Vercel CEO Guillermo Rauch on Building the Internet for Agents
- 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.
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.appdomain.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.”
Full transcript
Vercel recently raised $300 million in a Series F funding round. It brings the cloud company's valuation to $9.3 billion.
Vercel seems to be winning. You guys are at a $9.3 billion valuation. The Information reported you guys at $600 million in annualized revenue.
When you're a developer or a company of any size, you can deploy these agents and make your business more efficient. We started out the year with 3% of deployments on Vercel coming from coding agents. Now 60% are agentic.
Oh my God.
It took us about 10 years to get to a billion deployments, and then in the next 10 months, we've added another 1.5 billion deployments. So what took us 10 years, we've now surpassed in just a few months.
Almost every other thing I see on X every day is just a Vercel deployment. Sometimes you see the largest companies in the world deploying on Vercel, including companies like OpenAI and Meta. Agents are now making the choices, and they're choosing right. The lethal combination there is that you buy Vercel as the platform, and then you never have to buy software again.
G, welcome to Sorcery.
Thanks for having me.
Well, thanks for having me in your really cool studio.
We're here in San Francisco, on the seventh floor of the Vercel headquarters.
I mean, this is insane. You have multiple studios in here—
Yeah.
—as well. I was talking to your crew before, and you have this acronym of, like, you have to do everything at the best. What is that?
Well, we have a couple of company values. One is ITG: iterate to greatness.
Mm-hmm.
So once upon a time, recording out of Vercel was not this good. We're always iterating to greatness. The other one is "dig deep." Whether you get a customer complaint or you find a bug, go as deep as humanly and agentically possible.
Say more about that.
When I started Vercel, I was a framework and dev tools builder.
Mm-hmm.
My passion has always been: can I build a better hammer? Can I build a better power drill, but for software developers?
Back when I was a kid in Argentina, I would be given a project. This is kind of fun: I started out doing freelance gigs as a software developer. Any time I would be given a job, I would do the job, and I would also build my own tools to do the job, for two reasons. One is that I loved it, and the other one is that with each incremental job, I would be able to do it more easily and faster. In fact, it ended up being true. My lifetime has been about creating better tools. When I started the company, the first tool that became super, super popular was Next.js.
Mm-hmm.
We created the framework that almost everybody in the world has now used to build modern web experiences. But I also see the company itself as a product, and the best products are built with the best tools and the best frameworks.
Over the years, people have asked me a lot, "What is your role at the company these days?" I tell them, "I do a little bit of everything. I do QA, I talk to customers, et cetera," but I invest a lot in the machinery that powers the company itself—the machinery that yields innovation in the future.
I think one of the key constituents of this framework of how I run the company is our values and our mission statement—the classic things. We put quite a bit of effort into who we are and what our values are.
Is this how you became and continue to be very popular on the internet? Nick and I were talking about this, but you are a celebrity.
I'm a nerd celebrity of sorts. But I'll tell you, one of my personal values is transparency and working in public. A lot of my personal traits have permeated through the company as well: open source, open models, building in public, transparency.
Mm-hmm.
The reason I think I've had success communicating with people over the internet is that I just tell it like it is. People have asked me if I have ghostwriters for my X posts. I also spend a lot of time with young founders. Another value of mine and of our family is giving back.
Mm-hmm.
I spend a lot of time trying to give back to the ecosystem, and I invest in a lot of young founders. Almost every week I get, "Who writes the tweets for you?"
Really?
"Can I hire a firm?" I tell them, "Look, it's just..." It sounds like the worst advice because it's so simple, which probably makes it good advice. Just be yourself. Tell your story, share your truth, and don't overthink it too much. I think that's how some of the bangers have happened online.
But you go viral very constantly. Do you check your X analytics? I know I'm going off on a tangent, but I'm curious now.
No, no—
I—
I do, actually, for the sake of... Another, maybe, toxic trait of mine is that I'm constantly scrutinizing products: how fast they load, how well they represent the UI, whatever.
The only reason I've spent a lot of time in X analytics is that I know the engineer at X who worked on that feature, and I've given him a bunch of feedback. But you kind of feel it. You don't need analytics.
This is another—this is not really a value, but an operating principle of mine: you have to be mindful of the data and the anecdata, especially when running a business in a fast-moving market like AI.
Mm-hmm.
You can't just go off, like, "Oh, yeah, the last quarter's little spreadsheet that the data analysts put together." We were talking before about how every 2 weeks the board is getting reshuffled, and so you have to marry the anecdata, the vibes, and the instincts with what you're seeing from the arithmetic of your business.
I try to go off how I feel, how people feel, and, you know, you do have to adapt and move really fast in this world. I think what's proven really valuable to Vercel has been that we have a mission about enabling builders and enabling people to ship what's next. That means we constantly have to be rethinking and reinventing who we are, how we work, and so I think there's a sense of almost working in real time with the world and having this real-time conversation.
Based on that real-time conversation, you inform your next move.
I don't know if you'd be surprised or not, but when I've had these interviews—and I've interviewed many different CEOs and investors—a lot of them are getting their news and their pulse on what's going on in the AI industry just from X alone.
Yeah.
You'd think it'd be somewhere else, but it's like, no—
Yeah, you would think it's their research teams—
Yeah.
—and there are large enterprises with technical advisors to the CEO who write them reports and memos prior to meetings. But no, a lot of us are just in that global conversation, making our contributions, and then seeing how customers react.
So Vercel does play a role in this. You aggregate and distribute some of that data, but you've also mentioned before that we're in the middle of an exponential—or the exponential. What does that mean?
Yeah. For context, Vercel got really popular because of how easy we made it for anybody on the planet to ship an application.
When I started the company over 10 years ago—almost 11 now—if you sat down, even as an experienced software developer, and you wanted to put an app out into the world that really performed, was secure, and scaled, it was like a PhD that you had to take on cloud computing.
Mm.
At the time, I realized, hey, if it's hard for me to ship an app—and I've been studying the blade since I was pretty much 10 years old in Argentina, programming day in and day out—and if it was so hard for me to ship an app, I couldn't even fathom an enterprise trying to figure out cloud computing, frameworks and tools, shipping pipelines, and all of that.
I started a company with that simple premise: make it easy to ship. It took us about 10 years to get to a billion deployments. Deployment is one of my favorite words. Every time you ship a change to your app, that's a deployment.
In the next 10 months, we've added another 1.5 billion deployments. So what took us 10 years, we've now surpassed in just a few months.
The reason for this is that, one, shipping has gotten easier. Not easier in the sense that the code is actually pretty damn hard, but now we have coding agents writing the code, and I think nobody ever saw that coming.
The people building the AI, the people close to the AI, think what's been the supersonic tsunami is that the machines have taken over. They're better than us at figuring out these algorithms and this code, but they need to ship that stuff somewhere.
When you write the code... Code is kind of a funny thing. I remember telling people that another one of my internal principles is: code is a liability.
Code can have vulnerabilities. Code is how you could delete stuff. You can leak customer data. You have to treat code with responsibility.
Now these coding agents are superhuman at writing the most amazing code. But again, code doesn’t do anything on its own. It needs to live somewhere. It needs to be deployed.
Vercel has emerged as the place, this universal field, where all of the coding agents are plugging into and pushing their code out so that it can be used by anybody. You probably are using Vercel every day of your life, because when you see things on X go viral, it now makes me quite proud to see this. It took us so long to build it, but almost every other thing I see on X every day is just a Vercel deployment.
Sometimes you can tell because, by default, we give people a domain name. It’s like “.vercel.app.” Sometimes you see the largest companies in the world deploying on Vercel, including companies like OpenAI and Meta.
