A venture investor is pointing to one platform’s internal chart as evidence that open-weight models are pulling meaningful share away from the two best-funded closed labs. Gavin Baker, posting on X on August 23, 2026, said open source token share at Vercel rose from 28 percent to 62 percent over the prior two months. He credited the underlying chart to Vercel co-founder Guillermo Rauch.
This is a different claim from the open-weight adoption figures AI Insiders has previously covered on gateway-wide token share and production traffic. Those numbers described aggregated activity across many routing services. Baker’s 28-to-62 figure describes usage at a single company, drawn from a chart that company’s own leadership produced. Nothing here has been verified by an outside measurement firm, and Vercel has not published methodology alongside the number.
Baker called the shift notable given that combined OpenAI and Anthropic token volume also accelerated in July. His interpretation: total demand for tokens and the infrastructure behind them grew faster than the frontier labs’ own growth, meaning open models are not simply cannibalizing existing usage but expanding the overall pool. He also said he suspects xAI’s Grok is gaining share even faster than open-weight models broadly, pointing to similar patterns he says appeared in expense-management platform Ramp’s data. Ramp’s figures were not detailed in his post.
The economic argument is the part worth sitting with. Baker’s position is that open source gaining share is good news for compute providers specifically because it squeezes what model sellers can charge, even as the silicon and power needed to generate each token stay fixed: producing an open-weight token, in his framing, burns the same compute a closed frontier one does. “Nothing about open-source AI inference is ‘free,’” he wrote. That logic reframes rising open-weight usage as a demand signal for chips and data centers rather than a threat to the AI buildout.
Baker also offered a forecast rather than a fact: he expects closed frontier models to eventually account for 60 to 90 percent of the economic value generated by AI tokens while representing only 15 to 25 percent of tokens processed. That is his prediction, not a reported outcome, and Baker discloses elsewhere that he invests in AI companies, so the argument doubles as a case for why infrastructure and even open-weight-adjacent bets remain attractive to him.
Readers should treat the 28-to-62 figure as a single vendor’s snapshot, not a market census. Anyone sizing inference demand for the next quarter should ask their own gateway or hosting provider for a comparable breakdown before extrapolating Vercel’s curve onto the broader market.
Gavin Baker, posting on X on August 23, 2026.
