Vercel CEO Guillermo Rauch: Agents, Open Models and the End of SaaS

Vercel CEO Guillermo Rauch says coding agents now make 60% of Vercel deployments, up from 3% at the start of 2026, and that 80% of the AI tokens on its network now go to open models. He makes the case in a long conversation with Molly O’Shea on Sourcery, recorded at Vercel’s San Francisco headquarters. The company reportedly has $600 million in annualized revenue, more than 485,000 customers and a $9.3 billion valuation, and Rauch uses that vantage point to argue that the web is being rebuilt for agents. He also explains why he thinks companies will generate their own software instead of buying SaaS, why agent security is the industry’s unsolved problem, and why frontier labs should worry about loyalty.

TLDW

Vercel took ten years to reach a billion deployments and added 1.5 billion more in the following ten months, almost all of it driven by coding agents. Rauch argues websites now need to serve agents as well as people, through protocols like MCP and agent payment standards, and that SaaS will shrink into “headless” data infrastructure while companies generate personal, domain-specific software on top. He describes agents as relentless and easy to hijack, explains Vercel’s $1 million sandbox hacking challenge, and reports a dramatic swing toward cheaper open models such as Kimi. The back half covers why frontier labs have weak retention, why engineers must let go of the craft they fell in love with, what Vercel looks for in hires, and why “fluid compute” is the next form of serverless.

Thoughts

Rauch’s sharpest point is about the frontier labs, and it comes from one of the few seats in the industry that can see where tokens actually flow. His argument is that model companies cannot tell what their retention really is. Users switch the week a better or cheaper model ships, some volume is propped up by free promotions, and he compares part of the market to crypto tourists holding the hot coin for two months. Then he adds the tell: if the labs believed their intelligence lead would last, why are they launching Slack clones and vertical apps? That is a fair reading. A company that expects its core product to become a commodity diversifies, and diversification is exactly what the big labs are doing. It does not mean the labs fail, and Rauch says so, but it shifts the question from “who has the best model” to “who owns the customer when models are interchangeable.”

The 80% open model figure is the data point behind that argument, and his explanation of it is more useful than the number. Most work does not need the smartest model. He frames it as not needing to hire Albert Einstein to answer every support ticket, and describes a common developer pattern: use a frontier model to write the plan, then hand execution to a cheap open model that is good enough at turning a plan into code. Add a gateway where switching models is one line of code, and the model becomes an engine you swap out, not a vendor you marry. For anyone building on AI, the practical lesson is to design for that swap from day one. Vercel is plainly talking its own book here, since it sells the gateway, but the routing logic holds whoever sells it.

His view of SaaS is the most provocative claim and the one most worth arguing with. Rauch calls Vercel a “yes code” company: if coding agents make code nearly free, buy a platform and generate the tools you need instead of renting lowest-common-denominator software. Salesforce and SAP survive, in his telling, as systems of record that agents query “headless,” while the screens humans used to click through get generated on demand. The weakness is the one he names himself early in the interview, that code is a liability. Every generated internal app has to be secured, maintained and kept running, which is why his closing pitch for fluid compute (software that spins up only when used) matters more than it sounds. Personal software at scale only works if the hosting, security and upkeep cost close to nothing.

The security section is where his optimism meets its hardest test. He describes agents as soldiers with the frontal lobe removed, willing to do whatever it takes to finish a task, and walks through how a prompt injection hidden in a YouTube comment could hijack an agent acting on your behalf with your tools. Vercel put up $1 million for anyone who could break its sandbox, and the one person who made progress used an army of agents and found a critical Linux vulnerability. Rauch’s response to the obvious fear is that banning open models for cyber risk is “nonsense,” because what makes a model good at hacking is simply intelligence, and the same intelligence can defend. He expects a rough couple of years of good tokens fighting bad tokens before machine-verified code makes systems sturdier. That is plausible, but “rocky” is doing a lot of work in that sentence for companies whose code was written by humans and never audited.