We’ve been at this juxtaposition of all of this, which, to your point, has given us a very unique lens into what’s happening. Not only are we seeing all of the activity coming from the coding agents, we’re also distributing the tokens on the internet.
A lot of companies and developers are coming to us to get their AI tokens, and we offer every model under the planet. You can get OpenAI, Anthropic, open models, closed models, and proprietary models. We’ve emerged as this index that tells you where AI is going.
Are you still coding?
Yes. Now we can get into hot-take territory. One of my tweets that went viral the other day, I called it “Reject Non-Understanding.”
What do I mean by rejecting non-understanding? There are 2 schools of thought today. On one hand, you can tell your AI agent to do whatever, and you don’t even look at what it’s doing. You press a button, ship it, and it ships it to Vercel or wherever.
There’s another route, which is the one that I recommend to my children. I was up doing math homework with one of my children last night. It’s to use the AI, but also understand what it’s doing.
Now, you asked specifically, “Do I write the code?” I read code. What this means is that I’m asking the AI more and more to explain back to me what it did, which is a little bit of the job.
I’ve compared it to being an engineering manager. The engineering manager at any big tech firm in Silicon Valley over the last 10 or 20 years has also been a subject of debate, but for the most part, they help their engineers get organized, help them execute, and get blockers out of the way.
Obviously, the best engineering managers are amazing engineers themselves. Occasionally, they write code. Sometimes they write code to test out what you did and poke holes. Sometimes they might write a little benchmark. Sometimes they may do QA of the final product. Sometimes, in collaboration with the product manager, they might bring feedback from the field.
They’re looking out for you so that you don’t make mistakes. I think the role that most people are now taking with these coding agents is that of an engineering manager.
Again, the best engineering managers that I’ve ever worked with, first of all, love code. They love writing it, understanding it, and so on. I still do it. I also shipped something this weekend that you could call a proof system, a mechanism by which we can make sure that the AI doesn’t ship slop and doesn’t yeet your data out into the world by mistake.
I purposefully read lots of the code. I wrote down a lot of the examples, because there’s a lot of sensitivity around code. AI and code will run more and more of our lives, whether we like it or not.
Formula 1—there was a massive event because they had to ship a software update into the cars, and the drivers were complaining. You realize these moments make you think, “Holy shit, these Formula 1 cars are being run by code.” How the engine works is determined by code.
You’re going to go to the hospital, and all of those systems are run by code. It’s so important that we take these AI safety issues, the quality of the code we produce, and how we verify it extremely seriously.
To be honest about this, how many people—or what percentage of other companies or employees in companies—are actually doing that? I’ve heard people say they don’t even know how to read their code anymore because they don’t know what the coding agent is making.
I’ll tell you a wild stat first. We started out the year with 3% of deployments on Vercel coming from coding agents. Now, 60% are agentic.
Oh my God.
More and more of the software of the world is just robots telling other robots what to do. You can think of it as robots that take in English and yeet out code for other robots to run.
Of that 60%, it’s hard to say, but I will tell you there’s a really cool thing happening right now. When I was a kid in Argentina, I had to work so hard just to write my first line of code.
First of all, a lot of software for writing code was proprietary and financially impossible for me to purchase. I couldn’t pay for it; I couldn’t afford it. Second, most of the manuals were written in English.
It’s a little thing, but most of the world doesn’t speak English. At the time, I had to teach myself English as a way of teaching myself how to code. I had all of these obstacles in the way.
Then I taught myself English, and I had to get around all of these problems: How do I get software for free? How do I host it for free? And all of these things.
That led me to open source and the big epiphany that has defined Vercel: The tools that developers and agents should use should be free and open, so that everybody in the world can learn. It’s such a simple thing.
When you think about why Vercel has emerged as this coding-agent favorite, it’s because over the past 11 years, we’ve open-sourced everything we know. It turns out that when the models get trained, 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 the agent.
It was positive karma that we put out into the world, and it’s also good for business. Essentially, what’s happening that’s really good is that if you were to say, “I want to ship my first line of code today. I want to ship my first app. I want to ship my first idea,” you don’t have to go through the troubles that I had to go through.
That’s miraculous, and I want that to continue to happen. I don’t want to be the guy who says, “You have to understand every line of code and every precise mathematical principle behind these systems.”
But I can tell you, perhaps as another hot take, there are systems, and then there are capital-S Systems. There are the systems that are going to run and host our data, our most critical infrastructure, our security, and our national security.
I do think that there’s a place for agentic-powered engineers, and there’s a place for vibe coders, which is the term that the industry has given to the more playful, “Let me just yeet out whatever from the AI” type of behavior.
To unpack and go deeper into this era of exponential growth, you’ve been running Vercel for almost 11 years now—the birthday is in November. Can you explain those shifts and how they’ve impacted developers, their workflow, and their workday? How has it also impacted you as a business and what you offer them? You’ve now had to open the kimono to multiple different kinds of new products to keep up.
Totally. In the beginning of time, there was the web, and it was made out of pages and human-centric interfaces. You open up a website or an application, and you’re the one consuming. It’s machine to human.
The era that we’re going into now is more agent to agent. Imagine a place where you can deploy interfaces for agents.
A simple example is when an agent goes to vercel.com. I don’t give it the fancy experience. People should go to vercel.com; they’re going to see this beautiful triangle with a gradient. Our engineers and our agents carefully chiseled this beautiful animation.
I don’t give the agent that. The agent doesn’t care. The agent wants straight signal. What we give the agent—and this is a Vercel Cloud capability—is highly precise information about our business that the agent needs to learn.
So you’re creating your website now for agents?
Correct. We’re helping everybody in the world make this transformation as fast as they can, because if they don’t make that transformation, they’re going to be left out of the most exciting economic opportunity of our times.
Mm-hmm.
More and more of the traffic on the internet will be agents doing things for us.
I’ll give you a very simple example. If you ask Muse to build you a website, it’ll use Vercel. It can go off and say, “I’m going to visit Vercel, get some information about how to build software, and then build something for this user.”
I need to present my best face to that agent and to the many, many other agents that are going to…
And this is true for every business, right? The example that’s super top of mind for people is DoorDash, a customer of Vercel, just shipping an agent where you can ask for a burrito over text message. You could argue it’s a gimmick. Actually, some people will use it, whatever. The key point is that this is an agentic interface.
There are going to be people talking to an agent, and then that agent will go and talk to DoorDash. Funny enough, a lot of the internet was built to block bots, which is kind of nuts. A lot of websites are saying, “Oh, you know what?” They’re using technology that was built for a different era that actually says, “No, no, no, this is a bot. I don’t want this traffic.”
Vercel is taking the opposite posture. We’re saying, “No, no, no. The moment is now to help every business in the world retool themselves for agents. For agents and by agents.”
That’s a really good point. I just put out this interview that I did with John Collison of Stripe at the BrowserBase Navigate conference. We were also talking about this before, but they’re the browser for agents to browse the internet—or the connector for agents to browse the internet.
I was talking to John Collison about this, and John said, “Why do we need bot detection anymore?” Everything’s a bot. What are we detecting?
What you do need, and what we provide, is bot intelligence. It’s so important to know if the bot is having a good experience, which is kind of nuts to think about, right?
The transformation that Vercel made is to go from developer experience—which was our first customer, a developer—to now the agentic developer experience. It’s very important that you understand, “Okay, am I getting ChatGPT on my website?”
Yeah.
