Vercel is making a sharper case for how enterprise AI should be built: keep the model separate from the agent, and treat the surrounding stack as interchangeable infrastructure rather than a single vendor bundle. In an interview with TechCrunch AI following the company’s ShipNYC event, CEO Guillermo Rauch said customers are moving past early experimentation and into production tradeoffs around cost, governance, and reliability.

That matters because Vercel is no longer just talking about frontend deployment. According to Rauch’s comments to TechCrunch AI, the company now handles 6 million deployments a day, with half triggered by coding agents, and more than 1 trillion tokens passing through its AI gateway daily. If those figures hold, they suggest Vercel is positioning itself as a control layer for how AI applications are built and shipped, especially for teams that do not want to lock themselves into a single model provider.

The broader argument from Rauch is strategic as much as technical. As AI labs add more end-to-end features, including tooling that can generate and publish software directly, infrastructure vendors such as Vercel face the risk of being squeezed out. Rauch’s answer is to argue for a modular market in which the “model, harness, data platform, sandbox, gateway” remain plug-and-play components. For builders and enterprise buyers, that is really a debate about lock-in, economics, and who controls application behavior in production.

From prototype hype to production constraints

Rauch told TechCrunch AI that 2025’s AI enthusiasm was centered on broad experimentation, while the current moment is more operational. In his telling, companies first unleashed agents widely, then ran into the realities of production use: security boundaries, data access, audit trails, and the practical cost of running large volumes of inference.

He identified two “killer apps” for agents inside Vercel’s experience. The first is the coding agent, which he said is already driving a large share of token usage. That tracks with the wider market focus on developer tooling, where products such as Devin and Cursor have become shorthand for autonomous or semi-autonomous code generation workflows. The second, he argued, is the internal enterprise agent that helps run the company itself.

That second category is where Vercel is trying to differentiate. Rauch described internal users, such as sales staff, querying operational data directly instead of waiting for dashboards or custom reporting projects. The point is less that agents replace SaaS systems like Salesforce than that they sit across them, turning stored company data into a conversational interface and action layer.

For enterprise teams, that promise is attractive only if access can be controlled. Rauch’s comments suggest Vercel sees governance, not model intelligence alone, as the gating factor for broader deployment. The more an agent can do across departments, the more buyers care about permissioning, traceability, and preventing sensitive data from flowing into the wrong training pipeline.

Vercel’s product argument: control plane over model loyalty

To support that position, Rauch pointed to two Vercel tools discussed with TechCrunch AI: Eve and Vercel Sandbox. He described Eve as a framework for laying out agent instructions and skills in natural language, and Vercel Sandbox as a constrained environment where an agent can operate under policy controls for data access and data egress.

Those details matter because they show where Vercel wants to sit in the stack. The company is not presenting itself as a foundation model developer. Instead, it is trying to own the runtime, orchestration, and safety boundaries around models. That is a familiar infrastructure playbook, but applied to AI agents rather than conventional cloud workloads.

Rauch also tied the sandbox argument to an enterprise fear that has become central to AI procurement: accidental exposure of proprietary code or internal data. He used coding tools such as Devin or Cursor as examples of environments where, under the wrong settings, sensitive codebases could be used in ways a customer did not intend. The specific scenario he described was anecdotal and framed through a conversation with an Airbus executive, not as a disclosed incident. Still, it reflects a common buying concern in enterprise AI: whether convenience features obscure data-governance risk.

Vercel’s message, then, is that agent infrastructure should make intelligence portable while keeping policy local. That is how the company is trying to turn AI from a feature layer into a reason to buy more infrastructure.

Why Rauch is highlighting model choice now

One of the most notable parts of the interview was Rauch’s description of changing customer behavior toward model providers. According to him, last year many teams chose a single lab, often OpenAI or Anthropic, and tried to build most of their AI roadmap around that partner. Now, he said, customers increasingly see the stack as modular and choose models based on workload needs.

Rauch specifically said Vercel is seeing growth in Gemini usage because teams optimizing for production are looking at price and performance rather than headlines. He also said open models are gaining traction, naming DeepSeek and GLM-5.2. Those remarks are important but should be read carefully: TechCrunch AI reported them as executive comments from Rauch, not as independently verified market-share data.

Even so, the logic matches what many product teams are confronting. Once an application reaches production scale, the question is no longer just capability. It becomes whether a cheaper or faster model can handle enough of the work, whether a higher-end model should be reserved for fallback cases, and whether routing across providers can cut cost without breaking quality.

