Vercel Wants to Be AI's AWS — and That Means Taking On the Labs Directly

Vercel has spent years as the quiet backbone of modern web deployment, but the company is rapidly becoming one of the most consequential players in the AI infrastructure space. With 6 million deployments happening every day — half of them now triggered by coding agents — and more than 1 trillion tokens processed daily through its AI gateway, Vercel sits at a scale most startups can only dream about. CEO Guillermo Rauch sat down following the company's ShipNYC conference to explain how Vercel sees its role evolving, and why the architecture decisions made right now will define the next decade of software.

Rauch identifies two dominant 'killer apps' for agents at this moment: coding agents, which are already driving enormous token consumption worldwide, and internal corporate agents that help businesses operate more efficiently. The second category is less discussed publicly but arguably more transformative. Rauch used the example of a Vercel sales representative whose job involves growing existing accounts — a person whose effectiveness was historically bottlenecked not by talent, but by an inability to quickly query live business data. Agents, Rauch argues, are forcing companies to open up data silos that legacy SaaS platforms have deliberately kept locked away, and that shift will have profound long-term consequences for the enterprise software industry.

To address the security and governance challenges that come with deploying agents in production, Vercel built two new tools: Eve, a framework for defining an agent's instructions and capabilities in natural language, and Vercel Sandbox, which isolates agents in a controlled environment while still allowing them to operate intelligently. The Sandbox product addresses a specific fear Rauch has thought about seriously — the risk that improperly configured developer tools could expose sensitive proprietary codebases to external training pipelines. He recalled a conversation with an Airbus executive about decades of specialized aerospace engineering code that could theoretically be vacuumed up by a poorly vetted IDE.

On the competitive landscape with major AI labs, Rauch noted a clear market shift away from exclusive partnerships with a single provider. Where developers once committed entirely to OpenAI or Anthropic, they are now thinking in modular terms — treating the model, the harness, the data platform, and the sandbox as interchangeable components. Gemini, Rauch noted, has seen particularly strong growth in production environments because of its price-to-performance ratio, and open models like DeepSeek are also gaining traction. The pattern, he says, mirrors how software engineering has always worked: libraries, building blocks, and composable systems rather than monolithic stacks.

The most pointed tension Rauch addressed was the encroachment of the labs themselves into infrastructure territory. When OpenAI released tools allowing users to publish websites directly without leaving its platform, it entered space that Vercel has long occupied. Rauch framed this as both a challenge and an opportunity — users thinking of ChatGPT as a website-building tool might naturally discover Vercel as the next step. But the deeper strategic question, as he sees it, is whether the industry moves toward a coupled model where intelligence and deployment are bundled by a single lab, or toward an open ecosystem where each component is independently sourced. Vercel is betting — and building — for the latter.