Gavin Baker: Open-Weight Models Will Dominate Tokens, Favoring AI Infrastructure Over Model Labs
Summary
- • Investor Gavin Baker predicts a hybrid multi-model future with open-weight tokens the overwhelming majority
- • He argues model-layer fragmentation and lower margins raise compute demand and favor infrastructure
- • He says below-market 2024-25 compute contracts will reprice higher, accelerating cloud and neocloud revenue
- • He names a major cyber event as the most plausible near-term bear case
Details
Baker predicts a hybrid multi-model future where open-weight tokens are the overwhelming majority of tokens
This is an investor's forecast, not an established outcome. He expects dominant frontier labs to still do well but perhaps not be as dominant as seemed possible in May.
Baker argues fragmentation and lower model-layer margins benefit AI infrastructure and applications
His reasoning is that cheaper, more competitive models push down what the model layer can charge. That shifts value toward the companies supplying compute and the companies building on top of models.
Baker contends lower margins drive higher compute demand as long as someone pays for training
He describes Nvidia's Jensen Huang as the 'trainer of last resort', implying Nvidia could underwrite frontier training if other funders pull back.
Baker expects below-market compute contracts signed in 2024-25 to roll off and reprice higher
He argues this should accelerate cloud and neocloud revenue. He also suggests external financing for the buildout may be near its peak.
3Fourteen Research estimates B200 clusters earn a 37% IRR even if rental prices fall 40% per year
Baker cites this to argue GPU cluster economics stay attractive despite steep price declines. It is a third-party model estimate, not a realized return.
Baker says Anthropic's $413B of contractual commitments equals roughly $31B in annual cloud spend
He argues this is internally financeable given Anthropic's reported ARR. He was responding to the weekend MOU controversy, and the figures are his own estimates.
Baker says $161B of the commitments sits with Broadcom, financed by Broadcom and private credit
In his framing, this portion does not depend on Anthropic's own balance sheet, which he uses to downplay concern about the commitment size.
Baker argues fragmentation is best for humanity and anything slowing the buildout is net positive
He says actual overbuild is the biggest risk, so he sees slowdowns as protective. He also describes the market as structurally short watts, wafers, and capital.
Baker names a major cyber event as the most plausible near-term bear case for AI
He adds that AI cyberdefense is ramping, which he presents as a partial offset to that risk.
Insight = investor's argument or forecast, Market Impact = effect on industry segments, Financials = funding and spending figures, Research = third-party estimate, Strategy = positioning view, Security Alert = risk scenario
What This Means
This is one prominent investor's view, not a settled outcome. Baker argues that if open-weight models carry most AI usage, value shifts from a few dominant model labs toward the companies that supply compute and build applications. His case rests on assumptions that can be tested: that compute contracts reprice higher, that cluster returns hold up as rental prices fall, and that someone keeps funding training. Readers should treat his Anthropic commitment figures as his own estimates rather than disclosed financials.