Vercel CEO Guillermo Rauch is arguing that the next fight in AI infrastructure is whether the model and the agent stay coupled or come apart. Speaking after the company's ShipNYC conference last week, Rauch said Vercel now processes 6 million deployments a day, with roughly 50% of them triggered by coding agents, and more than 1 trillion tokens flowing through its AI gateway daily. Those numbers put Vercel at the center of how AI-written software actually reaches production.
The pitch is that 2025 was a prototyping year and 2026 is the cleanup. Rauch says Vercel ran hundreds of internal agents through the transition and hit the same wall enterprises are hitting now: data access, auditability, and cost.
Two use cases, in his telling, are pulling ahead. The first is the coding agent, which is driving the bulk of global token consumption. The second is the internal corporate agent — the one that lets a sales rep ask which five accounts added the most seats in the last two weeks, without waiting on a Q1 dashboard project. Rauch framed the second category as a productivity unlock that had been blocked for years inside Vercel itself, where the R&D side moved fast but the Salesforce-facing side did not.
Key facts
- 01Vercel handles 6 million deployments a day, half of them triggered by coding agents.
- 02More than 1 trillion tokens flow through Vercel's AI gateway daily.
- 03CEO Guillermo Rauch says customers are moving off single-lab lock-in toward mixing OpenAI, Anthropic, Gemini, DeepSeek and GLM-5.2.
- 04Vercel launched Eve, a natural-language agent framework, and Vercel Sandbox to cage internal agents and control what data they can access.
- 05Rauch positions Vercel as 'the AWS of this generation,' fighting to keep models and agents decoupled rather than bundled by a single lab.
To make internal agents deployable, Vercel shipped two products. Eve is a framework that lets teams define an agent's instructions and skills in natural language. Vercel Sandbox puts the agent in what Rauch calls a cage, applying policy on what data the agent can touch and what data can leave. The stated concern is concrete: a developer installing the wrong AI-powered IDE — Rauch named Devin and Cursor as examples of the category — and inadvertently sending a codebase out for training. He cited a conversation with the president of Airbus, whose C++ aerospace code represents decades of proprietary work.
On model selection, Rauch says the single-lab bet is over. Customers who standardized last year on OpenAI or Anthropic are now treating models as swappable modules alongside harness, data platform, sandbox, and gateway. Gemini is picking up production share on price/performance, and open models including DeepSeek and GLM-5.2 are showing up in the mix.
That multi-model reality is the wedge under Rauch's larger argument. If enterprises want the freedom to swap models, they need infrastructure that treats the model as a component rather than a platform. That is the layer Vercel wants to own.
It also puts Vercel into direct competition with the labs themselves. Rauch pointed to OpenAI's recent move to let users publish sites directly from within its own environment — a natural extension for a company whose users increasingly treat ChatGPT as a starting point for building on the web. He framed the encroachment as inevitable and as an opening: if the model recommends a host, Vercel wants to be the recommendation.
The company's positioning is explicit. Rauch says agents are forcing SaaS incumbents to open up their data, which he expects to erode the moats that firms like Salesforce built by trapping customer information inside proprietary systems. If agents cannot get at the data, the agents do not work. If the data opens up, the switching costs collapse.
There is a counterweight. Rauch's thesis assumes that model providers keep their capabilities modular and that customers keep wanting choice. If a single lab ships an agent stack that is materially better end-to-end — model plus harness plus deployment plus tools — the decoupled world Vercel is betting on becomes a harder sell. OpenAI's site-publishing move is a small early signal of that risk, and Anthropic has shown similar ambitions with its own tooling. The bet also depends on enterprises trusting Vercel-style sandboxes over lab-native environments once the labs offer comparable controls.
The AWS analogy Rauch reached for is the frame worth watching. AWS won by selling primitives that developers assembled themselves, at a time when the alternative was buying a full stack from IBM or Oracle. Vercel is making the same argument one layer up: sell the primitives — gateway, sandbox, agent framework — and let customers pick the intelligence. If the model-agent decoupling holds, the infrastructure layer becomes the most valuable real estate in AI, and the labs end up as suppliers rather than platforms. If it doesn't, Vercel is a very fast deployment company competing with vertically integrated giants. Six million deployments a day suggests the market is, for now, voting with Rauch.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
