Vercel’s Guillermo Rauch Says the Next AI Battle Is Over Control, Not Just Models

Vercel has spent years becoming the invisible layer under a huge share of modern web development, but the company is now positioning itself as something bigger: a central piece of the AI software stack. According to CEO Guillermo Rauch, the real fight in artificial intelligence is no longer just about building better models. It is about deciding where intelligence lives, who controls the data around it, and whether agents remain loosely connected tools or become locked inside the ecosystems of major labs.

In a wide-ranging conversation after Vercel’s ShipNYC conference, Rauch described a rapidly changing market in which developers are moving beyond experiments and into production. That shift, he said, has exposed the practical problems that determine whether AI agents are genuinely useful at scale: data governance, auditability, security boundaries and the economics of choosing between competing model providers.

Vercel, best known for helping developers deploy applications without managing infrastructure, says it is now seeing roughly 6 million deployments every day. About half of those deployments are triggered by coding agents, according to the company, while more than 1 trillion tokens pass through Vercel’s AI gateway daily. Those numbers help explain why Rauch believes the company sits at a strategic choke point in the AI economy.

His core argument is simple: the market is shifting away from a world in which developers commit to a single model provider and toward one in which every piece of the stack can be swapped. That includes the model, the agent harness, the data platform, the gateway and the execution environment. In that world, Rauch says, infrastructure providers like Vercel become more important, not less.

The moment AI moved from demos to deployment

Rauch said the industry’s mood has changed over the past year. The early phase was dominated by prototypes, ambitious experiments and a belief that autonomous agents could be rolled out almost immediately across business functions. That enthusiasm did generate results, he said, but it also revealed the operational constraints that only appear once software is used in real workflows.

Within Vercel, he said, hundreds of agents were created organically by employees during the company’s exploration phase. That internal experimentation produced an important lesson: the flashy demos were not the hardest part. The difficult work started when those systems had to operate in production, under real security, compliance and performance requirements.

Rauch framed the shift as a move from “can we build it?” to “can we trust it, govern it and scale it?” That transition, he argued, is where many AI deployments stall. Teams often discover that the hardest questions are not about model quality but about access controls, logging, permissioning and traceability.

Why coding agents became the first breakout use case

In Rauch’s view, one of the two major “killer apps” for agents is coding. Development tools that can write, refactor, test and ship software are already responsible for a large share of token usage across the industry. But the success of those tools creates a second problem: every additional line of code generated by an agent has to go somewhere, and that means companies need robust deployment infrastructure.

He argued that coding agents are so influential because they sit close to the highest-value work in software organizations. They save time, accelerate iteration and make it easier for small teams to ship quickly. But their utility also raises the stakes around privacy and intellectual property, especially when they are connected to proprietary codebases.

Rauch pointed to the risk that a developer tool could accidentally expose sensitive source code to cloud systems used for training or processing. In his telling, the concern is not theoretical. Enterprises in sectors like aerospace, manufacturing and enterprise software may have decades of embedded technical knowledge in their repositories. A poorly configured assistant, he warned, could move that knowledge beyond the company’s control.

The second killer app: internal agents

The other major use case, Rauch said, is the internal corporate agent: systems that help employees navigate data, execute workflows and make decisions faster. These tools are less visible to consumers than coding copilots, but they may be more valuable over time because they address one of the oldest bottlenecks in business operations — getting the right information to the right person at the right moment.

Rauch described a familiar example inside sales teams. A representative responsible for expanding existing accounts might want to know which customers have added the most seats recently or which accounts are showing the strongest growth momentum. In older enterprise systems, getting that answer could require a formal analytics project, a dashboard request or a long wait for engineering resources.

Agentic software changes that by allowing workers to ask direct questions in plain language and receive actionable answers immediately. In Rauch’s view, that is where AI begins to change how companies function internally, not just how engineers write code.

“The bottleneck for people like her has not been creativity or intelligence,” Rauch said of a sales worker example inside Vercel. “It’s been data.”

He also said that these tools can make company-wide improvements possible because the same underlying technology can support both customer-facing applications and internal productivity systems. In his words, the agent architecture is the same; only the APIs and use cases differ.

Vercel’s answer: Eve and Sandbox

To make agents more usable in real companies, Vercel has developed two pieces of infrastructure it believes are especially important: Eve and Vercel Sandbox. Together, they are designed to make agent behavior easier to describe, control and audit.

Eve is a framework for defining an agent’s instructions and capabilities in natural language. Rather than forcing teams to encode everything in rigid software logic, the system aims to let organizations express what an agent should do and what skills it should have in a more flexible format.