Because I want ChatGPT on my website. I want it to get the best possible information. I want it to get contextual information. People are not just distributing static information on the internet; they’re also distributing capabilities. You also want to expose your tools, your products, and your services to the agents.
We partner with Stripe and Shopify to offer agents the ability to transact. Another astonishing fact about the Vercel business is that we see agents now purchasing our services. One of the most popular tools in the Vercel toolset is a capability that helps agents buy stuff from our marketplace. For example, if an agent needs to buy a web browser, it can purchase BrowserBase through the Vercel Marketplace.
I’m enabling that transaction end to end without any human intervention. It’s like the agent is visiting the virtual Vercel shelf and saying, “Oh, I need that.” Then I say, “Okay, it’s this much.”
What’s really important for business owners is that you need to have this intelligence about what’s happening with all of these agents: Grok, ChatGPT, Claude, the coding agents, and even the search engine crawlers.
Because one important thing that people need to know is that models are getting trained all the time, but they have certain cutoffs in training. When ChatGPT first came out and you would ask it, “Hey, ChatGPT, what day is it?” it would say, “Oh, it’s 3 years ago,” or something like that, right?
What we have done as an industry is give the agent the ability to search the internet. It’s still very important for businesses to be really good at putting out their content. In fact, I’ll give you an example. When you put out your podcasts and your shows, you want to make sure that there’s a transcript as well.
When an agent—I’m not even going to say “I,” because I’m not going to be doing it—when I say to my agent, “Hey, what did Molly say about so-and-so?” if you basically fed the internet with that information, you’re going to be there more readily available to the agent. This is true for basically every business online.
That's why we have a newsletter. Subscribe to sorcery.vc.
Yes.
Um, no, that’s a really good point. To talk about that shift, we had SEO, which was really important before. Then we went into the age of chatbots, or that kind of thing, chat logs, and we had GEO. I don’t know what happened with that. Are you saying we need a new thing for agents? What would that be?
It’s not just giving them the information. That’s sort of what AEO is about, which I think is really, really important. It’s also that you have to make sure that your product runs agentically, that the agents can use it.
The fundamental protocol—or programming language, we can call it—when the internet first came around was hypertext. It delivered HTML, which some older people in the audience might recall as MySpace codes. The internet for humans was HTML. The internet of agents is new protocols.
One is called MCP, which basically allows you to say, “Hey, agents, these are all the tools that you can use from my service.” For example, the Vercel MCP says, “Here’s how you deploy software. Here’s how you list your projects.” We’re creating a presentation of Vercel that is very ergonomic for the agent.
There are new protocols also emerging. For example, one is the Machine Payments Protocol, or the Universal Commerce Protocol. These are going to be the ones that you mentioned: web browsers.
When I ask my agent, “Hey, go and purchase the same thing that I purchase every month,” or, “Hey, go and get me this really cool…” I take a photo of this. “Go get me this really cool microphone,” it can go and browse a website. Some companies will be out there saying, “Hey, agent, you can just buy it far more quickly in a more streamlined fashion because I’m speaking your language. I’m speaking your protocol.”
I think what we’re going to see over the next few years is a complete transformation of every product and every service toward this. By the way, you see this even from the largest enterprises. I recently spoke at the Salesforce conference.
Salesforce was built as a product that was awesome for humans. You go and see your list of contacts, your CRM. It’s that vision of a pixel-centric interface for humans.
Nowadays, the way that we work at Vercel is that I can go into my agent, which we deploy on Slack, and say, “Hey, tell me about Molly and how frequently she deploys on Vercel, because I’m going to have a meeting with her and I’m her account executive. Give me a quick report on some of the problems that she’s reported.” I never actually went to Salesforce.
Right.
I just went to the agent, and the agent went to Salesforce. What we’re helping companies figure out is, first, how do you build agents so that you can streamline your operations?
Almost everything we do at Vercel, we do with agents, so I don’t have to learn these 20 different tools. You may think, “How big of a change is that?” What I can tell you is that agents are so much easier to use. It’s such a disruption of how products are built.
It’s such a disruption of the Silicon Valley ethos, even, because you would go to any company in Palo Alto, and they would be designing interfaces for humans. They would all be gathered together in a room, and the whiteboard would say, “Okay, what does G have to see when he wants to prepare for a meeting?” “Well, let’s put Molly’s profile picture up here and a list of 5 support tickets over here.”
It was very fixed in time and space, and it was decided basically by a product design team somewhere in the world. At Vercel, we have an agent that helps prepare our Salesforce and our customer support teams.
The interfaces are more generative. If I’m asking about an individual, it renders a profile of that individual. If I’m asking about a company, it gives me more contextual information about that company.
We’re going more and more into a world that, if you’ve watched “Minority Report,” feels more like the information is just coming to you depending on everything: context, your level of experience. I may not want to give a beginner sales development representative the same kind of sophistication that I’m going to give my finance team when I give them an outlook of our business.
These are all things that agents can solve for that websites, traditional pages, and pixels could not solve for. We’re helping this new generation of software be easy to build and ship to these companies.
There are 2 questions I want to ask off of that. One is, what does the end state of the agent era look like? I had Chenxi from Merge on a couple of months ago.
Yeah.
She’s a huge fan. She loves Marc Benioff. She’s a huge fan of Salesforce, and when she saw them go headless, she was like—
Yeah.
“He sees the future.” He’s seeing beyond what no one’s really going to need to go into their Salesforce. They can just interact with it and use it headless.
We’re seeing this now, as you’re saying, with agents, Shopify, and Stripe. Someone was talking about this on X: the agent is buying directly from the brand, so it’s not going to Amazon.
Yeah.
So, maybe that's helpful for small businesses.
Yeah.
What does this end state look like?
Well, I'll tell you my super-opinionated point of view: I almost take the opposite stance. We were talking about your Bending Spoons thing that went viral. I don't believe in aggregating piles and piles of software and having what you could think of as lowest-common-denominator software.
I believe in a future of personalized software. Vercel makes it possible for you to prompt with your coding agent and generate software that's perfect for you. You can build the podcast's schedule and note-taking app, and an agent that fits just your needs.
But most importantly for enterprises, their ROI on creating tools that are domain-specific is much, much, much greater. In fact, there was this entire category of SaaS that you could call no-code, and I'm on the opposite end of the spectrum. I'm a yes-code company because I'm now saying coding agents basically make the code free.
So why not buy—well, the lethal combination there is that you buy Vercel as the platform, and then you never have to buy software again because you can just generate it to your needs. So where do the Salesforces and SAPs of the world fit in?
Well, software needs data. I can generate the most beautiful meeting prep, report, or any of these awesome things that you can deploy on Vercel—websites, applications, agents—but they need to get their data from somewhere. I think the sources of record—the place where customer history lives, where contacts live, where my meeting notes live, where my support tickets live—are critical data infrastructure.
I'm not going to call it software because you used the word headless. What does headless mean? Headless means that we no longer go to the head ourselves. The agent can go and get the data out of the system and then either create new software, create new visualizations of the data, et cetera.
In a way, the entire world of software is changing, and in a way, I think some of the fundamental pillars will still be there.
