AI infrastructure companies are working closely in partnership with the top AI labs, but their interests towards open and cheaper models might be completely divergent.
That’s the thesis laid out by Gavin Baker, Managing Partner and CIO at Atreides Management, in an X post breaking down what he calls the “mega bull case” for AI infrastructure. Baker’s argument is simple on its face but has fairly large implications for how investors should be thinking about who actually captures the economic value being created by AI: the labs building the smartest models, or the companies running the pipes and chips underneath them.
Baker frames the setup this way: “The mega bull case for AI infrastructure would be if market share shifted away from certain frontier labs with 90%+ inference margins toward cheaper models, whether open-source or closed.” His reasoning is that cheaper, more efficient models increase the return on investment for companies actually spending money on AI, since they get more intelligence for every dollar spent. That improved ROI, in turn, drives more token demand overall, because usage tends to expand when the cost per unit of intelligence falls.
The redistribution that follows is the crux of the argument. If frontier labs are currently capturing outsized margins on inference, and that share moves toward cheaper alternatives, the money doesn’t disappear, it just changes hands. “Margin dollars would effectively get redistributed from the frontier labs to AI infrastructure providers,” Baker writes. Under this framework, the infrastructure companies that win are the ones with the lowest cost per token, while the model builders that win are the ones squeezing the most useful output out of every token they generate. Baker sums it up bluntly: “Lower margin % at the model layer = more margin $ at the infra layer all else equal.”
Why Jensen Huang keeps talking up open source
Baker connects this directly to NVIDIA CEO Jensen Huang’s repeated public enthusiasm for open-source models, an enthusiasm that has sometimes puzzled observers given NVIDIA’s fortunes are tied to the AI labs spending enormous sums on its chips. Baker offers an explanation: “There are many reasons Jensen is so focused on open source, but this is likely the most important one as I think he is probably less worried about a monopsony these days.”
The monopsony concern is worth unpacking. A monopsony is a market with a single dominant buyer, and NVIDIA’s fear in earlier years of the AI buildout was that if two or three frontier labs ended up controlling the vast majority of AI usage, those labs would have outsized leverage over NVIDIA as a supplier. A world with many competitive model providers, open or otherwise, spreads that buying power across a wider base and reduces the risk of any single customer dictating terms. Encouraging open-source adoption also serves NVIDIA’s stated worldview that compute demand should keep expanding regardless of which lab or model ends up on top, since open models tend to proliferate widely and get deployed on a broad range of hardware.
SpaceX and Meta’s vertical integration edge
Baker’s post also flags a structural shift already underway: two of the biggest players in AI infrastructure are simultaneously building some of the best models in the industry. “With SpaceX and Meta being vertically integrated and possessing the #3 and #4 models respectively it is more possible than ever,” he writes.
The SpaceX reference is to Grok, which now sits under SpaceXAI following SpaceX’s absorption of xAI earlier this year. Grok 4.5 landed fourth on the Artificial Analysis Intelligence Index, behind only Claude Fable 5, GPT-5.5, and Claude Opus 4.8, while undercutting most of its rivals on price. Baker goes further than the official ranking, though. “Note that Grok 4.5 is ahead of Fable for some useful tasks at a much lower cost, so ranking them #3 is conservative,” he writes, suggesting the standard leaderboard position understates where Grok actually sits once cost-adjusted usefulness is factored in.
Meta occupies a similar position with its own frontier model efforts, giving it a foot in both camps: a hyperscaler with massive infrastructure spend and, at the same time, a lab producing models competitive enough to matter at the frontier. For companies structured this way, the redistribution Baker describes isn’t purely theoretical. A dollar of margin pulled away from an outside frontier lab and pushed toward efficient, lower-cost inference can land inside the same corporate balance sheet, rather than needing to travel to a separate infrastructure vendor at all.
The redistribution hasn’t happened yet
Baker is careful to note that this remains a forward-looking scenario rather than a description of where the market stands today. “This is not happening yet,” he writes. “Cheap, mostly open source tokens are likely the majority of volume today but the majority of economic value is still accruing to the most intelligent models.”
That distinction between volume and value is the piece doing the most work in his argument. Plenty of token volume already runs through cheaper and open-source models for tasks that don’t require frontier-level reasoning, but the highest-value use cases, the ones enterprises will pay a premium for, are still concentrated with the handful of labs producing the most capable models. Baker’s bet is that this gap narrows as cheaper models close the capability distance, pulling economic value down toward the infrastructure layer along with the volume that’s already there. He closes with appropriate hedging: “Might change though. We will see.”
The framing gives investors a cleaner way to separate two things that often get bundled together in AI market commentary: excitement about intelligence gains at the model layer, and excitement about infrastructure spend. Baker’s post suggests these two trends could eventually pull in opposite directions when it comes to who keeps the margin, even as both are, for now, growing side by side.