Sandbox, meanwhile, is Vercel’s isolation layer. Rauch described it as a controlled environment where an agent can operate with enough freedom to be useful, while still being constrained by policy. It is meant to prevent agents from wandering across the wrong data sources or leaking information beyond approved boundaries.

Why sandboxing matters

For Vercel, Sandbox is more than a technical feature. It is a response to a broader industry concern that AI tools can unintentionally expose sensitive information when they are given too much access. Rauch said this is especially relevant for coding assistants that may interact with entire repositories or enterprise systems.

The issue is not only malicious behavior. It is also about configuration mistakes, incomplete understanding and misplaced trust. In a company setting, users may install tools that seem helpful without fully appreciating what data those tools can read, send out or retain.

Rauch used the example of large industrial organizations with highly specialized codebases to illustrate the risk. For firms in aerospace, for instance, source code can represent years of accumulated engineering knowledge. If that code is fed into a developer tool with the wrong settings, the result could be a serious loss of control over sensitive intellectual property.

That is why the company sees policy enforcement and audit trails as essential. Companies want to know what an agent accessed, what tools it called, which permissions it used and whether the workflow can be reviewed after the fact.

The end of single-vendor thinking

One of Rauch’s strongest claims was that the market is moving away from the idea that companies should build everything around a single AI lab. In the previous phase of the boom, he said, many businesses chose one primary partner and assumed the rest of the stack would follow. That era is ending.

Instead, buyers are beginning to treat AI systems like standard software architectures. They may use one provider for a model, another for orchestration, another for data, and another for deployment. The result is a modular stack that can be adjusted as prices, capabilities and workloads change.

Rauch said production deployment is the main reason for this shift. Once companies care about throughput, reliability and economics rather than novelty, they start comparing models on a cost-per-performance basis. That opens the door to more variety in the market.

Rauch said the data is increasingly showing demand for a broader mix of model providers, including Google’s Gemini models and open alternatives such as DeepSeek and GLM-5.2.

He argued that the growth of those systems is not necessarily a sign of hype or branding power. It is a sign that companies are choosing tools that perform well in practical environments.

Why Gemini and open models are gaining ground

Rauch’s comments suggest that the competitive landscape is changing in ways that are not always obvious from the headlines. While OpenAI and Anthropic remain dominant names in consumer and enterprise AI discussions, he said other options are attracting more attention inside production teams because of economics and flexibility.

Google’s Gemini family, he said, has strong price-performance characteristics that appeal to teams optimizing at scale. Meanwhile, open models such as DeepSeek and GLM-5.2 are drawing adoption from organizations that want more choice and control over deployment decisions.

That broader model mix matters to infrastructure companies because it means customers need abstraction layers. If a business wants to switch models based on cost or latency, it needs systems that make that possible without rebuilding the whole application.

In Rauch’s telling, that is where Vercel’s value proposition becomes stronger. The company is not simply a host for applications. It is a bridge between models and the products built on top of them.

When the labs become competitors

The rise of model providers into adjacent territory is creating tension across the industry. As AI labs expand their capabilities, they increasingly overlap with infrastructure companies that have long owned adjacent pieces of the stack.

Rauch acknowledged that this is happening already. He pointed to product moves from OpenAI that make it easier to publish web experiences directly within the company’s environment. From one perspective, that is a natural evolution for a platform that wants to keep users inside its own ecosystem. From another, it is an encroachment on traditional hosting and deployment tools.

Vercel’s view is that this kind of overlap can still work in its favor. Rauch argued that if ChatGPT becomes known as a tool for building websites, then users will increasingly think about web hosting, deployment and performance — areas where Vercel wants to be the recommendation that follows.

But he also made clear that there is a broader strategic question at stake. If model providers continue to absorb more of the workflow, will agents remain independent software components, or will they become tightly coupled to a single vendor’s platform?

The real question: coupling or modularity

For Rauch, this is not just a commercial issue but an architectural one. He believes the future of AI will be shaped by whether models and agents are integrated into a single closed system or separated into interoperable components.

His preferred answer is the modular one. He sees the current moment as a return to software engineering principles that have long guided the broader computing world: use building blocks, not monoliths; choose the best module for the job; replace parts when better ones appear.

That philosophy also explains why Vercel is emphasizing open protocols. If the market standardizes around interchangeable parts, then companies can avoid lock-in and continue moving quickly as the AI ecosystem evolves.

Rauch said Vercel wants to be “the AWS of this generation,” a shorthand for a platform that underpins the application layer while still preserving openness and flexibility.

That is a bold comparison, but it captures the scale of the ambition. Vercel is not just selling hosting, in its own view. It is trying to define the infrastructure category that will support the agent economy.