This episode is brought to you by Brex, my favorite. You become what you spend on, and I refuse to spend my time on work that shouldn't exist. Expense reports, receipt chasing, and manual closes. The companies building what's next, from Vercel, OpenAI, Anthropic, Granola, and Deepgram all made the same call. They all run on Brex. Brex is the intelligent finance platform that combines cards, expenses, and banking into a single stack with agentic finance built in. AI agents that handle expenses automatically enforce policy before spend happens and close your books in minutes. That's why Sorcery runs on Brex, so I can spend time on building and not busy work. It's time to get Brex AF. Learn more at brex.com/sorcery. That's B-R-E-X.com/S-O-U-R-C-E-R-Y. Bye. From handshake to signature faster, Harvey, the AI platform built for legal work, is trusted by 80% of the AM Law 100 and over 20% of the Fortune 500. Learn more today at harvey.ai. That's harvey.ai. AI needs more than chips. It needs power, land, and infrastructure. Zone develops next-generation data center campuses, partnering with AI companies, site developers, and technology leaders to bring compute online faster and at scale. Zone is building the foundation of the AI frontier. Visit zonefrontier.com to learn more. That's zonefrontier.com to learn more.
As you've seen these different value-aggregation buckets of where infrastructure is going, how have you oriented the business around that to capture that value?
Infrastructure is extremely valuable. It has this compounding value and return because when you enable anybody to ship and make source code free, essentially, that code has to be deployed to some infrastructure.
The other thing that's really, really important in this age is security and creating guardrails for what the agents can do. The infrastructure to support agents is extremely vital. Vercel has created new services purpose-built to secure and contain the agents.
A good example is that now everybody can create software. How do you prevent this software from accidentally going out into the internet if it's supposed to be internal to your organization? We created an enterprise platform where you can plug in Codex, Claude Code, and all of these systems, and then securely deploy.
Some people might remember the term intranet, where you basically had an internet that's just for your company, right? We built that for the agent era.
Another really important primitive is the news that went viral around OpenAI hacking Hugging Face. What happened there is that they had this agent, and the agent was given a goal. In the pursuit of that goal, agents—for better and for worse—will do whatever it takes to fulfill their mission.
They're like a soldier whose frontal lobe has been removed, and they're like, "Let's go, go, go, go, go." People will come to appreciate this capability because when you deploy these enterprise agents, when you build your own agents, you're going to realize that now they can act autonomously for longer.
In fact, if I were to summarize what has changed about the world since ChatGPT to today, it's autonomy. With ChatGPT, we had to be in the loop: "Hey, ChatGPT, tell me this. All right, now tell me that. All right, now tell me this." We put that in a loop, and now we call that the agent, and we're training them to go for longer and longer and longer.
What can happen is that you say, "Hey, agent, go and close this sale," and the agent can take a very open definition of that. They're so resourceful, and they're being optimized to fulfill their mission and to go for a very, very long time without intervention, because that's also a better user experience.
If I say, "Find me Molly's email because I need to send her a happy birthday note," the agent can say, "I'll do whatever it takes to find this email and send her the note." Forty minutes later, it tells me, "Oh, I did it."
Maybe it went to the web, maybe it went to Crunchbase, or maybe it found it in a comment on YouTube where you said, "This is my email." This is actually the stuff that happens with agents. But the user experience is great because I didn't have to do all that. I just told my agent to do it.
The risk is exactly that: what are the guardrails for what's okay for this agent to do in the fulfillment of this mission? How can we prevent other people or external influences from altering what that mission was supposed to be in the first place?
I'll extend that example. I say, "Find Molly's email and send her a birthday note." It goes to YouTube. Sometimes you've probably done this: you've Googled an email, and then you find some shitty website or whatever. You find someone, copy and paste the email address, right?
Now imagine that in that YouTube comment, someone plants a prompt injection. A prompt injection is an instruction that's hidden, that's evil, and meant for my agent. My agent was just scrubbing YouTube, and now it got somewhat persuaded to do something evil. But it's acting on my behalf. It's acting with my tokens, my tools, and my resources.
That risk exists. So the long answer is that Vercel is also building the sandboxes and capabilities so that companies can operate agents, deploy agents, and really trust them. They can get the good experience from them, but really trust them to do the thing they were supposed to do.
Side note: Have you ever seen a sandbox that didn't have sand spill on the outside of it? Why are we calling it this?
Great question.
They're not enclosed environments.
Great question. The word sandbox is being used a lot, and sometimes it's kind of a technical misnomer in that it's not what people are calling a sandbox. This computer has access to the entire freaking world. It's not really a sandbox, right?
When OpenAI hacked Hugging Face, there were elements that were sandbox-like, and there were elements that truly weren't. But I'll tell you, the thing that concerned me about that incident is that the first escape of the agent came from a really good sandbox.
Going back in history, in the Linux community—in the Unix community, actually—one of the earliest words that was used was "jail." FreeBSD was a community that coined, I think, "jail," which, in my opinion, is arguably a better word, right? Because with a sandbox, you're right: you just escape. You just step out.
It spills everywhere.
So FreeBSD created a concept of jails—
Mm.
—to refer to sandboxes, which perhaps would be a term that people would hate, so let's not bring back "jail" necessarily. But to your point, the idea is that you are containing the agent within a controlled environment.
It's more like having the agent be a rat in a cage—it's fully contained, it can't break the glass, and you're watching it spin. In practice, I can tell you that, talking to prospects and talking to people we're helping figure out agents, people have not really figured out agent security yet.
A very, very important part of that is having very secure sandboxes, which is itself a really big challenge because coding agents are also extremely good at finding vulnerabilities in code.
We love coding agents. They're great at writing code. But overall, in order to be good at writing code, you're good at understanding the rules of code. And to understand the rules of code, you also understand how to break the rules of code. And so what's emerged is that these coding agents are really good at hacking, and that means that enterprises right now really need to be prioritizing protecting themselves from the potential AI hackers of the future.
So Vercel takes a stance here: we will not lock you in to any model provider. We will give you the tools so that you can use coding agents, secure your environments, and use every model under the sun, which is very attractive to enterprises because sometimes it emerges that a certain model may be better at cybersecurity than another. It's also economically better because you make all of the model providers compete for your business. It's also better from a quality standpoint. Whoever is the new hot model today might just keep the pole position for 2 weeks.
That is happening.
It's exactly happening. I was going to say, Vercel sponsors Formula One. We sponsor the Mercedes team. Mercedes is such a GOAT. I think they just win all the time. But other than that, the other teams are always flipping: 2, 3, 3, 2, 4, 3, 2. The AI landscape is a lot like that.
One day, you have Anthropic. The next day, an open model comes in. It's 20 times cheaper and is 98% as good or even better in some regards. And so Vercel has emerged as the place where not only people are coming to us to build and ship agents, they're also coming to us to get this optionality, or this best-of-all-worlds situation with AI.
Yeah. Wow. We didn't even talk about this, but the use of agents is exponentially increasing compute.
Yes.
How do you deal with that?
What's going to happen over the next few years is that agents are going to use far more of the internet and far more of the available resources than humans could, just from a fundamental physics standpoint. Think about it: it used to take me an entire afternoon to write a few files of code, and then I would deploy them. That would use a little bit of compute to build the software, and then it would go out onto the internet, and I would have to wait until humans visited and used the application.
Just picture that. You write code by hand, you ship it, we run some compute to build and test it, you put it out into the world, and then other humans come in through your door and use it. That triggers some compute and some networking. Now it's the agents writing code. That's why we added 1.5 billion deployments in 10 months. And now the agents are browsing the internet and using all of these resources, so the demands on infrastructure are going vertical.
You ask what is a good place in the stack to be, and I love being in the infrastructure space because, first, every week you wake up to an astonishing new record of usage. I was just looking at our data on our overall network utilization, and every day is a new record. But also, it's hard to even say when this ends. We have so many problems to solve with AI.
