| Ticker | Name | Research | View | What he said | At |
|---|---|---|---|---|---|
| MU | Micron Technology | QT · SA · STK · FA | Positive | "Hanging on to the memory stocks for dear life." Every prior cycle said sell, but this may be the first true capacity cycle since the mid-90s — the one cycle you absolutely do not sell; takes the over on every bullish memory number. | 8:30 |
| NVDA | Nvidia | QT · SA · STK · FA | Positive | He was very early (talking to Jensen back in 2000). Competition is coming, but Nvidia "is still going to be the dominant part of the market"; one of only two working switched scale-up networks. | 16:33 |
| TSM | Taiwan Semiconductor | QT · SA · STK · FA | Positive | The "flinty old men and women" guarding Morris Chang's legacy won't expand as fast as Jensen wants (maybe +5% vs his "double or triple") — enforcing the real-world wafer constraint that helps everyone avoid a bubble. | 10:46 |
| AMZN | Amazon | QT · SA · STK · FA | Positive | Trainium is "by far" the most underestimated custom silicon — aggressive design choices, and Amazon's Trainium runs one of only two functioning switched scale-up networks in the world (the other being Nvidia's). | 16:57 |
| CRWV | CoreWeave | QT · SA · STK · FA | Positive | The neocloud category is "absolutely durable." Their GPUs are utilized 2–3× more per hour than a bottom-of-the-barrel provider's, so they can charge a real premium. Regrets being conflicted out of investing at a $1.1B valuation. | 19:51 |
| NBIS | Nebius Group | QT · SA · STK · FA | Positive | Named among the "high-quality neoclouds" (with CoreWeave and Crusoe) whose superior GPU utilization is a durable, underappreciated competitive advantage. | 21:05 |
| CLS | Celestica | QT · SA · STK · FA | Positive | Cited as one of "the best ways to invest in TPUs" (a TPU hardware supply-chain play) — "I owned one of those." [Transcript "Elastica"; mapped to Celestica.] | 17:15 |
| GOOGL | Alphabet (Google) | QT · SA · STK · FA | Neutral | Made "very conservative design choices" with TPU V8 and won't submit TPUs to its own MLPerf benchmark — but "I would never bet against Google," and TPU V9 "is going to be amazing" with more aggressive choices. Mixed → Neutral. | 17:21 |
| AVGO | Broadcom | QT · SA · STK · FA | Neutral | "I would never bet against Broadcom" (Google's TPU design partner) — a respectful aside, not a direct pick in this conversation. | 18:32 |
| INTC | Intel | QT · SA · STK · FA | Neutral | A historical reference point: 20+ years ago TSMC staff thought catching Intel was "a beautiful dream… probably for our grandchildren" — and did it in one lifetime. Passing mention, no stance. | 11:05 |
| TTD | The Trade Desk | QT · SA · STK · FA | Neutral | A missed-deal war story: "Trade Desk could have invested over $50 million" in the CoreWeave round at $1.1B — he was conflicted out and still gets "a little sad" about it. No stance on the stock. | 19:27 |
| BX | Blackstone | QT · SA · STK · FA | Neutral | The interviewer's firm (it put $7.5B behind CoreWeave). Noted only in the orbital-compute aside — terrestrial-data-center exposure would be "a lot of trouble with Blackstone." No stance. | 18:53 |
| Atreides Management | Atreides Management (his fund) | — | Neutral | His tech/deep-tech hedge fund, named for Dune ("the spice must flow" — spice = high-bandwidth memory). Run from Boston. The lens for every view here. | 0:00 |
| OpenAI | OpenAI | — | Positive | He'd "take the over" on OpenAI + Anthropic combined hitting $200B of revenue in maybe 12–18 months — code generation turned out to be the killer app to monetize AI. | 12:33 |
| Anthropic | Anthropic | — | Positive | Paired with OpenAI in the "take the over on $200B combined revenue" call; Claude Code cited as giving better answers than the regular model even for investing questions. | 12:33 |
| Crusoe | Crusoe Energy | — | Positive | "I love Crusoe… we have a large position in Crusoe." A high-quality neocloud whose GPU-utilization premium he expects to make the business durable. | 19:36 |
| Lambda | Lambda | — | Positive | Listed alongside CoreWeave / Crusoe / Nebius as part of the durable neocloud category. | 19:30 |
| xAI | xAI (Grok) | — | Neutral | Mentioned only as a "favorite LLM" to send orbital-compute papers through to distill ("Grok, etc."). No investment view. | 22:38 |
| Accredo Health | Accredo Health | — | Negative | A formative bad trade (2011–12): a NYT front-page exposé sent it down 90–95% and it never recovered — the "unknown unknowns" / always-priced-in-risk lesson. | 3:28 |
| Nextel International | Nextel International | — | Negative | His other formative loss: a large, over-levered position; he wrote his only-ever board letter demanding a buyback, and the company went bankrupt ~15–18 months later. The "be very careful of high leverage" lesson. | 3:52 |
"View" is Gavin Baker's stance in this conversation (Positive / Neutral / Negative), not a price rating. Research: QT Qualtrim · SA Seeking Alpha · STK Stock Analysis. Private names get no ticker (Research —). The transcript's "Elastica" is read as Celestica (CLS); the paired "Momentum" name is unresolved and intentionally omitted (no ticker invented). He also references "a prominent potential provider of orbital compute" he's "not supposed to" name — not turned into a ticker.
