Models & Open Source
$26.5B Chases 6% of Businesses on Open Weights
Three deals put $26.5B into open-weight AI infrastructure. Only 6.1% of AI-using businesses touch these platforms — about $4.3B per point of adoption.
Three deals in three weeks have put roughly $26.5 billion behind open-weight AI infrastructure: Stripe’s reported $7.5 billion for OpenRouter, Nvidia’s $6 billion arrangement with Poolside, and a reported $13 billion Nvidia bid for Hugging Face, as TechCrunch tallied this week. Set that against the demand side and the bet looks stranger than the headlines suggest: only 6.1 percent of AI-using US businesses touched a model-serving platform in July, per Ramp’s August AI Index, up 0.2 points month over month.
Divide the capital by the adoption and you get the figure that frames the trade. About $4.3 billion of acquisition value per percentage point of business adoption of the platforms that serve open and Chinese-developed models. For comparison, Ramp puts Anthropic at 43.5 percent of US businesses and OpenAI at 39.7 percent — the incumbents hold roughly seven times the penetration that the acquired layer serves.
Buying the pipe before the water arrives
Nvidia’s reported approach to Hugging Face, first detailed in coverage of the $13 billion figure, would hand a chip vendor the largest US distribution surface for models it does not make. That is the strategic logic and the strategic problem in one sentence.
The engineering-side data is even thinner and moving faster. Jellyfish’s August report, covering 276,000 engineers, finds open-weight model selection at under 2 percent of engineers in a given week, with the overall share of companies using open-weight models or routers “roughly doubling over the last six weeks.” Within that small population the concentration is Chinese: Kimi at 51.8 percent of open-weight users, GLM at 51.2 percent, DeepSeek at 8.6 percent, Qwen at 7.9 percent.
Two percent doubling every six weeks is either a rounding error or the early part of a curve, and $26.5 billion is a wager on the second reading. Nvidia’s motive is legible: as model builders ship their own inference silicon, owning the largest US distribution surface for open models is a hedge against hyperscaler concentration. Its own Nemotron family of open models already exists and, per TechCrunch, has not achieved wide uptake — buying the hub is faster than winning the leaderboard.
The counter-case is in the same data. Ramp’s analysis notes that first-time AI buyers “are still using the American model companies,” so open-weight growth is coming from advanced spenders rather than new demand, and the platforms are capturing spend that already exists rather than creating it. Fireworks CEO Lin Qiao told TechCrunch her company processes 40 trillion tokens a day, more than either Gemini’s or OpenAI’s APIs — a claim that is unaudited but, if true, means the open layer’s token volume already outruns its dollar share by a wide margin. Cheap tokens make impressive volume and modest revenue.
What breaks the thesis
Price convergence is the fragile input. The open-weight case strengthens only if frontier prices stay high, and the frontier keeps discounting — the same dynamic visible when Anthropic’s priciest tier took just 8 percent of its own token volume because businesses refused the premium. Jellyfish’s Nik Albarran frames the trigger precisely: companies turn to open models for control and configurability today, and “if the prices continue to go up from the frontier labs, more and more companies will be forced to at least consider it.” Prices going down, not up, is the acquirers’ real risk.
The second risk is that supply is contractual, not permanent. OpenAI’s decision to stop serving models to Cursor from November 12, reported by Reuters via CNA, is a reminder that an aggregation layer’s inventory can be withdrawn by the parties it aggregates. That risk cuts toward open weights, which cannot be revoked once downloaded — but it also caps what any hub can promise, since the checkpoint you already hold does not need the hub. This paper priced GLM-5.3-Flash’s 42x blended cost advantage over Opus 4.8 from published rates, and nothing about that arithmetic requires an intermediary.
The third is that consolidation removes the neutrality that made these assets valuable. A hub owned by a chip vendor has an interest in which chips your weights run on, which is the same conflict this paper flagged when Hugging Face was fielding $13 billion bids while charging a 38 percent premium over a neocloud for a B200 hour. Neutrality is a service; it is not obviously compatible with strategic ownership.
The fourth risk is that the acquired assets are worth less separated from their communities than attached to them. A hub’s value is contributor behavior, not code, and contributor behavior responds to ownership. Nothing in these deals guarantees the people uploading weights stay after the logo changes, and no acquirer has yet demonstrated it can hold an open ecosystem through a change of control.
For operators the decision is narrower than the headlines. If your workload is high-volume, repetitive inference — support chat, classification, extraction — the routing math already favors an open-weight tier, and Cursor’s data on routed work reinforces it, as today’s brief on cost per commit shows. If your workload is long-horizon agentic coding, frontier models still win on Jellyfish’s own numbers, and switching now buys you migration cost against a price gap that may close on its own. The thing worth doing this quarter is cheap: keep an abstraction layer between your application and any single hub, because the $26.5 billion just bought the right to change your terms. The macro backdrop from today’s lead on the chip-tariff exemption applies here too — the cost of serving any weights, open or closed, starts with imported silicon.