Read the source · this Framework is our reading of their public work · Atreides Management →

Synthos Research · Frameworks · Thesis Snapshot

Gavin Baker: how he actually thinks

This is not a profile. It is a working model of Gavin Baker's worldview — his causal beliefs, his own stated triggers, what he changed his mind about (dated), and where he is silent — reconstructed from his claims as a guest across other people's shows (2020 → 2026) and tested against a holdout window the model never saw. Baker is the council's AI-infrastructure specialist — and the Framework that taught us the most by failing. It failed its first eval on an integrity guardrail, failed its second because his vendor-level calls flip faster than a frozen model, and passed its third under a new two-lane eval built for exactly that: 70.6% direction overall, with the doctrine and the trades graded — and disclosed — separately.

A Framework may only believe what the claim record can prove · what a Framework is · how voices earn tracking · methodology

direction fidelity: 70.6%
structural lane: 64.3%
tactical lane: 100%
evals to pass: 3

Model fidelity: 70.6% direction — the first Framework graded on two lanes, after two disclosed failures

We froze this model on his claims through April 19, 2026 and graded its predictions against what he actually said over the following three months. This is the third eval of this Framework, and the history is part of the result: the first failed when the predictor fired mechanisms that adversarial verification had killed (an integrity trip, not a fidelity score); the second scored 50% because his vendor-level stances flip in weeks — the model held his old Google-TPU bullishness in the exact window he turned contra-consensus bearish. The fix was not a softer bar. It was a sharper instrument: two-lane grading — doctrine (structural) gates the pass; fast-flipping calls (tactical) are predicted with flip-risk awareness and disclosed separately.

12 of 17 gradeable predictions got his direction right — 70.6%, above the 60% bar even under the old strict rule. Structural lane: 9 of 14 (64.3%), which is what gates the pass. Tactical lane: 3 of 3, reported but never gating.
Zero killed-edge firings this time — and the best proof the guardrail is real cost us a point: the model went cautious on a neo-cloud earnings event because the mechanism that would have called it correctly had been killed in verification. It took the miss rather than use machinery it wasn't allowed to trust. That trade — a lower score for a cleaner model — is the whole discipline.
The strongest hit: on the Vera Rubin production announcement the model called it bullish-Nvidia, deepening the systems lead — direction, mechanism and vocabulary all matching his actual claims.
Correct silences: 95% — macro prints and Fed events correctly drew no answer; this model has no macro machinery and invented none.
Thin invariant layer, disclosed: one formal invariant. His durable beliefs live in the edge structure; the corpus (guest appearances, not his own show) is dense on mechanisms, light on repeated credos.

1 · How the world works, according to Baker

A scaling-laws worldview with a value-investor's price discipline bolted on — the two halves of his career in one graph.

Scale wins — and compute is the whole game

The doctrine everything else hangs from

Ten times the training data means requiring proportionally more compute (CPUs, GPUs, memory, networking), so the largest companies with the most compute and data win. The investable edge sits underneath: storage, memory and networking that raise GPU utilization are the key investment area.

Without proprietary data, models are commodities — his one formal invariant

The only belief stated often enough to formalize — and it is ruthless

Foundation-model companies without unique proprietary data become commodities — 2023–2026

most model startups are zeros with zero chance — 2023–2026

Nvidia cannot be beaten head-on

Held since his own 2024 bear-to-bull reversal (dated below)

you can't out-Nvidia Nvidia in a head-on assault — 2024–2026

AI is an extinction event for seat-based software

The bearish half of his AI thesis — and his fastest-burning fuse

as AI replaces human labor, per-seat (per-human) revenues shrink; meanwhile larger context windows also killed enterprise fine-tuning — and the timeline he puts on it is tactical, not structural: Software CEOs have 2-3 months to adapt or die.

Price paid determines return — the Fidelity half

The discipline that keeps the AI maximalism honest

price paid determines return — though paying 40x can still work if earnings compound fast as margins later expand. And on why classic value stopped working: it was arbitraged away once emotionless quants entered.

2 · His strongest causal chains

The scaling cascade

Data scales → compute scales: requiring proportionally more compute (CPUs, GPUs, memory, networking)
Compute concentrates → winners concentrate: so the largest companies with the most compute and data win
Cheaper inference doesn't break it — it feeds it: massively increases AI ROI via Jevons paradox

The capex-race escalation ladder

The race feels existential: because winning LLMs feels existential
So the giants climb the ladder: issue debt, then cancel dividends, then stop buybacks — his pre-registered sequence for what desperation looks like, step by step, watchable in real time.

3 · What would change his mind — his own stated triggers

The Nvidia tell

Receivables vs revenue

His own stated warning light: If the AR-growing-faster-than-sales trend continues past the July quarter, that's reason for concern. A precise, dated, checkable trigger on his highest-conviction name.

The ROI gate

Agents must materialize

The bull case on AI capex is gated on agents shipping — and the rate-limiting factor is compute. Agents failing to materialize breaks the Blackwell-ROI leg.

A dated tech call

Optics move inside

Coherent optics will have to come inside in 2025, no later than 2026 — a falsifiable infrastructure prediction with an expiry date on it.

The escalation ladder

Debt, dividends, buybacks

Each rung of issue debt, then cancel dividends, then stop buybacks is itself a signpost — a mega-cap canceling its dividend for capex would confirm the existential-race read at high conviction.

4 · What he changed his mind about — dated

Three recorded reversals — including full flips on the two biggest names in his universe. This is why his Framework needed the two-lane eval: the doctrine holds still while the name-level stances move.

Then · Feb 2024Bearish Nvidia: capital and compute become commodity, someone builds a competitor GPU

Now · Jul 2024Bullish — the systems-and-software lead means head-on competition fails

A full reversal in ~5 months with no stated trigger in the record — the revision style the eval had to learn to grade.

Then · Feb 2024Bearish Google: obviously, massively more threatened by AI, mismanaged

Now · Nov 2024Bullish — owns the TPUs and the YouTube data

And then, inside our test window, he flipped again — contra-consensus bearish on TPUs while backing Amazon's Trainium. Name-level conviction with him is a moving target by design; that finding is now built into how we grade him.

Then · Nov 2021early stages of a powerful new blockchain-driven cycle

Now · Nov 2024Conviction decayed to a shrug — not reversed, just quietly abandoned

5 · Where he is silent

Macro forecasting — never asserted: no recession timing, no Fed-path-to-multiple machinery (actively rejected), no labor-data inputs. In eval, macro events correctly drew silence at 95% precision.
Money printing → dollar devaluation — never asserted; the council's debasement debate simply does not run through him.
Antitrust / breakup risk — never asserted, even for the mega-caps he covers daily.
Traditional valuation machinery — DCF, WACC, beta: actively rejected. Price discipline for him is entry-multiple versus earnings compounding, not discount rates.
One retail exception: the physical-store mechanism — they lower CAC, improve LTV/CAC, cut fulfillment costs — a leftover from his consumer coverage that the graph keeps because he still asserts it.