I think we've only scratched the surface of what's possible, from discovering cures to illnesses to the personal agent landscape, just adding more value to people's lives. Education, right? I was teaching math to my oldest kid last night, and I was like, “Ah, this app is kind of janky and not personalized enough.” This will be an agent. In fact, one of our most successful startup AI customers is building an incredible teacher assistant for older kids. I think it's more for teenagers in college, but it's happening, right? Education will also be disrupted with AI.
There are all these industries that have barely been touched by AI, so I can't even fathom the demands on compute, the energy grid, and the network. Obviously, I think there's a major cybersecurity concern that we need to get ahead of. We're investing a lot there. And so, yeah, it's a very exciting moment to build, and it's a very exciting moment to be in infrastructure. Also, I think for humanity itself, we've only scratched the surface of the wins that will come from this.
The cybersecurity concern is probably the next line item under the high-growth headline. Everybody is concerned about it. You don't know what to do with rogue agents. I was just speaking with Fred Wilson of USV. He's on the board of Coinbase, and I was like, “Fred, you've endured the crypto world. I'm sure you've probably clicked on a couple of links you didn't mean to.”
And he was like, “Yeah.” I said, “I've clicked on, like, 10.”
And that was with humans creating the attacking campaigns. It's like the classic scheme that promises you a bounty. Imagine once the agents have a perfect idea of you, your biases, and your context, so they can create a much more targeted attack.
Vercel hosts a lot of the largest crypto companies in the world and fintechs. We have Polymarket and Calcia as customers, and what's top of mind for a lot of these people is not just social-engineering-type attacks, but how do you protect your code?
Yeah.
Because the code that everyone is running on was written by humans. There was this funny moment in the past 3 years where a lot of my peers have been very skeptical of coding agents, and almost all of them have been proven wrong, right? There's very little—it's a very impactful amount, but it's very little that humans can add at this point to the coding agent capability.
There was a time when people were skeptical that agents could produce secure code. A lot of people would troll the agents because they would make obvious security mistakes. But the reality is that humans make obvious security mistakes, and in fact, a lot of the stuff that the world is suffering from today is that we're using coding agents that are now superintelligent. They're reading the human code and they're horrified. They're like, “What the hell were you guys doing over here?” And they're finding stuff that is just so bad.
To get ahead of this problem, we put $1 million toward a sandbox challenge. We told the world, “Vercel Sandbox, we believe, is one of the most secure sandboxes, if not the most secure sandbox, in the world. To strengthen it further, we want you to poke holes.” And we invited the entire world to attack us.
Only 1 person with an army of agents was able to make progress, and we're basically processing his contribution now and sharing it with industry peers. He and I are best buddies now, and we're DMing on X as we speak. For sure, we'll be collaborating with him for a long time to come because, as I mentioned, the contribution is small but impactful. Great example.
Yeah.
This gentleman has an amazing cybersecurity brain. This is not someone's child saying, “Hey, Claude, can you hack the sandbox?” and Claude did it, right? This is very sophisticated cybersecurity skill plus an army of agents.
What we're doing is—because we invited people with this bounty, we now have what is called the white-hat sort of hacker, right? They're like a good Samaritan. We're now sharing this finding with basically everybody in the world because this hacker found a critical vulnerability in Linux, and almost the entire world runs on Linux.
And so what Vercel is doing now is creating the sort of cybersecurity harnesses and incentive systems such that we can find out first and protect our infrastructure. Someone asked me on X, “What's going to happen in the future if all of the human code that we've written to date is insecure? What's going to happen?”
I said, “Look, I think in the short term, there's going to be almost like a war between four good tokens and four bad tokens.” The reality is that North Korea, China, enemies, adversaries, bad people, good people—we all have agents now.
I personally disagree with the whole “let's ban open models because of the cybersecurity risk.” I think that's nonsense because the thing that is making these models good at cybersecurity is just that they're intelligent.
Mm.
You would basically be prohibiting access to intelligence, and we can deploy this intelligence in a defensive way, like we just did, and then help the entire industry. But I think there's going to be a rocky next couple of years because we're going to have to strengthen a lot of these foundations.
Now, a really positive silver lining: I mentioned this week that I worked on a proof system, and a lot of my colleagues are researching this now. We can make agents, who are inexhaustible in their energy and capability, basically create a kind of code that we can prove, almost mathematically, is correct, and humans couldn't do this.
A very small percentage of humans in the audience could do it. I don't know if you saw that OpenAI proved Navier–Stokes, and none of the mathematicians can understand this proof. I opened the GitHub repo, and OpenAI could also not understand it. It's just so much code in this mathematical proof.
I'm not saying that our machine code will necessarily look this way in the short term, but it's gonna move more toward mathematical certainty. By the way, Vercel has been going down this path of strengthening our infrastructure with more verifiable code for the past few years, but the surface is enormous. Most companies have not done this yet. Most of the code is still human-written, and it's not correctly verified, which creates risk for the world.
But again, I think I just want to deliver a message of optimism. I think it's all gonna be good.
Everything's gonna be okay.
I swear, everything's gonna be okay.
We haven't talked about one of the most talked-about subjects yet, and that is the open-versus-closed war.
We're talking about a lot of wars.
We didn't know this was a war podcast, but here we are.
It is.
Well, I do a lot of interviews with defense companies. Now here we are. But okay, of the many tweets that you tweet that go viral, one of them was about how you and Vercel see everything that's going on in the world on the infrastructure side: the difference between open-weight and closed-source usage. So what are you seeing there now?
For context, Vercel operates one of the largest token distribution networks in the world. In the internet that I described earlier—the internet of pages—we had content distribution networks, or CDNs, and we would transmit pages. Now, the really interesting new thing is tokens. It's a unit of intelligence, and you need networks to distribute them, to get your tokens, to power your agents, to send inputs, and to get outputs back.
Developers come to Vercel to get tokens to power their agents, and we're basically one of the largest token suppliers in the world. We aggregate all of the model providers and all of the open models in the world, or most of them, so we get a lot of really good data. We then give it back to the world under Creative Commons, so anyone can use this anonymized volumetric data to understand where the future is headed.
The thing that's astonishing is that, since the beginning of the year, when we started the Vercel AI Gateway, as we call this network, it was mostly OpenAI, Anthropic, and Google. That's where developers and enterprises were getting their AI. In fact, Google Gemini had an amazing run, where most enterprises had found the perfect balance between price and performance with Gemini models. I think they did a great job with that, and they continue to be very strong there.
Then we had the rise of coding models. I think maybe people don't know this, but a lot of the tokens in the world today are being consumed by developers and by what you can think of as software factories—the agents that these developers are commanding. This is one of the biggest sources of token demand in the world.
At the beginning of the year, most of the traffic was going to proprietary models—OpenAI, Gemini, xAI, and so on. Now the situation is almost entirely flipped. Eighty percent of the tokens in our network—and this is a tweet that went viral—are from open models.
There are a few reasons behind this. Open models used to be terrible at coding. I had personally never used them. Even though I'm the biggest open-source person in the world, and I created some of the most popular open-source software in the world, I will continue to fight for open. Our mission in this AI era is what I call: I will enable you, and I will enable the largest companies in the world, to own your open intelligence stack. I want you not to be tied and anchored to a model, and to have control over your data.
But I couldn't live up to that myself because I was like, “Well, the best coding model isn't open. I can't even use this open thing.” Then some of these Chinese labs started putting out stuff that was extremely competitive with the frontier. The Kimi models—I actually used Kimi again yesterday, and I thought, “This model is so good.”
Of course, it's not at the top on every single benchmark or level of execution. There are a lot of things that it cannot do, but enterprises realized that a lot of what they do—maybe 80% of their traffic, and maybe that's what the data is showing—requires a certain level of intelligence. I don't need to hire Albert Einstein to solve every problem, answer every support ticket, spell-check my stuff, and classify my data.