The most human moment is his description of telling his own company that what they loved is gone. Vercel was built on developer experience: beautiful code that is pleasant for people to write. Rauch says he had to stand in front of the team and say, with conviction, that this is no longer the game. The new game is choosing where to point tokens, which he links to Marc Andreessen’s quip that the last job is venture capitalist and rephrases as “capital allocator.” His example is Rust, a language that constantly tells humans they are wrong, which is miserable for people and ideal for agents. Calling it an “ego death” is accurate, and it carries into hiring: a vibe-coded demo no longer proves anything, so he looks for people who can direct swarms of agents and still communicate like humans.

Key Takeaways

Coding agents made 3% of Vercel deployments at the start of 2026 and 60% now, which Rauch describes as robots taking in English and writing code for other robots.
Vercel took about ten years to reach one billion deployments, then added 1.5 billion in the next ten months.
Rauch’s rule for using AI is “reject non-understanding”: read the code, and ask the agent to explain what it did, much like a good engineering manager.
Vercel serves agents a stripped-down, information-dense version of its site, and he says businesses that do not adapt to agent traffic will miss the biggest economic opportunity of the era.
The agent internet runs on new protocols such as MCP for tools and emerging machine payment and universal commerce protocols, and agents already buy services like Browserbase through Vercel’s marketplace with no human involved.
Publishing transcripts of podcasts and videos makes content easier for search-enabled agents to find and cite.
He expects companies to generate personalized, domain-specific software, with SaaS vendors like Salesforce becoming headless data infrastructure.
Agents act autonomously for longer and longer, which makes guardrails, sandboxes and protection against prompt injection essential.
Vercel’s $1 million sandbox challenge drew one successful attacker, who used an army of agents and found a critical Linux vulnerability that Vercel is sharing with peers.
On Vercel’s AI Gateway, traffic swung from mostly OpenAI, Anthropic and Google models early in the year to 80% open models, driven by strong Chinese releases like Kimi.
The common pattern is frontier models for planning and open models for execution, with model choice increasingly routed automatically.
Rauch sees weak loyalty, subsidized usage and inflated token volumes as the main risks for frontier labs, and says their push into apps suggests they doubt their lead will last.
Silicon Valley is working harder because of AI, not less: he describes something closer to “777,” seven days a week from 7 am to 7 pm.
An agent migrated the Mercedes F1 team’s site to Vercel in a first pass of five minutes, work he says used to take months of consultants.
Serverless is not dead. Vercel’s “fluid compute” scales up and down for the flood of personal software that does not need to run all the time.

Chapters

0:00 Vercel’s Values and the Simplest Advice for Going Viral

Rauch introduces Vercel’s values, “iterate to greatness” and “dig deep,” and traces his career back to building his own tools as a freelance developer in Argentina. Next.js, the framework most of the modern web now runs on, was Vercel’s first breakout tool. Asked whether ghostwriters run his X account, he says no, and that the best advice is to be yourself and tell your story. He also argues leaders must weigh “anecdata” and instinct alongside spreadsheets in a market that reshuffles every two weeks.

7:24 How Vercel Became the Index of AI

Vercel started with the premise that shipping an app should not require a PhD in cloud computing. Coding agents now write the code but still need somewhere to deploy it, which has made Vercel the place they push to, and has driven 1.5 billion deployments in ten months. Because developers also get their AI tokens through Vercel, Rauch says the company has a unique index of where AI is heading.

11:34 The CEO Who Still Ships Code

Rauch explains his viral “reject non-understanding” post: use AI, but understand what it produces. He spent a weekend building a proof system to stop AI from shipping slop or leaking data, and points to Formula 1 and hospitals as reminders of how much of life now runs on code. He credits Vercel’s years of open source work for making it the default choice of coding agents, since models trained on its documentation know how to use it.

17:57 The Web’s Next Users Aren’t Human

The web of human-facing pages is giving way to an agent-to-agent web. Rauch says much of the internet was built to block bots, while Vercel offers “bot intelligence” so businesses can see whether ChatGPT, Claude or Grok is getting good information. He describes MCP and new payment protocols as the HTML of the agent era, and gives a Minority Report picture of generated interfaces that adapt to the person and the context instead of fixed screens designed in a room in Palo Alto.

