Vercel, known for cloud infrastructure that lets developers deploy agents without managing servers, has quietly become a central player in AI software. The company reports 6 million deployments per day, about half of them triggered by coding agents, and more than 1 trillion tokens flow daily through Vercel’s AI gateway.
After Vercel’s ShipNYC conference last week, CEO Guillermo Rauch spoke about how the company sees the current phase of AI, and how platform providers like Vercel can end up competing with major model labs. The following is a condensed account of that conversation.
From prototyping to production realities
Rauch described last year as a period of prototyping—everyone experimenting and ‘‘unleashing’’ agents. Vercel’s internal experience, with hundreds of organically developed and deployed agents, shifted the focus to production challenges. The company identified two primary production use cases for agents:
- Coding agents, which account for a large share of token usage and produce large amounts of software that need to be hosted, and
- Internal corporate agents that help run the business, raising questions about secure data access, auditability, and tracing tool calls and access controls.
Eve and Vercel Sandbox: natural-language instructions and a controlled environment
To address those needs, Vercel created Eve, a framework where an agent’s instructions and capabilities can be defined in natural language. Vercel Sandbox complements Eve by putting an agent in a ‘‘small cage’’: the agent retains freedom to act and demonstrate intelligence, while policies control which data it can access and which data may leave the sandbox.
Rauch emphasized that Sandbox’s biggest advantage is data control. He gave an example about IDEs: if a developer tool is misconfigured, it could send an entire codebase to the cloud for model training. He recalled a conversation with the president of Airbus about decades of specialized C++ aerospace code that could be at risk if the wrong developer tool were installed.
A concrete internal use case
Rauch described a sales representative focused on growing existing accounts. The bottleneck historically was access to data—questions like "which accounts are growing fastest?" or "which five accounts added the most seats in the past two weeks?" previously required waiting for a Q1 dashboard project. With Eve, those kinds of queries and insights become instantly available, increasing productivity across the company.
Rauch argues agents are forcing companies to open up their data in ways that conflict with many SaaS giants’ business models, which rely on locking in customer data.
Changing relationships with major AI labs and a modular approach
On the customer side, Rauch said last year many firms chose a single lab partner (for example OpenAI or Anthropic), but now customers see the system as modular—model, harness, data platform, sandbox, gateway—each piece plug-and-play. That shift favors providers who can interoperate with multiple models and infrastructures.
Rauch noted growth in Google Gemini use in production settings because of competitive price/performance, and also mentioned rising adoption of open models such as Deepseek and GLM-5.2.
Are models and agents becoming coupled or decoupled?
Rauch framed a central industry question: will intelligence come from a single, coupled source, or will teams assemble modules and build on top—mirroring traditional software engineering? Vercel aims to be the AWS-like infrastructure for this generation and advocates for open protocols.
Direct competition and platform feature overlap
Rauch acknowledged that model labs are adding infrastructure-like features. He pointed to OpenAI’s recent tools that publish directly to the web from within the OpenAI environment as a natural next step. Those moves can compete with existing infrastructure platforms, but they also present openings: if users start treating ChatGPT-like services as website builders, the model’s recommendations about hosting can drive traffic to infrastructure providers such as Vercel.
According to Rauch, the industry is deciding whether models and agents will be tightly coupled or remain modular, and infrastructure providers are positioning themselves for a future of open, interoperable components.
When this was said
The remarks were made by Guillermo Rauch in an interview following Vercel’s ShipNYC conference, which the company held last week. Reported platform figures: 6 million deployments per day (roughly half initiated by coding agents) and more than 1 trillion tokens flowing daily through Vercel’s AI gateway.
