Vercel has quietly turned into one of the busiest companies in AI, and its CEO thinks the whole industry is about to split in two. In an interview with TechCrunch AI after the company’s ShipNYC conference, Guillermo Rauch laid out the numbers behind that claim: 6 million deployments a day, half of them triggered by coding agents, and more than 1 trillion tokens flowing through Vercel’s AI gateway every day. What stands out here is the bigger fight he’s picking. Rauch wants to decouple the model from the agent, and he’s betting Vercel’s future on it.
What’s Actually Changing
Last year was the prototyping phase. Rauch describes it as “unleash the agents, everyone can build.” This year the mood shifted to making agents work in production, and that shift exposed two use cases that actually pay off.
- The coding agent. It’s driving most of the world’s token use. But as Rauch points out, when you generate that much software, “you need somewhere to put it.” That’s Vercel’s home turf.
- The internal company agent. The one that helps you run the business. The hard part isn’t intelligence, it’s security: how do you let an agent touch company data, audit what it did, and keep a trail of every tool call?
Vercel’s answer is two products. Eve lets you write an agent’s instructions and skills in plain language. Vercel Sandbox puts the agent “in a little cage,” where it stays smart but you control what data it can see and what data can leave.
Why the Data Question Matters Now
Rauch’s sharpest warning is about data leaking into training. Install the wrong dev tool in the wrong setting, he says, and your entire codebase can get shipped to the cloud. He recalls talking to the president of Airbus about decades of proprietary C++ aerospace code, and the risk that “someone comes in and installs the wrong developer tool and boom, all the code goes out.”
This is significant because it reframes the SaaS business model. Rauch argues that many software giants built their kingdoms on trapping your data, and “that’s incompatible with agents.” Agents force companies to open up their systems through APIs. That’s a long-term threat to any vendor whose moat is lock-in.
Are Companies Still Marrying One AI Lab?
Short answer, according to Rauch: no. Last year, buyers picked a single partner and built everything on OpenAI or Anthropic. Now they treat the stack as modular. Model, harness, data platform, sandbox, gateway, every piece plug and play.
That has a real consequence for the competitive map:
- Gemini is quietly winning production workloads. Not because of headlines, Rauch says, but because “when you’re optimizing for production, you start looking at price/performance, and Gemini models have awesome price/performance characteristics.”
- Open models are climbing. He names DeepSeek and GLM-5.2 as taking off.
- The labs are now competitors, not just partners. When OpenAI shipped tools to publish websites directly, it stepped onto Vercel’s turf. Rauch spins it as an opening, betting that people who ask ChatGPT about web hosting get pointed back to platforms like his.
The Real Bet: Decoupling Model from Agent
Here’s the thesis underneath everything. Rauch says the industry is deciding one question: do you get all your intelligence from one place, or do you buy building blocks from different providers and assemble them yourself? He wants the second world, the one that “software engineering has always been.” His line: “We’re going to be the AWS of this generation, so obviously we’re fighting for a world of open protocols.”
What to Do With This
If you’re building with AI right now, three practical takeaways:
- Stop betting the company on one lab. Design your stack so the model is swappable. Price and performance shift fast, and Gemini plus open models are real options today.
- Treat data governance as a build requirement, not a cleanup task. Know what your agents can access and what can leave. Sandboxing isn’t optional at production scale.
- Watch the lock-in vendors. If your tooling traps data, agents will expose that friction. Favor open protocols and API-first systems.
Rauch is talking his own book, and “AWS of this generation” is a bold flag to plant. But the production data he’s sitting on gives the argument weight. The coupling question he’s raising is the one every AI team will answer over the next year, whether they mean to or not. More detail is in the full TechCrunch AI interview.