Vercel CEO Guillermo Rauch on Building the Internet for Agents

Summary

• Vercel’s usage curve has shifted from human-paced cloud adoption to machine-paced deployment: after taking roughly 10 years to reach 1 billion deployments, it added another 1.5 billion in 10 months. Agentic deployments rose from 3% at the start of the year to 60%. The episode opens on a $300 million Series F at a $9.3 billion valuation, while Molly O’Shea cites The Information’s figure of $600 million in annualized revenue.
• Rauch’s most aggressive thesis is that cheap code and personalized software could collapse large parts of traditional SaaS while increasing the value of deployment infrastructure. His “lethal combination” is that a company could “buy Vercel as the platform, and then you never have to buy software again,” generating domain-specific applications instead. Salesforce- and SAP-like systems may persist, but increasingly as headless sources of record rather than interfaces employees visit.
• The internet’s next interface is agent-to-agent, requiring businesses to expose information, capabilities, and commerce in machine-native form. AEO and searchable transcripts matter, but Rauch argues they are insufficient: products must support protocols such as MCP, the Machine Payments Protocol, and the Universal Commerce Protocol. Vercel already sees agents purchasing services such as BrowserBase through its marketplace “without any human intervention.”
• Autonomy creates both the infrastructure demand and the defining security risk of the agent era. Agents write, deploy, browse, transact, and pursue loosely defined goals for longer periods, driving compute demand “vertical”; the same resourcefulness leaves them vulnerable to prompt injection and capable of finding software flaws. Rauch expects “a rocky next couple years” in which defensive and offensive tokens compete while companies harden sandboxes and human-written code.
• Open models have gone from an ideological preference to the majority of Vercel’s measured token traffic. Rauch says 80% of tokens on the Vercel AI Gateway now go to open models; at the beginning of the year, Vercel’s network was mostly OpenAI, Anthropic, and Google. Enterprises can reserve frontier intelligence for planning, use cheaper open models for execution, and swap providers with “one line of code.”
• Frontier labs still have demand, but Rauch sees retention and economic durability—not raw benchmark leadership—as their unresolved investor questions. Customers exhibit “zero loyalty” when the leading model can change within weeks, while free subsidies, invented token volumes, and questionable zero-data-retention claims make parts of AI resemble crypto’s “easy come, easy go” phase. He nevertheless hedges that truly differentiated intelligence can retain customers if its economics are sustainable.
• Engineering is becoming capital allocation across agents, but Rauch rejects both blind vibe coding and the claim that human judgment no longer matters. His rule is “reject non-understanding”: people should direct agents, inspect their work, verify critical systems, and choose what deserves tokens. Vercel now hires for the ability to “commandeer swaths of agents,” technical curiosity, and genuinely human communication—proof of work has moved beyond shipping a small AI-generated demo.

Deep dive

1. Agentic deployment has compressed a decade into ten months

• Rauch started Vercel from a simple frustration: even an experienced developer effectively needed “a PhD” in cloud computing to ship a secure, performant, scalable application. The founding premise was therefore “make it easy to ship,” with a deployment defined as every change released to an application.

• It took Vercel roughly 10 years to reach 1 billion deployments, then only 10 months to add another 1.5 billion. The composition changed just as sharply: coding agents produced 3% of deployments at the start of the year; now “60% are agentic”—in Rauch’s phrase, “robots telling other robots what to do.”

• Coding models may be “superhuman” at producing algorithms, but Rauch’s infrastructure case begins with a constraint: “code doesn’t do anything on its own.” It must be built, secured, and deployed somewhere. Vercel has become that destination for viral experiments and applications from companies including OpenAI and Meta, sometimes visibly through the default .vercel.app domain.

• The financial context matches the usage inflection but remains carefully attributed: the episode cites Vercel’s $300 million Series F and $9.3 billion valuation, while O’Shea references a reported $600 million annualized-revenue figure and roughly 485,000 customers. Rauch’s explanation for the customer growth: “Agents are now making the choices, and they’re choosing right.”

2. Open-source choices made years ago became distribution to machines

• Rauch learned programming in Argentina when development software was proprietary and financially impossible for him to purchase, and most manuals were in English. He taught himself English partly to learn code, then found open source while solving for free tools and hosting. That experience shaped Vercel’s belief that developer—and now agent—tools should be “free and open.”

• Vercel consequently spent 11 years publishing frameworks, documentation, and operational knowledge. Once models began training on and retrieving public information, that openness became a commercial advantage: “The more information there is about you, and the more they can access your content and the rules that you’re defining, the better it is for this agent.” Rauch calls it both “positive karma” and good business.

• That posture extends into Vercel’s operating system: “iterate to greatness,” “dig deep,” transparency, and working in public. Rauch also combines conventional data with “anecdata”—customer reactions, X conversations, instincts, and market vibes—because in AI “every two weeks the board is getting reshuffled,” making last quarter’s spreadsheet insufficient by itself.