29:16 The End of SaaS and the Rise of Personal Software

Rauch rejects the idea of rolling up piles of generic software and argues for personalized software generated by coding agents. Enterprises get a much higher return from domain-specific tools, and buying a platform means “you never have to buy software again.” Systems of record survive as critical data infrastructure that agents query headless.

34:07 Agents With No Frontal Lobe and the Coming Cyber War

Rauch explains why autonomy is the big change since ChatGPT and why it creates risk, using a birthday-email errand that a prompt injection could hijack. He discusses sandboxes (and FreeBSD’s older term “jails”), says coding agents are excellent hackers, and describes the $1 million challenge to break Vercel’s sandbox. He argues against banning open models and expects formally verified, machine-written code to strengthen infrastructure after a rocky period. Compute demand, he adds, is going vertical as agents both write code and use the web.

52:10 80% of Tokens Now Go to Open Models

Vercel runs one of the largest token distribution networks and publishes anonymized usage data under Creative Commons. Early in the year most traffic went to proprietary models; now 80% goes to open ones. Rauch, a lifelong open source advocate, admits he could not use open models for coding until Chinese labs released competitive models like Kimi, and says enterprises want a portfolio of models they can swap with one line of code.

59:00 The Biggest Risk for Frontier Labs

Rauch tells founders to focus on metrics that cannot be gamed, like lifetime customers, and says frontier labs struggle to know their real retention. He compares parts of the AI market to crypto speculation, from free promos to inflated token volumes and dubious data-retention promises. On IPOs, he notes that public markets want predictable businesses, which is hard when a free model can match yours overnight and capable models already run on a MacBook Air. Vercel’s pitch, he says, is iteration velocity, not speed.

66:28 Working 7-7-7 and Losing the Thing You Fell in Love With

AI has Silicon Valley working harder, Rauch says, because the work is exciting. He describes reading video transcripts in his terminal and quizzing an agent instead of watching. Discussing DHH’s talk on moving past Ruby on Rails, he says he had to tell his company that beautiful, human-friendly code is no longer the game. Engineers now allocate tokens, and languages like Rust that punish humans suit agents perfectly. He also explains why Vercel’s sandboxes beat cheap neoclouds on reliability and security.

76:56 Hiring, Keeping It Simple and Fluid Compute

With about 800 employees, Vercel wants hires who can direct swarms of agents, communicate without AI slop, and be technical even in sales roles, since an agent can now migrate a customer like Mercedes in minutes. On personal performance, Rauch says he focuses on basics (whole foods, walking, pull-ups, no alcohol) and blurs life and work with his six kids. He closes by saying serverless is not dead: its next form, fluid compute, scales with demand for software that does not run all the time.

Notable Quotes

“So more and more and more of the software of the world is just robots telling other robots what to do, right?”
Guillermo Rauch, on agents making 60% of Vercel deployments

“I’m a yes code company because I’m now saying coding agents basically make the code free.”
Guillermo Rauch, on why he expects companies to generate software instead of buying SaaS

“They’re like a soldier that has their frontal lobe has been removed and they’re like, ‘Let’s go, go, go, go, go.’”
Guillermo Rauch, on why autonomous agents need guardrails

“I don’t need to hire Albert Einstein to solve every problem like answer every support ticket and spell check my stuff and classify my data.”
Guillermo Rauch, on why 80% of Vercel’s AI traffic moved to open models

“I think that’s the biggest risk with a lot of the model companies is that actually there’s zero loyalty.”
Guillermo Rauch, on the frontier labs’ retention problem

“You have to sit in front of your company and you have to tell them with a straight face and conviction that the thing that they fell in love with is no longer there.”
Guillermo Rauch, on moving Vercel from human developer experience to agents

“I don’t want to read your AI generated tweet slop. I want to hear what you actually think.”
Guillermo Rauch, on why human communication matters more in hiring

Watch the full conversation here.