What the numbers say about Vercel’s position

The company’s scale helps explain why Rauch believes it has leverage in the current AI transition. Six million deployments a day is a significant workload by any measure, and the fact that half of those deployments are triggered by coding agents suggests that AI is already deeply integrated into the company’s core traffic.

Even more striking is the volume of token traffic moving through the gateway. More than 1 trillion tokens a day indicates that Vercel is handling a massive stream of model interactions, giving it visibility into usage patterns, performance demands and customer behavior.

Those figures also show how quickly AI usage can become infrastructure usage. What begins as a model selection question quickly turns into a networking question, a permissions question and a deployment question. Companies need places for agents to run, policies to govern them and systems to keep track of what they do.

In that context, Vercel’s strategy appears to be built around a simple thesis: once AI moves into production, the winner is not necessarily the company with the most famous model. It is the company that can help organizations ship safely, quickly and repeatedly.

Why the enterprise angle matters

Much of the public conversation about AI still focuses on consumer chatbots, creative tools and headline-grabbing model launches. Rauch’s comments point to a different center of gravity: enterprise operations. The businesses that feel the strongest pull from AI are increasingly those that want to change how work gets done inside the company.

That includes sales, support, engineering, operations and analytics. In each case, the promise is not that AI will replace entire teams overnight. It is that agents will remove friction, expose hidden information and reduce the number of steps required to act.

But to do that, agents need permission to touch real systems. And once they do, they become a governance problem as much as a productivity tool. Vercel is betting that this is where the market is headed and that its infrastructure will become part of the default stack companies use to manage that complexity.

The hidden pressure on SaaS vendors

Rauch also suggested that the rise of agents could force major SaaS platforms to change how they think about data ownership. Many enterprise software companies have historically benefited from keeping user information inside their own systems, where customers must navigate the vendor’s interface to get anything done.

Agentic software changes that equation. If a worker can ask an assistant to pull data, compare accounts or generate actions across systems, then the software layer that controls access to information becomes more valuable than the old interface layer.

That is one reason he believes agents are “forcing companies to open up.” In his view, a world built around trapped data is increasingly incompatible with a world built around dynamic, tool-using software.

The implication is that enterprises may need to rethink the trade-off between convenience and control. AI tools are useful precisely because they reduce friction, but that utility can only be realized if companies are willing to expose data in a structured, governed way.

The broader industry context

Rauch’s remarks arrive at a moment when the AI industry is wrestling with its own next phase. The initial wave of fascination centered on model capability. The current wave is about deployment, reliability and return on investment. That is a less glamorous but more consequential phase, because it determines which products survive beyond the pilot stage.

This evolution is visible across the sector. Companies that once marketed themselves purely as model providers are expanding into apps, development environments, hosting, search and productivity tools. Infrastructure companies, in turn, are adding AI gateways, policy layers and agent runtimes.

The lines are blurring, and Rauch’s comments reflect that. In his view, the competition is not simply between one lab and another. It is between open, modular systems and closed, vertically integrated ones.

He believes the market will reward flexibility. Teams want to choose the best model for a given task, switch providers when economics change and keep their own data and policies in the middle. That kind of architecture gives them leverage. It also gives infrastructure platforms a reason to matter.

What comes next for Vercel and AI infrastructure

Vercel’s long-term bet is that the company will remain central as AI becomes more operationally embedded. If agents continue spreading through code generation and internal workflows, demand for secure execution environments, observability and model-agnostic gateways is likely to rise.

That would put Vercel in the middle of a fast-expanding market in which model vendors, infrastructure platforms and enterprise software companies all want to own a piece of the agent stack. Rauch’s vision suggests he believes the platform layer can win by staying open while becoming indispensable.

Still, the path is not guaranteed. If major labs continue to absorb more deployment and workflow functionality, they could reduce the need for third-party layers. If enterprises become comfortable with a smaller number of tightly integrated vendors, modularity may matter less than convenience.

For now, though, the direction of travel appears to favor companies that can help organizations separate models from agents without slowing them down. That is the market Rauch is trying to build for.

Timeline of the shift from experimentation to production

The bottom line

Rauch’s message is that the AI market is entering a more mature and more contested stage. The excitement around models remains high, but the durable value may lie in the systems that make those models safe, flexible and practical in the real world.

For Vercel, that means competing not just with infrastructure rivals but with the labs themselves. It also means betting that the future of AI will be built on modular software, governed access and open protocols rather than closed ecosystems.

If Rauch is right, the biggest winners in the next phase of AI may not be the companies that build the smartest models. They may be the ones that help everyone else deploy them responsibly.