Vercel Ships Eve and Sandbox and Bets Against Single-Lab AI Infrastructure
6 million daily deployments and 1 trillion daily tokens are the foundation of Rauch's argument that models and agents must come apart
Vercel CEO Guillermo Rauch laid out his infrastructure thesis after ShipNYC last week: the model and the agent should not come from the same company, and a developer who locks both to one lab is making a bet they cannot unwind. At 6 million deployments a day - roughly half triggered by coding agents - and over 1 trillion tokens passing through Vercel's AI gateway each day, Rauch is not making an academic argument. Vercel runs the rails that AI-generated code deploys on, and his read of what enterprises hit in production is shaping what the company builds next.
2025 Was the Prototype Year - 2026 Is Where Enterprises Hit the Same Three Walls
Rauch's framing is that 2025 was a prototyping year: developers experimented, enterprises started pilots. 2026 is where those pilots hit production constraints. Running hundreds of internal agents at Vercel, his teams encountered the same three walls enterprises are hitting now - data access, auditability, and cost.
Two use cases are pulling ahead of the rest. Coding agents are driving the bulk of global token consumption - that 50% agent-triggered deployment figure reflects how much code now gets written without a human touching every line. A second category is the internal corporate agent, which Rauch illustrated from Vercel's own experience: the R&D side moved fast, but the Salesforce-facing sales side could not query its own data without a dashboard project queued in engineering. Giving a sales rep a natural-language interface to ask which five accounts expanded headcount last week - without a ticket, without a wait - is the gap Vercel's Eve framework is designed to close. For a deeper look at how agentic systems work in practice, see What Is Agentic AI - And Why It's About to Change How You Work.
Eve Defines Agents in Natural Language - Sandbox Decides What Data They Can Touch
Eve is a framework for specifying an agent's instructions and skills in natural language, cutting the configuration overhead of building that spec in code. Vercel Sandbox runs alongside it with a different job: it places the agent inside a policy layer that governs what data the agent can reach and what data can leave. Rauch's concern is concrete - a developer installs an AI coding tool, that tool begins absorbing the team's codebase for training, and months of proprietary work exits the building unnoticed. He named Devin and Cursor as examples of the tool category where that risk exists, and cited a conversation with the president of Airbus, whose C++ aerospace code represents decades of work that cannot be reconstructed once it leaks.
For developers deploying internal agents inside enterprises, Sandbox is an answer to the question every security team will ask before sign-off: how do we know what data the agent is sending, and where is it going? That question has not had a clean infrastructure-layer answer before. Framing the solution as a cage with enforced data policy rather than a vendor promise is the right instinct - whether Vercel's implementation survives enterprise security reviews is the real next test.
Gemini and DeepSeek Are Winning Production Share as Single-Lab Bets Unwind
On model selection, Rauch's numbers are direct: the single-lab bet is unwinding in production. Customers who standardized on OpenAI or Anthropic in 2024 are now treating models as swappable components alongside harness, data platform, sandbox, and gateway. Gemini is picking up share on price-performance grounds. Open models including DeepSeek and GLM-5.2 are appearing in enterprise stacks alongside closed-API providers. Anyone building on Cursor's agent-triggered workflow is already navigating this multi-model reality on a daily basis.
Multi-model adoption is the structural wedge under Vercel's positioning. If enterprises want the freedom to swap models, they need infrastructure that treats the model as a component rather than a platform - and Vercel wants to own that layer. Rauch used the AWS analogy explicitly: sell the primitives (gateway, sandbox, agent framework) and let customers pick the intelligence above them. Rauch's AWS analogy has limits, though. AWS won before the compute providers had platform ambitions of their own. OpenAI recently moved to let users publish sites from within its own environment - a direct step into deployment territory - and Anthropic has shown similar tooling ambitions. If one lab ships a model-plus-agent-plus-deployment stack that is materially better end-to-end, the decoupled world Vercel is betting on becomes a harder pitch. Six million deployments a day suggests that moment has not arrived yet - but Rauch knows it is the scenario he has to keep outrunning.