The world is literally opening up to choice and saying, “I'm gonna use the best model for each task.” For example, something that's really common with developers today is that you do need Einstein sometimes—or Einstein-level intelligence—to create a plan. I'm gonna write a document about how I'm gonna approach this problem, and I need as much intelligence as possible. But then I need execution. I need to translate that plan into action, into code, into instructions, into operations, and basically every open model is amazing at that task.
What I think the data is clearly telling us is that the future's gonna be a portfolio of options. Enterprises also love having the negotiation power to say, “Look, I'm not anchoring or verticalizing myself to a specific lab. I'm implementing tools and protocols that make the model more of a module, like an engine that you can swap out.”
If, by tomorrow, Anthropic is not doing so well in a particular dimension, I just change it out and go to the Vercel AI Gateway with 1 line of code. The other thing that's kind of nuts is that you would think replacing models would be this massive engineering effort, but in reality it's 1 line of code.
A lot of what the world is researching right now is whether we can make the model choice automatic. There are a lot of promising directions and examples of places where you don't even have to ask. Much of my career, and much of my success, has been about thinking backward from the user experience first. This world where you have to first choose a model in order to do a task is not very user-friendly.
I would much rather have my colleagues say, “Okay, this is my query. This is my prompt,” and then we can decide where to route it based on context and intelligence needs. Sometimes, for very high-end engineers, they need to choose a specific model because they know it very well and have to choose it. So it's a combination of broadening access to every model under the sun and also letting some people pick.
A clear commonality that I've seen in the conversations I've had—and one in particular with Bending Spoons—is that Luca mentioned that 99% of the time, they're using open models. They're not using frontier models.
So what role do frontier models play, and how does that impose a risk on them, given all this crazy momentum and these hyper-billion-dollar valuations? What is the risk to all that and to everything they've built?
The biggest risk is that it's hard for them to understand what their retention is. The best products in the world—I always give this advice to young founders—focus on the non-cheatable metrics of your business. You know what's cheatable? Top of funnel. Run ads.
There are a lot of startups that are kind of frowned upon. They just do these baity marketing campaigns, record videos, make silly jokes, or whatever. Cool. That's top of funnel. But are you getting a customer for life? Are you doing a great job for them day in and day out so that, by year 10, they're still with you?
I think that's the biggest risk with a lot of the model companies: actually, there's zero loyalty. In fact, we have amazing loyalty to our AI Gateway product because we have the best interests of the customer in mind. If they want the absolute top intelligence from a model, they can get it, and they stay. There's no reason to leave.
What I would worry about if I were them is: how fleeting am I? Am I—
By the way, you were talking about crypto. There are a lot of elements—and I want to be super clear about this to everyone who's listening—of AI that resemble some of the crypto highs—
Yeah.
—of easy come, easy go. There were people who had zero loyalty to a coin. They were just saying, “I love this coin because it's the hot coin for the next 2 months,” and then selling it. Or they were even lying or manufacturing usage. There are all kinds of crazy things going on with AI, including people inventing token volumes.
There are a lot of free promotions. There's a lot of fake ZDR—zero data retention. One of the things that we guarantee in our AI Gateway for enterprises is zero data retention. You come to me because I can guarantee contractually, with all my providers and with US-based inference providers, that no one copies your data or trains on your data.
There are a lot of things going on in the market that I think are very frothy. You hear about token volumes, and you hear about promises of training versus no training.
There’s a lot of weird stuff going on. That reminds me of crypto a little bit—the speculation element, et cetera. Another thing with crypto is that there were people who were hardcore believers, and they’re still doing really well because they’ve been in it for so long, they understand the technology, and they love it. And there are tourists.
What I would ask myself if I were one of these companies is: What percentage of my users are here just because I gave them free subsidies on a plan, or because I happened to have a good day where my model was good, but by next week it’s not good? There’s a lot of volatility there.
Now, I do think that if your model is really intelligent and you can sustain it economically—which is a big if—why wouldn’t you have customers? There’s always going to be a demand for frontier intelligence. But I do think that you can run a business that is primarily open with models, and you’re going to get better prices and more ownership over your intelligence stack, which creates more defensiveness for your own business against the model businesses.
The other thing is, if the model labs were so convinced that the frontier intelligence advantage is enduring, why are they going into random software businesses? So I think some of the—
Like all of them building different applications for all the different verticals.
Yeah, like tomorrow they launch, “Oh, here’s our Slack clone,” or, “Here’s our other thing,” and whatever. There’s some behavior that seems to indicate that they’re starting to diversify from just being a token supplier because they’re finding that they’re—
My wine is from the Champagne region of France, and only that one can be called whatever.
Champagne.
Yeah, exactly. And so they’re seeing, “My token is not that much better than that other token.” And so my goal is: Hey, let’s drive the token price to 0, let’s drive developer happiness to 100, and enterprise value of our customers and ourselves to 100. So we’re in that situation.
But I also would not say, “Okay, these companies are going to disappear tomorrow.” I think they’re going to do well for long periods of time.
Yeah, I have so many different directions I can go with this.
It’s a fun topic.
It’s a very fun topic. What’s going to happen? Are they going to IPO, or is it a money pit?
What I’ll say is, it’s hard to predict. The whole point of IPOs and public companies is that they can forecast accurately. What you deliver to the market is a solid engine of a business. I’m Costco. I went public. I can say my business works this way: I sell this many Costco subscriptions, I’m growing at this rate, and people’s demand for high-quality Kirkland bread and lox and whatever—I can understand it.
With some of these AI companies, it’s much harder to understand. I could understand that on both sides there’s nervousness, where the public-market investors are like, “I was super, super excited about this thing because they had this model, and then I woke up the next day and there’s a free model that’s the same.”
Local models are coming, too. There are models that are going to run on your computer and are going to do amazing things. It’s already happening. I have one running on this laptop. This is a MacBook Air, and it can run a small Qwen model that can do very interesting things.
And so this is, to me, the fun of this game. I got into technology because you have to have an extremely open mind, be super agile, and stay on your feet. What I always tell our customers is that the primary value Vercel gives your business is iteration velocity, not speed. Speed is—you could be just spinning in circles. I’m giving you the engine so that you can innovate really fast.
The worst thing you could do today is stay put, where you say, “I’m not going to implement AI in my business. I’m not going to build agents. I’m not going to try open models because I heard this thing on TV and I’m scared,” or whatever. You have to be super open-minded and agile as a business and, by the way, as an engineer. Our greatest engineers are the ones who did not stick to what they knew. They learned new ways to use agents and AI to make themselves better.
It’s true for everybody. It’s true for the workforce. If you want your job to stay great, relevant, fun, and engaging, you have to actualize. You have to incorporate these new tools, and so I think that’s the name of the game today. It’s a very rapidly moving game.
Another one of my bangers was this: I look around at my peers in Silicon Valley, and they’re— You would think, “Okay, AI has arrived, and we’re all working 2 hours a week.” The opposite is happening. I get text messages at midnight from the CEOs and CTOs of these companies, where they’re saying, “Oh, we’re going to launch this thing. We’re preparing this. Help me out with that.”
And everyone’s working. It’s not even 9 to 5. It’s more like 7 to 7, 7 days a week, from 7:00 a.m. to 7:00 p.m. And, by the way, the reason I tweeted this is that a lot of this is super fun. It’s not that we’re suffering out here. There’s so much good stuff happening, so many new things, and it’s exciting to build and exciting to ship. Maybe sometimes it’s too exciting and people get carried away, but it’s a fascinating time for technology.
