Vercel CEO Guillermo Rauch posted a chart (relayed here from a screenshot, since this environment can't fetch x.com directly) showing open-weight models' share of tokens processed through Vercel's AI Gateway hitting a record 62% on August 22, up from 28.4% roughly two months earlier on June 24. Rauch's own read on why this keeps climbing, stated as his own view rather than as measured data: "this is very likely just the start, because enterprise adoption is still early, and harnesses, CLIs, IDEs, SDKs, etc need to be adapted to be model agnostic" — a claim about infrastructure catching up to model choice, not about model choice itself, and one this post can't independently verify. Investor Gavin Baker picked the chart up separately and framed it as a bigger story than the raw percentage: "more data than open-source AI is taking share from OpenAI and Anthropic," and — worth taking at face value rather than skimming past — his own read is that this is good for AI infrastructure demand, not evidence of it shrinking.
What's actually being measured here
Vercel's AI Gateway is a routing layer that developers use to call different model providers through one API, and the company publishes a recurring "AI Gateway Production Index" report with real disclosed scale behind it: tens of trillions of tokens a month, across more than 200,000 unique teams, built on seven months of production traffic. The July edition of that index reported open-weight models at 29% of token volume, up from roughly one-ninth of total volume in April — a slower-moving, independently published number that lines up directionally with the newer 28%-to-62% jump Rauch's chart shows for the two months since. Worth being precise about what this data is and isn't: it's real, measured usage on one company's gateway product, not a survey of the broader AI market, and Vercel's own customer base — developers building web and JS-heavy applications — isn't necessarily representative of enterprise, consumer chat, or on-prem usage elsewhere.
The July index names specific movers: DeepSeek at 22.6% of token volume, in third place and within two points of Google, and GLM 5.2 breaking into the gateway's top models by volume within two weeks of release — concrete, checkable claims about which labs are actually capturing this share, rather than "open source" as an undifferentiated category.
The volume number and the money number tell different stories
The detail that actually complicates the headline: per Vercel's own July index, open-weight models handled close to a third of all tokens on the platform while accounting for under 4% of total spending — a result of running at roughly one-tenth the average token price across the gateway. That gap is the real story underneath the volume chart, and it's also exactly the point Baker makes explicitly in his own thread: "nothing about open-source AI inference is 'free'" — an open-weight token costs a provider roughly the same compute to produce as a closed one, so the low price reflects thin or negative margins at the model layer, not free inference. Baker's own predicted end state, stated as his opinion rather than as data: closed, frontier tokens end up representing 60–90% of economic value while making up only 15–25% of token volume — a claim Vercel's own volume/spend split (a third of volume, under a twentieth of spend) is already roughly consistent with, several months before Baker's prediction.
What's opinion, and what's data, in Baker's thread
Worth separating clearly: the token-share chart and the July index numbers are Vercel's own measured data. Baker's broader claims layered on top are his own analysis, clearly framed as such — that OpenAI and Anthropic's combined volume also accelerated in July, meaning open-weight gains aren't simply cannibalizing the frontier labs but riding a larger wave of total demand; and his suspicion that Grok is growing "even faster than open-source," which he attributes to data from payments company Ramp that this post hasn't independently seen. Both are plausible extensions of a real dataset, not claims with the same evidentiary weight as the Vercel numbers themselves, and are worth reading as an investor's interpretation rather than as confirmed facts.
What to expect next
- Watch for Vercel's next published index. The 62%/38% figure comes from a live dashboard chart rather than a written report with the same disclosed methodology as the July index; a formal August edition would make the newer number as checkable as the one before it.
- Watch whether the volume/spend gap narrows or widens. If open-weight models keep taking token share while staying near a tenth of average price, that's a margin story for model providers more than a capability story — worth tracking as its own trend line.
- Watch for the Ramp data Baker references. A second, independent enterprise-spend dataset on Grok's growth specifically would be a real corroboration or complication of this post's numbers, not yet something this post could verify directly.
References: Guillermo Rauch on X — record open-weight token share · Gavin Baker on X — analysis thread · Vercel — AI Gateway Production Index, July 2026 · related coverage: A Rumor on a Podcast, a Denial on X, and Then Dario Amodei Showed Up · The Week Washington Flirted With Banning Open Weights · The Price War Nobody Is Actually Fighting · Frontier Arcade: trends & predictions