Vercel CEO Guillermo Rauch on the AI battle: Decoupling models from agents

Vercel, renowned for its cloud infrastructure enabling developers to deploy AI agents without server management, has quietly emerged as a pivotal player in the AI software landscape. The company currently facilitates an astounding 6 million deployments daily, with half of these originating from coding agents. Furthermore, its AI gateway processes over 1 trillion tokens every day.

Following the company's recent ShipNYC conference, we engaged with Vercel CEO Guillermo Rauch to gather his insights on the current state of AI and how platform companies like Vercel navigate competition with major AI research labs. Below is a lightly edited transcript of our conversation.

There's a palpable shift in the community's energy this year; less emphasis on pilot programs and a greater focus on achieving practical, high-performing solutions. While this trend is undoubtedly evident among your clients, we're curious about Vercel's internal journey in adapting to this new reality.

Last year was primarily dedicated to prototyping—an era of boundless possibilities where agents were unleashed, and everyone was encouraged to build. We participated fully in this phase, gaining invaluable insights from hundreds of agents organically developed and deployed within our own company. This experience quickly brought us face-to-face with the practicalities and inherent challenges of deploying agents in production environments.

For me, the most significant lesson was identifying the "home-run" use cases, the two "killer apps" for agents. The first, unequivocally, is the coding agent, which accounts for a substantial portion of global token utilization. However, generating vast amounts of software necessitates a robust deployment solution. The second crucial application involves internal agents designed to optimize company operations. This category presents unique challenges, particularly concerning secure data access, comprehensive auditing of agent activities, and maintaining a clear trail of all tool calls and access controls an agent utilizes to complete a task.

To address these complexities, we developed the Eve framework, which allows users to define an agent's instructions and capabilities using natural language. Complementing this is Vercel Sandbox, a controlled environment where agents operate within defined boundaries. This "cage" grants agents the freedom to express their intelligence while allowing us to enforce strict policies on the data they can access and the information permitted to leave the sandbox.

What specific problems do these innovations help you mitigate?

For Sandbox, the paramount advantage lies in data control. A significant risk I often consider with AI is the potential for coding IDEs, such as Devin or Cursor, to inadvertently train on an entire codebase if configured improperly. I recall a conversation with the president of Airbus on this very topic: imagine decades of highly specialized C++ code for aerospace engineering, and then an incorrect developer tool is installed, leading to the entire codebase being uploaded to the cloud for training. This scenario highlights the critical need for robust data governance.

I'm particularly interested in exploring that second "killer" use case further. While coding agents are widely understood, what does an internal corporate agent look like in practical application?

Consider a sales representative in Vercel's office whose role is to expand existing accounts within our install base. Her primary bottleneck isn't her creativity, intelligence, or ability to forge relationships; it's data access. She needs to answer questions like, "Which accounts are growing fastest? Provide me with the five accounts that have added the most seats in the last two weeks so I can prioritize my outreach." Historically, she couldn't pose such immediate queries; she'd have to wait for a quarterly project to deliver a new sales dashboard.

Vercel experienced this data bottleneck for years, which was incredibly frustrating. On the R&D side, we pride ourselves on being one of the fastest-moving companies globally. Yet, when it came to our sales engine and Salesforce engineering, I personally felt quite incompetent, having never even opened Salesforce before joining the company.

Now, I feel empowered to make an impact across the entire organization because Eve can power both our customer-facing agents and agents designed to boost internal productivity. It's the same underlying technology, simply exposed via APIs. Agents are compelling companies to become more open with their data, which will have profound long-term implications. Many SaaS giants have built their empires by locking in customer data, a model fundamentally incompatible with the agent paradigm.

How do you perceive the evolution of client relationships with the major AI labs?

Last year, many clients committed to a single lab partner, intending to build everything on platforms like OpenAI or Anthropic. Today, there's a growing understanding of the modular nature of the AI stack—model, harness, data platform, sandbox, gateway—where each component is plug-and-play. Clients are now comfortable using OpenAI, Anthropic, or Gemini. We're observing significant growth in Gemini's adoption, even if it doesn't always dominate headlines, because companies are now optimizing for production environments. When production optimization is the goal, price-performance becomes a critical factor, and Gemini models offer excellent characteristics in this regard. Additionally, open models like Deepseek and GLM-5.2 are rapidly gaining traction, demonstrating that the data truly speaks for itself.

There are also areas where Vercel finds itself in direct competition with these labs, correct? Just recently, OpenAI launched new tools that enable direct web publishing without ever leaving their ecosystem.

It's a natural progression for them to offer hosting for smaller websites. Paradoxically, this presents a significant opportunity for us. Now, people will begin to perceive ChatGPT as a tool for website creation. If they continue to query the model about web hosting, the model will, in turn, recommend Vercel. However, you're absolutely right: as these models and platforms expand their capabilities, they inevitably enter into direct competition with existing infrastructure platforms.

At this juncture, I believe we are fundamentally deciding whether the model and the agent will be tightly coupled.

The core question is: will intelligence be sourced from a single, monolithic provider, or will it be assembled from modular components, libraries, or building blocks from various providers? This latter approach mirrors the long-standing principles of software engineering, and it's precisely the vision we are bringing to market. We aspire to be the AWS of this generation, and as such, we are staunch advocates for a world built on open protocols.