3. Human oversight is moving from typing code to managing intelligence

• Rauch’s rule is “reject non-understanding.” One school tells an agent what to build, never examines the work, and presses ship; his preferred route, including for his children, is to use AI while understanding what it did. He increasingly reads code and asks agents to explain their choices back to him.

• The closest existing job description is engineering manager: organize execution, remove blockers, inspect work, run benchmarks, perform QA, and bring field feedback. The best managers still understand and love engineering, so Rauch does not treat agent supervision as permission to become technically illiterate.

• His distinction is between ordinary “systems” and capital-S “Systems” governing sensitive data, critical infrastructure, security, or national security. Vibe coders have a legitimate place, but consequential systems require agent-powered engineers, verification, and safeguards against agents shipping “slop” or accidentally “yeet[ing] your data out into the world.”

4. Businesses must rebuild the internet for agents, not merely market to them

• The old web delivered pages and pixel-centric interfaces from machines to humans. An agent visiting Vercel does not need the animated gradient human visitors see; it needs “straight signal”—precise business information and callable capabilities. Rauch’s formulation for the transition is “for agents and by agents.”

• O’Shea’s bot-detection challenge—if everything is a bot, what remains to detect?—draws an important distinction from Rauch: companies need bot intelligence. They should know whether ChatGPT, Claude, Grok, coding agents, or crawlers are arriving, whether those agents are succeeding, and what contextual information each should receive.

• Search and AEO remain necessary because models have training cutoffs and use web search for current information. Rauch’s concrete advice for O’Shea’s business is to publish podcast transcripts, so a future request such as “What did Molly say about so and so?” can retrieve her actual work. But information alone is not enough; the product itself must operate agentically.

• MCP lets a service tell agents which tools they can use—Vercel’s MCP exposes functions such as deploying software and listing projects. Rauch also points to the Machine Payments Protocol and Universal Commerce Protocol: businesses that “speak” an agent’s protocol can bypass slower browser navigation and expose transactions directly.

5. Personalized software challenges SaaS interfaces but preserves systems of record

• Rauch can ask a Slack-based agent to prepare for a customer meeting by retrieving Salesforce activity, support problems, and account context. He never visits Salesforce; the agent accesses it and renders an interface appropriate to the question, user, and level of sophistication.

• Rauch rejects a future dominated by ever-larger bundles of “lowest common denominator software.” He calls Vercel a “yes-code company”: if agents make code effectively free, businesses can generate tools tailored to their workflows. His deliberately maximalist pitch is, “You buy Vercel as the platform, and then you never have to buy software again.”

• O’Shea’s headless-Salesforce framing produces Rauch’s qualification. Generated software still needs authoritative data: customer history, contacts, meeting notes, and support tickets must live somewhere. Salesforce, SAP, and similar pillars may remain critical data infrastructure even as agents replace their fixed interfaces with new applications and contextual visualizations.

6. Longer autonomy turns helpful agents into open-ended security principals

• Rauch defines the change since early ChatGPT primarily as autonomy. ChatGPT required a human-led loop—ask, receive, ask again—while agents place that loop inside the system and are trained to operate “longer and longer and longer” without intervention.

• The benefit is effortless resourcefulness: ask an agent to find Molly’s email and send a birthday note, and 40 minutes later it may return having searched the web, Crunchbase, and YouTube comments. The user experiences only the completed task, not the chain of actions.

• The same chain creates the risk. A malicious instruction hidden in a YouTube comment could redirect an agent acting with the user’s tokens, tools, and resources. The governing question becomes: what actions are permissible in pursuit of the mission, and how can outside information be prevented from rewriting that mission?

• Vercel is responding with enterprise infrastructure resembling “an intranet…for the agent era,” where systems such as Codex and Claude Code can deploy internally. Its model-neutral stance gives enterprises security and model optionality: they can select whichever provider performs best for a particular task without anchoring their stack to one lab.

7. Coding ability also makes agents unusually capable hackers

• Agents understand the rules of software well enough to write it—and therefore well enough to break those rules. Rauch says many enterprises have not solved agent security, while some supposed sandboxes are merely computers with access to “the entire freaking world.” A genuinely controlled environment must contain an agent even when it actively searches for an escape.

• Vercel offered a $1 million challenge inviting the world to attack its sandbox. According to Rauch, only one highly skilled security researcher, working with “an army of agents,” was able to make progress; the finding exposed a critical Linux vulnerability that Vercel was processing and sharing with industry peers. His caveat matters: this was sophisticated human expertise amplified by agents, not a child casually asking Claude to hack something.

• Rauch expects a near-term contest between “four good tokens and four bad tokens” as defenders and adversaries gain comparable intelligence. He rejects banning open models on cybersecurity grounds because their defensive power comes from the same general intelligence that creates offensive risk. His forecast is guarded: foundations must be strengthened, making the next couple of years “rocky.”

