may expand 5% while Jensen — visiting every 3 months — wants double or triple. If they double or triple, “Nvidia could probably sell 1, 1.5, 2 trillion dollars worth of chips next year. I really believe that.”
• He takes the over on OpenAI + Anthropic at $200B combined revenue near term. The engine: the shift from $250/month all-you-can-eat to usage-based pricing with frontier capability “locked behind harnesses” on enterprise plans — “wildly bullish” for token pricing, the same overage model that made cellular a great growth industry. Only “10 basis points of the world’s population” uses models properly amid an insane shortage; at 5% it’s “unimaginable.”
• Trainium is “by far” the most underestimated custom chip — Trainium 3’s ramp in the second half of this year will make Trainium to 2026 what TPUs were to 2025. Google made “very conservative design choices” with TPU V8, and the only two functioning switched scale-up networks for mixture-of-experts inference are Nvidia’s and Amazon’s. Google won’t submit TPUs to MLPerf — its own benchmark — which is “visibly driving Jensen crazy.”
• Neoclouds are durable, not a CapEx arbitrage. Running a cluster is Formula 1 — looks easy, would kill an amateur — and CoreWeave’s GPUs get utilized 2-3x more per hour than bottom-of-barrel providers, justifying the premium. His regret: The Trade Desk could have invested over $50M into CoreWeave at a $1.1B valuation but was conflicted out by Crusoe, which he says they still hold a large position in.
• The underappreciated short: terrestrial power and cooling industrials. Orbital compute becomes provably “possible, going to work, and economical in the next 2 years” and takes meaningful share by end of decade — the years before that will be “very painful” for industrial names that “massively flexed up capacity” for a build-out that “could really come to a screeching halt.”
• Process edge is overwhelmingly reading: he rarely meets public companies — “they never say anything that’s not in a transcript or 10-Q. And I can read much faster than they can speak.” And even non-coders get better investment answers from Claude Code or Code X than the regular model: coding may be “the ultimate AI app and it subsumes more and more.”

Deep dive

1. The production function: read, rarely meet

• The Peter Lynch DNA Baker kept from Fidelity: “if you like the store or you like the product, you’re going to love the stock.” The Lynch rule he still fights — sell losers, ride winners: “I’m extremely valuation sensitive, I’m very contrarian, I’m most comfortable on the 52-week low list. I’ve been hanging on to the memory stocks for dear life.”
• His edge is “overwhelmingly reading” — transcripts and primary sources, plus expert transcripts as “a great use of AI.” Companies “never say anything that’s not in a transcript or 10-Q. And I can read much faster than they can speak.” Lightning round: pattern recognition and sleep underrated, scuttlebutt overrated; on sizing, “you can either be a slugging percentage player or a batting average player” — pick your game and stay consistent.
• The scar tissue: Accredo Health fell 90-95% and never recovered on a NYT front-pager he says was “simply not true” — “there are always unknown unknowns.” And Nextel International, where he wrote his only-ever board letter demanding a buyback 15 months before bankruptcy after a price war between two larger competitors: “be very very careful of high leverage.”

2. Every memory cycle says sell — except the one that matters

• Memory prices up 60%, Micron margins high-60s vs a historical average closer to 16% — and the veteran’s candor: “based on every memory cycle we have had for the last 25 years, this is the time to be selling memory. 100%.” He was the Micron analyst in 2000.
• The exception is the mid-90s, “the last true capacity cycle” — and if this is the first true capacity cycle, “we may still be very early.” A glut will come eventually, but “eventually is doing all the work in that sentence.” On prior panels’ forecasts: “I take the over on every number that they gave.”

3. TSMC’s flinty old men are the bubble insurance

• Baker’s framing: every profound technology — railroads, canals, PC, internet — bubbles because of Mauboussin’s “breakdown in diversity,” and the bubble funds the build-out. This time, shortages of watts and wafers may prevent it: “smoother for longer is what we all want.”
• The wafer shortage persists because TSMC is “run by flinty old men and women in their 70s” guarding Morris Chang’s legacy — the people who, told 20 years ago that catching Intel was “a beautiful dream… probably for our grandchildren,” did it in one lifetime. Jensen visits every 3 months wanting capacity doubled or tripled; they may expand 5%. If they complied, “Nvidia could probably sell 1, 1.5, 2 trillion dollars worth of chips next year” — but “the other side of that might be very painful for everyone.”

4. Token maxing and the $200B question

• Jas’s pushback: if OpenAI + Anthropic hit $200B combined revenue in 12-18 months, does every S&P 500 company miss earnings on token spend? Jensen’s GTC goal — top engineers spending a minimum of half their comp on tokens — is untenable against S&P wage bills “without significant adjustments to the labor force.”
• Baker’s resolution: the shift to usage-based pricing. A $250/month subscription no longer buys frontier capability — “the best capabilities are locked behind harnesses and reserved for people on enterprise plans.” That’s “wildly bullish” for token pricing, the cellular-overage model that made telecom a great growth industry. And “if you’re not token maxing, you should be.”
• Echoing Alex’s scale check: “10 basis points of the world’s population is using these models the way they should probably be used” amid an insane shortage despite trillions spent — at 5%, “it’s unimaginable,” which is why orbital compute is a necessity. Coding may be the shortest path to ASI and AGI — “not just the killer app… the ultimate AI app and it subsumes more and more.”

5. Trainium’s year; neoclouds drive Formula 1 cars

• “Trainium by far” is the most underestimated custom chip: Trainium 3’s ramp in the second half of this year will make Trainium to 2026 what TPUs were to 2025. Google’s TPU V8 made “very conservative design choices,” and mixture-of-experts inference needs a switched scale-up network — only Nvidia’s and Amazon’s exist. Google won’t submit TPUs to MLPerf, its own benchmark. Caveat kept: “I would never bet against Google… never bet against Broadcom”; TPU bulls should show 13F positions in Momentum or Elastica.
• Neoclouds are durable, not transitional: running a cluster is Formula 1 — “if I tried… I would die” — and CoreWeave GPUs are utilized 2-3x more per hour than bottom-of-barrel providers. His 2005 retail analogy: running 1,000 clean, well-staffed stores across 50 states minted $50B market caps and only ~10 companies ever did it; clusters are harder.
• The regret told against himself: The Trade Desk could have invested over $50M into CoreWeave at a $1.1B valuation but was conflicted out by Crusoe — still a large, loved position — while Jas’s Blackstone later invested $7.5B.

6. Orbital compute and the unpriced short

• Timeline: orbital compute becomes clearly “possible, going to work, and economical in the next 2 years,” taking meaningful share “towards the end of this decade.” Built terrestrial data centers are always going to be valuable for training and RL, and he “can’t imagine” never building another in the next 7 years — but the interim is “very painful” for power and cooling industrials that flexed up for a build-out that “could really come to a screeching halt.”
• The hardware as described: a sun-synchronous satellite that is just a rack — 8ft tall, 2.5ft wide, 4ft deep — with a 300-400 ft radiator in its own shadow, laser-stitched into a virtual data center. Power from the sun, cooling from the dark side.