That is exactly the kind of architecture Vercel benefits from. If customers use OpenAI, Anthropic, Gemini, DeepSeek, and GLM-5.2 through a shared gateway and policy layer, the infrastructure platform becomes harder to displace. Vercel’s AI gateway, as described by Rauch, is part of that thesis.

Competition with the labs is becoming explicit

Rauch did not avoid the obvious tension: as major labs add application and deployment features, they move closer to the territory of cloud and developer platform companies. He referenced recent tools from OpenAI that can publish directly to the web without leaving OpenAI’s environment, framing that move as both direct competition and a discovery channel for Vercel.

His answer was effectively that the market is deciding whether the model and the agent will remain tightly coupled. In one version of the market, customers buy an all-in-one system where a lab provides the model, the tools, the deployment surface, and perhaps even the user-facing application. In the other version, customers assemble systems from components and swap intelligence providers as needed.

Vercel clearly wants the second outcome. Rauch told TechCrunch AI that the company is “fighting for a world of open protocols” and cast Vercel as trying to become foundational infrastructure for this generation of AI software. That is a strategic claim, not an established market fact, but it explains why Vercel is emphasizing interoperability so heavily.

For buyers, this is more than vendor rhetoric. Coupled systems can be simpler to adopt but harder to govern and more expensive to change later. Modular systems can reduce dependency on one provider but may require stronger internal engineering discipline. The split between those approaches is likely to define purchasing decisions across enterprise AI over the next 12 to 24 months.

Evidence, claims, and what is still unverified

The strongest facts in this story come from TechCrunch AI’s interview with Guillermo Rauch. The operational figures — 6 million deployments per day, half triggered by agents, and more than 1 trillion tokens flowing through the AI gateway daily — were attributed to Vercel via Rauch’s comments. They are significant, but they remain company-reported metrics rather than independently audited disclosures.

The same caution applies to Rauch’s statements about rising usage of Gemini, DeepSeek, and GLM-5.2, and about changes in how customers evaluate OpenAI and Anthropic. These comments are useful market signals from a platform operator with visibility into workloads, but TechCrunch did not present third-party validation in the source material provided here.

The product descriptions of Eve and Vercel Sandbox also come through Rauch’s framing in the interview. They indicate the direction of Vercel’s platform strategy, especially around policy and agent control, but the source evidence does not include detailed technical documentation, pricing, or customer deployment case studies. That leaves open questions about maturity, differentiation, and how broadly these tools are used outside Vercel itself.

What this means for builders and enterprise buyers

For builders, the immediate takeaway is architectural. If Vercel is right that the market is separating models from agents, product teams should think about their AI systems as layered services: model selection, routing, observability, permissions, and execution controls. That favors vendors that can unify deployment and governance without forcing one model choice.

For enterprise AI buyers, the interview underlines two purchase criteria that are becoming more concrete. First is data control: whether internal data can be queried and acted on without leaking into external systems or opaque training processes. Second is economic flexibility: whether applications can move between OpenAI, Anthropic, Gemini, or open models as pricing and quality change.

For the market, Vercel’s stance also shows how quickly “developer tools” are turning into enterprise workflow infrastructure. The same stack used to deploy apps is now being pitched as the place to run coding agents, govern internal agents, and mediate access to systems like Salesforce. If that holds, the competitive set around Vercel will expand well beyond hosting and frontend platforms.

What to watch next

The next useful signal will be whether Vercel publishes more concrete evidence behind Rauch’s usage claims, especially around the AI gateway and adoption of Eve and Vercel Sandbox. Customer references would matter more than topline token counts.

It will also be worth watching whether Vercel deepens integrations across model providers such as OpenAI, Anthropic, and Gemini while keeping support for open models like DeepSeek and GLM-5.2. If the company can make switching or routing between them operationally simple, its modularity argument gets stronger.

Another signal is competitive response. If labs keep expanding end-to-end deployment, coding, and hosting workflows, the line between model vendor and infrastructure platform will blur further. Whether enterprises accept those bundled environments or insist on more open control layers will shape who captures the most value.

Creati.ai perspective

Rauch’s interview is notable because it frames the next AI platform battle less around model supremacy and more around system design. Vercel is betting that enterprises will not want one provider to own intelligence, execution, deployment, and data boundaries all at once. That is a credible thesis, especially for regulated teams and large engineering organizations that expect model churn and want leverage over cost.

But the company still has to prove that its control layer is not just conceptually attractive but operationally indispensable. The reported scale numbers are eye-catching, yet vendor-reported usage does not automatically translate into durable platform power. The real test is whether Vercel can become the default place where AI agents are governed, not just deployed, as customers mix OpenAI, Anthropic, Gemini, DeepSeek, and GLM-5.2 in production.