Nvidia, Meta, and xAI Are Winning the AI Efficiency Era — Here’s Why the Margin Is Moving Down the Stack

The AI race is shifting from biggest model to cheapest intelligence — and investor Gavin Baker argues that’s the most bullish possible outcome for infrastructure, not the death of AI value.

What Happened

A CNBC report published July 10 crystallized a structural shift the industry has been feeling for months: the AI competition is no longer about who can train the biggest model. The new battleground is routing, cost, and control — systems that decide which model to use, when, and what company data or tools to bring in. Smaller, task-tuned models are routinely outperforming large general-purpose ones on specific tasks, and hybrid local-plus-cloud deployment is moving from experiment to enterprise standard.

Peter Fenton of Benchmark put a number on the trajectory: he predicts more than 90% of all tokens will come from open-weight models within 18–24 months — possibly by year-end 2026. If that forecast lands anywhere near correct, frontier labs running 90%-plus inference margins would face real structural pressure, as companies run capable models without paying provider markups. The shift isn’t speculative noise; it’s the one data point AI bulls and bears have both co-signed.

That rare consensus is where the debate gets sharp. AI bear Michael Burry — who has publicly shorted the AI trade including Nvidia and derided what he calls “tokenmaxxing” as a bubble — read the efficiency shift as validation of his thesis: frontier infrastructure is overvalued if inference margins collapse. Investor Gavin Baker then quote-tweeted Burry to argue the precise opposite. Same shift, inverted conclusion. Baker’s read is the bull case worth unpacking carefully.

The Structural Read

Baker’s argument has a precise internal logic, and it’s worth laying out exactly — because it’s easy to misread as simple AI optimism. It isn’t. It’s a specific claim about where in the stack value accrues when the model layer commoditizes.

The mechanism runs like this: if frontier-lab inference margins compress from 90%+ toward something lower — because open-weight or cheap closed models take volume share — the per-unit cost of intelligence falls. Cheaper intelligence means better ROI for end customers. Better ROI means companies deploy AI into more workflows, more tokens, more volume. The Jevons paradox applied to compute: efficiency doesn’t shrink the market, it expands total demand. Margin dollars don’t disappear; they redistribute. They flow away from the model provider and toward whoever has the lowest per-token infrastructure cost at scale. That is, structurally, Nvidia’s toll-booth position.

Baker also names the vertically integrated players — SpaceX/xAI and Meta — as structurally well-positioned: they own both the model and the distribution, run cheaper models that are competitive on capability, and don’t need to extract margin at the model layer because they monetize elsewhere. Jensen Huang’s consistent push for open source, Baker argues, is rational self-interest: lower model-layer margin percentages mean more inference compute overall, and Nvidia captures a toll on all of it.

That hedge is load-bearing. Baker is presenting a conditional thesis, not a realized fact. The Jevons dynamic and the margin-migration story are the would-be consequence if open-weight volume dominance materializes and cheap tokens drive enterprise deployment at scale. Right now, by his own read, the economic value in AI still pools at the frontier. The thesis is a forecast about where the gravity shifts — not a description of where it has already shifted.

Three Implications

How the AI Stack Reprices (If the Shift Lands)