A jargon-free summary of the thesis behind each pick — what it actually is and why he holds that view. (Plain-language companion to the table above; renders on each ticker's consolidated page.)
Micron makes memory chips (DRAM and the "high-bandwidth memory" that AI servers crave). Memory is famously cyclical: prices boom on a shortage, then crash when everyone builds capacity and a glut arrives. After 25 years of cycles, Baker says the playbook normally screams "sell" right about now — margins are at record highs.
But he's "hanging on for dear life" because of one exception. He thinks this is a true capacity cycle — like the mid-1990s — where demand structurally outruns the industry's ability to add supply, not just a normal inventory swing. In a capacity cycle the up-leg lasts far longer, so selling early is the mistake. He takes the "over" on every bullish memory forecast.
Nvidia makes the GPUs at the center of the AI boom. Baker was extremely early (he was talking to CEO Jensen Huang about chips back in 2000). He acknowledges rivals are coming, but expects Nvidia to stay "the dominant part of the market."
A key technical edge: to run today's biggest AI models efficiently you need a special kind of chip-to-chip plumbing he calls a "switched scale-up network," and Nvidia operates one of only two that actually work in the world (Amazon's Trainium has the other).
TSMC is the factory that physically manufactures almost all the world's advanced AI chips. Baker's whole "no bubble" thesis runs through it: the people who run TSMC are disciplined veterans who see themselves as guardians of founder Morris Chang's legacy and protectors of Taiwan, so they deliberately refuse to expand capacity as fast as customers like Nvidia demand.
That restraint is a real, physical limit on how fast the AI build-out can grow — a brake that didn't exist in past tech manias. By rationing the supply of chip-making capacity ("wafers"), TSMC keeps the boom from overheating into a classic bubble. That makes it both a beneficiary and the safety valve.
Beyond its retail and cloud businesses, Amazon designs its own AI chip, "Trainium," as an alternative to buying Nvidia. Baker thinks Trainium is "by far" the most underestimated of the custom AI chips because Amazon made aggressive, ambitious design choices (where Google played it safe with its competing TPU).
The clincher: running the newest AI models (which split work across many specialist sub-models, called "mixture-of-experts") requires that special chip-networking he describes — and only two such working systems exist worldwide, Nvidia's and Amazon's Trainium. That scarcity is why he rates Amazon's silicon so highly.
CoreWeave is a "neocloud" — a company that rents out GPU computing power to AI firms, instead of being a giant like Amazon or Microsoft. The common skepticism is that this is a commodity business with no lasting advantage.
Baker disagrees. He says running a GPU cluster well is "like driving a Formula 1 car" — it looks easy but is extraordinarily hard, and the best operators get their GPUs used 2–3× more productively per hour than weak rivals. That efficiency lets them charge a premium, which makes the advantage durable. He regrets being legally blocked from investing back when CoreWeave was valued at just ~$1.1 billion.
Nebius is another neocloud, which Baker groups with CoreWeave and Crusoe as "high-quality" operators.
Same thesis as CoreWeave: the durable edge is operational excellence — getting GPUs utilized far harder than a "bottom of the barrel" provider — which lets the best names charge more and earn a lasting premium, something he thinks the market underappreciates.
Celestica builds the hardware (servers, racks, electronics) that goes into data centers. Baker cites it as one of "the best ways to invest in TPUs" — that is, a supply-chain way to benefit from the build-out of Google's custom AI chips without owning the chip designer itself — and says "I owned one of those."
(Note: the auto-transcript renders the name as "Elastica," which is almost certainly Celestica given the TPU-hardware context. The paired "Momentum" name he mentions is unresolved and not included as a ticker.)
Google designs its own AI chip, the TPU, as an alternative to Nvidia. Baker's view is genuinely mixed. On the cautious side: he thinks Google made "conservative" design choices with the current TPU (V8), and pointedly notes Google won't submit its TPUs to MLPerf — the very chip-benchmark Google itself created — which he reads as a tell.
On the positive side: "I would never bet against Google," and he expects the next chip, TPU V9, to be "amazing" with more aggressive design. Net-net it's a balanced, wait-and-see stance rather than a clear buy or sell — hence Neutral.
Summary & timestamps derived from the public YouTube video (transcript in transcript.txt) for personal study. Not investment advice. © Sohn Investment Conference / Atreides Management for source material.