Founder, Managing Partner & CIO of Atreides Management · ex-Fidelity OTC Portfolio — running synthesis of his video/podcast appearances, with per-transcript breakdowns and a stock index.
Trainium is "by far" the most underestimated custom silicon — aggressive design + one of only two functioning switched scale-up networks (the other is Nvidia's).
Neocloud category is "absolutely durable": CoreWeave's GPUs run 2–3× harder per hour, so it charges a premium. Regrets being conflicted out of investing at a $1.1B valuation.
Hangs onto memory "for dear life" — this may be the first true capacity cycle since the mid-90s, the one cycle you don't sell; takes the over on every bullish memory number.
Was very early (talking to Jensen in 2000); competition is coming but Nvidia stays the dominant part of the market — one of only two working switched scale-up networks.
He'd "take the over" on OpenAI + Anthropic combined hitting $200B of revenue in maybe 12–18 months — code generation was the killer app to monetize AI.
The flinty custodians of Morris Chang's legacy ration wafer capacity (~5% vs Jensen's "double or triple") — the real-world physical brake that helps the AI build-out avoid a bubble.
Mixed: TPU V8 made "very conservative design choices" and won't enter its own MLPerf — but "never bet against Google," and TPU V9 "is going to be amazing."
Historical reference: 20+ years ago TSMC staff thought catching Intel was "a beautiful dream… for our grandchildren" — and did it in one lifetime. Passing mention.
A missed-deal war story: The Trade Desk "could have invested over $50 million" in the CoreWeave round at $1.1B; he was conflicted out. No stance on the stock.
His other formative loss: a large, over-levered position; he wrote his only-ever board buyback letter ~15 months before bankruptcy — the high-leverage lesson.
In one line: a deep-tech / semiconductor specialist who reads primary sources at scale; his core call is that AI is supply-constrained by watts and wafers — and that TSMC's deliberate capacity discipline is what may let the AI build-out avoid a classic bubble.
Watts & wafers, not hype, gate the AI build-out. The binding constraints are power and chip-making capacity. TSMC's veterans deliberately ration wafer capacity (guarding Morris Chang's legacy / Taiwan's stability), expanding ~5% a year vs the "double or triple" Jensen wants — a real physical brake that may let this cycle avoid a classic bubble ("smoother for longer").
Memory looks like a capacity cycle, not an inventory cycle. Every memory cycle of the last 25 years says sell at these margins — except the mid-90s, the last true capacity cycle. He thinks this may be that one: "the cycle where you absolutely do not want to sell." Hangs onto Micron "for dear life" and takes the over on every bullish memory number.
Custom silicon: Trainium > TPU. Amazon's Trainium is "by far" the most underestimated — aggressive design + one of only two working "switched scale-up networks" (the other is Nvidia's) needed to inference mixture-of-experts models. Google's TPU V8 made "conservative" choices and won't enter its own MLPerf — though he'd "never bet against Google… or Broadcom," and TPU V9 will be "amazing."
Neoclouds are durable; utilization is the moat. CoreWeave, Crusoe (large position), Nebius, Lambda. Running a cluster is "like driving a Formula 1 car"; the best run GPUs 2–3× harder per hour, earning a real, lasting premium — not a commodity.
Orbital compute solves the watt shortage by the end of the decade. Possible/economical within ~2 years, meaningful share by ~2030. The painful late-decade short: terrestrial power/cooling industrials that "massively flexed up" for a build-out that could "come to a screeching halt."
Process > access. "Reading is overwhelmingly the most important part" — primary sources (transcripts, 10-Qs, expert calls) at scale, not management meetings ("I can read much faster than they can speak"); pattern recognition and being early to correct frameworks; size down high leverage (the Nextel lesson).
Transcripts
One dated page per appearance — each has its full stock table, talking points, and the saved transcript. Newest first.
Gavin Baker (Atreides Management) appearances discovered via YouTube search (Gavin Baker), not yet processed — newest first. Limited to the last ~2 years.