I was listening to DHH. He has this presentation at Ruby on Rails. I don’t know if you saw that, and he was like, “I’m in this state of delirium.”
Yes.
I don’t watch videos, so I—
You listen to it?
I do the craziest things because I host a fireside chat every Friday with our company.
Mm-hmm.
And I show them how I watch videos.
How do you watch videos?
I get the transcript on my terminal. I skim it really fast.
And then you ask your agent questions and process it?
Watching a video takes too long. I did several peeks at the video just to get some of the body language, but mostly HSR. I read the videos because otherwise it’s too slow.
So weird.
Yeah, totally. No, it is.
Well, he was saying he’s in a state of—he’s not in AI psychosis. He’s in delirium and—
AI delirium, for sure.
Euphoria because it’s just so exciting. It’s like—
Yeah.
The Internet is being invented as we speak.
Hundred percent. The rules are being rewritten. That’s why I was telling you I was shipping software on Sunday, and part of that is very much in the same wavelength as what DHH is talking about, because DHH, much like me, loved the idea of creating great developer experiences for humans.
I realized several years ago that what got us here was not going to get us there. If there was one big move that I made, it’s that I got the AI psychosis, delirium, whatever you want to call it, mosquito bite many years ago. Not to give myself too much credit, but it takes courage to do that, because you have to sit in front of your company and tell them with a straight face and conviction that the thing they fell in love with is no longer there.
The thing that DHH and I fell in love with was beautiful code, easy to write for humans with our little monkey extremities, right? Primate, evolved extremities. And that’s just not the game anymore.
Mm-hmm.
But what’s awesome is there’s a new game—and, well, the game that has always been here is that you have to make the agents work on the right things. Capital allocation.
I don’t know if you’ve heard of Marc Andreessen, but at some point, he went viral in a self-serving way that was hilarious. He said, “The last job is venture capitalist.” Sadly, I think he’s kind of right. I wouldn’t say venture capitalist; I would say capital allocator.
What you do as an engineer is allocate tokens, and we all have the same wall-clock time. None of us, to my knowledge, can distort time. We all have the same clock running, and we have to invent, create, and do good for the world. Now we have this superpower of tokens. What are we going to put them toward? What are they going to investigate? What languages and tools are they going to use?
That’s what DHH’s presentation was about. He’s like, “Ruby on Rails got me here. I think other things will get me there.” I think right now it might be this programming language called Rust, which was also super controversial when we introduced it a few years ago inside Vercel because it’s a language that is—
They named it beautifully. I picture it in my head as this rusty, not-pliable tool, and the things that are made for a good developer experience were these soft, nice fabrics. Now you have to grab this rusty, hard engine that is always telling you that you’re wrong.
One of the things Rust does is tell you all the time that you’re wrong in a really cryptic way. It’s cryptic because it’s almost like it’s approaching mathematical notation. It tells you, “You violated these 5 lemmas, and you’re wrong and you’re an idiot.” That’s what the compiler tells you. And that’s what DHH’s point was.
In the past, you as a human needed to swallow that pill. It took a specific type of character to swallow that pill. Almost like when you watch Magnus Carlsen lock in for a chess game, so locked into the rules of the game—as a human, you need to lock into the rules of this language. And now it's coding agents that do it, and we basically live in the promised land.
We ascended to: Oh my God, all of the code in the world is going to be secure. This was where my optimism was coming from. All the code is going to be fast. It's going to be secure.
Incidentally, Vercel built a platform where this kind of code is amazing to deploy. We built almost the perfect platform for this moment, which I'm really happy about. But the world has changed, and as a developer, you have to—it's almost like an ego death. You have to forget what got you here. Going back to the Marc Andreessen thing, choosing what to work on is still something that you weigh in on, and choosing with your creativity is very, very important.
How good is AI at real professional work? Mercour's AI Productivity Index, or APEX, tracks it with every major frontier model release. Built with partners like Harvey and Cognition, see where the frontier actually stands at mercour.com/apex. Today's episode is sponsored by VC-X by Fundrise, the public ticker for private tech, allowing investors of all sizes to invest in venture capital. Learn more at getvcx.com. Public is a multi-asset investing platform where you can automate your investing with AI agents. Create your own custom agent with a simple prompt. Agents can monitor markets, execute trading strategies, and manage your cash across stocks, bonds, options, crypto, and more. Wanna buy at market close and sell the next day's open? Just tell your agent. Automate your investing at public.com. Founders scale faster on Deel. Set up payroll for any country in minutes, hire anyone anywhere, get visas handled fast, and get back to building. Visit deel.com/sorcery. That's D-E-E-L.com/sorcery.
I have to ask you, because there are some really quick money grabs in this universe, whether it's neoclouds—
Yeah.
—all these kinds of new providers; the coding agents are spewing off money. How do you stay focused in that when all of that stuff is going on? Or even these—
Quality.
—data companies are making $2 billion.
It's so easy. I'll tell you an example. The sandbox product—to give people a really simple explanation: as a human, I join a company, and I'm given a laptop and do my work. I install Excel and whatever. It's my personal computer to do work. Agents also need computers. They're the sandboxes. We don't see them; they live in the cloud.
What's nice about this world is you can see the computer: Oh, it's Apple; it's beautiful and very valuable. In this new world, you buy the Vercel Sandbox and give it to the agent. Regarding the relative quality of the sandbox product—not to pat ourselves on the back, but I want to give the world this kind of quality—the products and services we're creating for this agentic era, this agentic infrastructure, are the highest quality in the game.
I hear every day from customers who migrate off of, “Oh, yeah, I went to that neocloud, and they gave me a deal on this and whatever.” It's not reliable, it's not secure, and it's not easy to use for agents. There's still a massive quality gap between all of these providers, and I commend everyone in our industry who's not just yapping and launching into every other business for whatever reason, because they're committed to this line of business.
We've been launching sandboxes in the cloud for 11 years, because every single time a Vercel deployment is made, we had almost the same challenges even before agents existed. We had to create a secure cloud computer for the code to be built in, so we came into this market with a very unique perspective. We came with a lot of experience, and we're delivering a great product. I think you're going to find that quality is still the name of the game.
Yes, everyone can show you they can create something really quickly with AI, but they still need to maintain it. They need to put those tokens to work for a lifetime of commitment. They have to make it reliable and secure and stuff. This is why I always remind the team: a lot of the world has changed, and some of the things that make companies win and what makes people win are the same.
Well, Vercel seems to be winning. You guys are at a $9.3 billion valuation. The Information reported you guys at 600 million in annualized revenue. That's growing. You have 60% of AI coding agents using the platform, and I think it was 485,000 customers. You named a few already, and it's a crazy list.
The customer growth of Vercel has been exceptional for the market that we're in. Maybe no other infrastructure company in the public market today has seen this kind of customer acquisition that we've seen. And that was based on this idea of making infrastructure more accessible to the world, and it turns out agents are now making the choices, and they're choosing right.
So I've heard that you're also involved in the hiring process. You have around 800 employees?
Yes.
And you're hiring, so let's talk about what you're looking for in those hires so we can help when people watch this video.
AI leverage is extremely key to the hiring process. First of all, it used to be that if you had a static resume, it was already kind of hard to hire you. You needed to show—borrowing from the crypto world—proof of work. What have you shipped? But now the barrier to shipping has also gotten lower. So I'm not even interested in, “Oh, yeah, I have this little demo,” or, “I vibe-coded this thing.” I'm actually looking for your ability to commandeer swaths of agents and really get leverage back from this new ecosystem.