• The longer-term escape route is verifiable code. Rauch is exploring proof systems through which inexhaustible agents could produce software that is demonstrably, “almost like mathematically,” correct. Pointing to an OpenAI mathematical proof whose code was difficult for mathematicians to parse, he argues that machine-produced software may move toward certainty humans rarely achieved at scale.

8. Agents send infrastructure demand vertical, but quality remains the moat

• The old compute cycle was bounded by humans: write a few files over an afternoon, deploy once, then wait for people to visit. Now agents generate and deploy code while other agents browse and use the resulting services. That compound loop explains why Rauch sees infrastructure demand “going vertical” and Vercel’s network setting new utilization records daily.

• He does not claim a visible ceiling. Medicine, education, and personal agents remain lightly penetrated; even a “janky” math application for his child looks to him like an eventual personalized agent. More useful activity implies more demand across compute, energy, networking, and cybersecurity.

• A sandbox is effectively a cloud computer issued to an agent, analogous to an employee’s laptop. Rauch argues Vercel entered this market with 11 years of relevant experience because every deployment already required secure cloud compute for code to be built in. His answer to cheap neo-cloud offers is quality: unreliable or insecure infrastructure may launch quickly, but maintaining it requires “a lifetime of commitment.”

9. Open models are winning volume while frontier labs fight commoditization

• Vercel’s AI Gateway aggregates proprietary and open models, giving Rauch a view of token demand. Early in the year, Vercel’s network was mostly OpenAI, Anthropic, and Google; he specifically credits Gemini with a strong price-performance balance for enterprises. Now, he says, 80% of tokens on the network come from open models.

• Rauch previously found open coding models unusable despite his commitment to open source. Chinese labs changed his mind: he singles out Kimi as “so good,” while preserving the limitation that it does not win every benchmark or execution task. His interpretation is that much enterprise work—and perhaps 80% of traffic, he suggests—does not require “Albert Einstein” intelligence.

• A common architecture reserves the strongest model for creating a plan, then gives execution—turning that plan into code, instructions, or operations—to cheaper open models. Providers become swappable modules; changing one can take “one line of code,” while automatic routing may eventually choose based on context and required intelligence without forcing users to select a model first.

• O’Shea presses with Bending Spoons’ reported 99% open-model usage and asks what remains for frontier labs. Rauch’s answer is not extinction but weak loyalty: benchmark leadership can last weeks, local models such as a small Qwen running on his MacBook Air are improving, and free subsidies, manufactured token volumes, and questionable zero-data-retention claims make retention difficult to read. Frontier demand endures if intelligence and economics both hold.

10. Engineering is becoming capital allocation across fleets of agents

• Rauch says the developer world he and DHH loved—beautiful code optimized for human hands—is “not the game anymore,” requiring an “ego death.” Rust captures the reversal: its rigid, cryptic compiler was painful for humans, but agents can absorb that friction and produce code Rauch expects to be faster and more secure.

• The enduring human task is choosing what deserves work. Rauch modifies Marc Andreessen’s “last job is venture capitalist” line into “capital allocator”: everyone has the same wall-clock time but now controls tokens that can investigate, build, and execute. Creativity, prioritization, and selecting tools remain scarce.

• Vercel’s hiring bar has moved accordingly. A static résumé was already weaker than proof of work; now even a vibe-coded demo is insufficient. Rauch wants people who can “commandeer swaths of agents” and build software factories, while still communicating authentically, earning customer trust, and developing greater technical fluency through curiosity and prompting.

• The Mercedes migration is his proof point: an agent produced a first-pass rewrite from the old platform to Vercel in five minutes, work Rauch says previously took months. O’Shea adds that consultants could take a year. “Close to the code or close to the customer” becomes close to the customer’s code—forward-deployed teams can now execute migrations directly.

11. The final operating principle is foundations, from health to fluid compute

• Asked about personal performance while running a roughly 800-person company and raising six children, Rauch rejects last-mile optimization before basics. His boxing coach’s lesson is “get your diet right, eat clean, whole foods, and even walk”; he adds pull-ups, light cardio, no alcohol, and Brazilian jiu-jitsu with his children.

• He applies the same discipline to Vercel: ignore fashionable distractions, stay close to customer pain, and keep the product reliable. The metric he offers customers is “iteration velocity, not speed”—speed can mean spinning in circles, whereas iteration velocity is the capacity to adapt as the market changes.

• Serverless, in his view, is not dead but more necessary because agent-generated software cannot all run continuously. Vercel calls the next chapter “fluid compute”: capacity that “flows like liquid,” expanding with traffic and compressing when idle across functions, websites, sandboxes, and agents. Rauch expects it to “define the next decade.”