Maybe you can call this agentic engineering. You can call it building software factories. I'm looking for people who are at that level of execution.
I still really value the communication quality of people—their communication competence. I believe that the future is going to have 2 extremes. There are going to be human-centric interactions that have to be the most humane and highest-quality interactions that you can possibly find. If you go to vercel.com, you're going to find this beautiful human-centric experience. It's for your eyes, and it communicates our values, our craftsmanship, and our design.
The same goes for how you speak to people. I don't want to read your AI-generated tweet slop. I want to hear what you actually think. I really care about people being able to be great human communicators, because that's going to matter a lot.
Especially—we employ fantastic AEs who are selling our agent platform to the Fortune 500s, and they're establishing trust-based human relationships. I really care about how they show up and how they communicate, and their technical understanding, by the way. I expect a much higher level of technical competence. Why? Because with curiosity and the ability to prompt, every problem, every subject becomes tractable to you.
It doesn't mean that you're going to become as competent as a Rust engineer with 20 years of experience—they didn't even exist—or like a C++ engineer with 20 years of experience in C++. But now an AE can be forward-deployed and speak way more eloquently about our technology, its impact, its APIs, and its data flow—all the things that were more in the realm of engineering.
I borrowed a quote from a friend of mine who was CEO of Splunk: the 2 most valuable people at the company are the ones who are close to the code or close to the customer. Code is the engine of everything; it's the thing that powers the products that we sell. The customer is who we live for; it's who we serve.
But now it's almost code at both extremes, because more and more we're forward-deploying our sales force, our sales engineers, and our forward-deployed engineers—meaning I'm helping people migrate over and sell in record time. We did this example with Mercedes. When we started partnering with them, they asked us, “How can you make our F1 team better?”
Being a performance maniac, I said, “Well, we can make your website much faster, because when people go to buy your merch or want to get the latest results from the race or whatever, I want to make you faster and I want to make you agentic.”
That process would have taken quite a bit of time. But the first iteration of the migration of their old stack to the new one actually took 5 minutes. We used an agent to rewrite the code from the old platform to the new one. Five minutes. When I saw the agent's result, I was like, “This used to take months.”
In fact, a lot of what could have been a slowdown to our business in the past was that enterprises were fully bought into the vision of the cloud and acceleration, et cetera, but everyone dreaded migrations.
Yeah. They would hire consultants for this.
That's right.
And it would take a year.
Cloud consultants—
A year.
And Scrum masters.
Yeah.
Have you ever heard of Scrum masters?
Yeah.
And agile coaches.
Mm-hmm.
Anyone would talk about the migration. I can now forward-deploy an engineer and do the migration.
Mm-hmm.
And so that's why it's no longer even just close to the code or close to the customer; it's close to the customer's code and close to the code—or close to the agent and close to the agent, right? And so I expect a much higher level of technical proficiency and curiosity from everybody at the company.
We're still hiring humans very happily, so that's not going away.
Fantastic. Um, so as we close out, one of my partners is Brex, and they're all about performance, spending smarter, moving faster. But I take this, because we're on the topic of performance—
Big Vercel customer, by the way.
Really?
Yes.
Wow. Well, I have to ask this question in a personal way. You run a very large company. You've been doing this for 11 years. You have some of the best customers you could possibly have. You also have 6 kids. Congratulations on 6.
Yeah, we just had our new baby girl last week.
Aw, what's her name?
Lisa.
Aw. How do you think about performance in your personal life? I usually ask this because I think of it in terms of who you surround yourself with or who you're inspired by. Who are those key mentors who keep you aligned and on track?
It could also be some of your closest friends who are CEOs of companies, while you're gathering information and trying to figure out what's actually going on in the world. Who are those people for you?
A lot of them are unknown, in the sense that I remember my first boxing coach, who taught me the basics of fitness, nutrition, and performance that were so simple. I've always loved boxing and martial arts because, as humans, we get in our own way all the time.
This might be controversial to some of your guests or audience, but so much of performance optimization focuses on the last mile, 1% optimizations: buy this supplement, do this thing, follow this longevity protocol. No one has gotten the 90% right—the basics of getting your diet right, eating clean, whole foods, and even walking. Walk and do pull-ups. I do the most basic things in the world, and I try not to deviate from that.
I try to tune out all of the nonsense about supplements and, God forbid, peptides. There are people who have gotten all that right and should totally optimize beyond that, and it's also their preference. To each their own. But my message has always been—and I hope to instill this in my kids—just work on your foundations.
This is also advice for our team and for every software company out there. It's so easy to get distracted by nonsense and take your eye off the foundations. There's a customer, there's a pain, and there's a product that serves them. It has to work great, and most of the work that we have to do is based on what customers and their agents are telling us.
Most of the work is actually pretty simple. I think what's really hard for us as humans is staying focused and keeping it simple. For example, this morning I did light cardio for an hour. I try to do basic things.
For my family, I always try to set the example. The reason I try to eat well, exercise, not drink alcohol, and do all these things is that your kids are always watching, and they become a blueprint of you. We do Brazilian jiu-jitsu together, and I hope to pass on these lessons.
Even with the broader Vercel team, I'll give you an example. We celebrated Oktoberfest, but I told the team, "Oktoberfest is fine, but I haven't had a beer in years at this point." I would love to portray a message of, "Can we make both things fun?" I also don't want to be the preachy guy who's just like, "Never drink," or whatever.
We combined it with activities and fun things, including fitness activities. I don't know if you've ever played that game where you have to hold really heavy jugs. We did a competition to see who could hold the jugs the longest. We have bodyweight fitness equipment at the office, so I try to be consistent in my life.
I blur the lines between life and work. How can you possibly have 6 kids and do this?
Yeah.
It's because I blur the lines. I'm teaching my kids what I know. I'm teaching them engineering. I bring them to the office. If they go through the interview process and the 6 stages of grief that we put people through—yes, we have a high bar—if they meet it, by all means.
I bring them to the office, and I'm living and working all at once.
Well, G, this was such a fun conversation, and I could probably ask you 10 more questions. There was one I didn't ask, which Darren from Tembo sent me, and he wanted to know: Is serverless dead?
No, the opposite. What's happening with compute is that we need so much of it, and what's happening with software is that there is so much of it being deployed. The reason enterprises love Vercel is that there is no way on earth they could deploy as much software as their employees are producing if all of it were running all the time.
The beauty of serverless was that when the software isn't being used, we spin it down. There are so many people building personal software. They use it every once in a while, or they deploy it whenever they need it. So you need servers to come online and go offline intelligently.
We call the next evolution of serverless fluid compute. It's a better generation of technology, but it's also a better name for what's happening. You need compute that flows like liquid. If you get lots of traffic, it expands. If you get no traffic, it compresses.
It's heterogeneous in the sense that it can work for sandboxes, it can work for functions, it can run your website, and it can run your agent. I've never been more excited about the next chapter of serverless, which is fluid compute, and I think it's going to define the next decade. I'm excited to see how that pans out, Tembo.
Awesome. Well, again, I'm not going to try to ask another question.
You can. I can go on and on.
Thank you so much.
Thank you, Molly.
And I'm definitely going to clip out every single time you said "code." You probably hit 1,000.
Awesome.
Hey, it's Molly. If you enjoy our interviews, check out our newsletter, Sorcery.vc, where we deliver a once a week top deals and tech headlines email, and also go deeper on our podcast interviews. Subscribe to Sorcery today. And don't forget to subscribe to the podcast on YouTube, Spotify, Apple, or wherever you listen. Link in description to sign up.