Guillermo Rauch (Vercel) at the Cerebral Valley AI Summit

You recently crossed the 10-year mark, I believe, with Vercel. That's right. And it seems like maybe you're gearing up for the next 10. Any reflections on kind of the first 10?

Yeah, broadly, I think the first 10 years of Vercel are reflective of a large transformation that's happening in the industry. The last 10 years were about pixels broadly, user interfaces that need to load really fast, that optimize for better conversion, 10 blue links on Google searches and optimizing for SEO. The next 10 years will be the 10 years of tokens and agents. Going from pages to agents is a soundbite. And it'll be marked by new kinds of software engineering approaches, perhaps moving from developer experience to the agentic developer experience. You know, companies like Vercel have grown out of creating the best possible DX, the best possible APIs and succinct code snippets that human developers would adopt. The next 10 will be marked by appealing to agents. There is the Y Combinator meme of make something people want. I think we'll be transitioning into make something agents want. And yeah, in the next 10 years will be around agents, which, perhaps don't have the user interfaces that we've been accustomed to till now, uh, less so much on rendering a bunch of pixels and UIs and clicks and more on autonomy and, uh, intelligence at your fingertips.

Yeah. So you wrote a great blog post, highly recommend checking it out, um, about the AI cloud. Um, kind of your, yeah, you, Vercel has been known historically as, uh, providing this kind of front-end-focused cloud. What is different about the AI Cloud? Why are, why are you kind of defining this as the next phase of Vercel?

Yeah, if you look at the history of the cloud with, you know, AWS, Amazon Web Services, things were quite different back then. The thing that was hard to do was creating highly performant web experiences. Amazon.com showed us that for each 100 milliseconds of page speedup, they would get a 1% improvement in conversion rates. The kinds of software that people will be creating in the future are less motivated by rendering some pixels on the screen really fast. We're noticing that the compute that people are devoting to creating these kinds of applications are spending a lot of time thinking. They're streaming interfaces to clients. They're hyper-personalized and contextual. So my argument is that the Hello World application of the cloud will no longer be a website or web application, at least in the form that we've been creating them till now. Another large transition is that even though the cloud has been so successful at sort of virtualizing the data center and automating the provisioning of compute and resources, it's also been a huge source of problems. The average experience of any of your clouds is that you get paged, Recently I went viral on X because my Internet of Things mattress woke me up in the middle of the night because US-East-1 went down.

What was it like, by the way? It didn't really like just—

It was 3:00 AM and my mattress is warm and then I opened the app and it tells me AWS is down. And so of course there is ways and Vercel does a lot of them by default to create no single points of failure, automatic recovery, and things like that. But the average experience of the cloud is still that you get paged, that you have to pay for really expensive dashboards of error rates, that you have to configure and wire up alarms and page schedules for your team and Slack integrations. And I think we'll be going into a world of full autonomy of the cloud. I call it the self-driving car. Moment of the cloud. And I call it moving from problems to solutions because the artifact that the user of the cloud will want is not a dashboard that shows them their error rate. It's gonna be either a pull request with a proposal of a fix. Think of it as a self-repairing cloud, a proposal of a fix of something that you got wrong. It might be a recommendation, right? So sometimes you have to enact policy. You have to say, well, we're under attack, therefore we should be blocking this whole range of IPs coming out of Russia because they're not our customers for this particular product. And then the other one will be investigations. When something goes wrong and the human does need to get paged in, there should be plenty of context. The LLM should have cooked a lot on what went wrong so that you don't have to do so much, you know, from scratch investigation.

How far are we from that vision? Like, are we, are you already seeing that internally at Vercel where you're getting paged with context of the problem?

Absolutely. So we have rolled out a number of these products to our customers. This started out as internal agents. This is something that I advocate to for every company that today, or perhaps you were born before AI, but even if you've been born in the age of AI, you should be building your own agents, right? Phase 1 of AI was you buy ChatGPT Enterprise, you buy v0, things like this, and you distribute a bunch of seats across your organization. But I really think everyone should be exercising the muscle of building their own agents. For example, for SRE or Site Reliability Engineering, we had this Slack-based incident workflow, whereas any, where any, anything would go wrong, we have to go in and start chatting, you know, from scratch and try to understand the root cause. And we instead built an agent that basically analyzes and scans all of our sources of truth, logs, telemetry, et cetera, and can provide a really good investigation. And now we're giving this agent to our customers.

How far does this go in this agentic web future where pages, pages or interfaces might have their own computer, code themselves, or fix themselves? Are you kind of preparing for that world as well?

Absolutely. I think, well, every interface will change pretty dramatically, right? I think we're moving into a world that's a lot more event-driven or workflow-driven. It used to be that you had to go to a website or application to take action, whereas in reality, action can be taken already on your behalf. Getting Paged is a really good example because it's not like you have to open a website and start clicking around some signal happens, some anomaly gets detected. Vercel recently rolled out our automatic anomaly detection product with AI investigations for this reason. The human should not be the one triggering the action in most cases. I would even argue that this goes beyond the prompt box because every interface becoming a prompt box makes the assumption, perhaps wrong assumption, that some human will open up a website and then type something in, whereas even that might be unnecessary for us. A whole slew of applications. But overall, we're moving into this world of workflows and events that then make a set of agents respond. And then I'll, I'll say the page will likely become the output rather than the input. Whenever the agent needs your intervention or produces an artifact, that'll likely take the form of a page or some kind of UI. So it's not that UIs go entirely away.

Well, so are you picturing a— painting a vision where the current web, the open web of today kind of goes away? Most humans aren't browsing pages. Like most humans are interacting through agents or chat interfaces. And then there's emergent UI.

Is that the way to think about it? We call this emergent UI generative UI because UI can be produced last mile or streamed in from agents. And v0 being a great example of this because anytime you go in and you type a prompt, We generate UI on demand, but imagine little v0s being embedded within all of your agents to produce just-in-time UI to your customers.

So right now, is v0 mostly used in the category of vibe coding or disposable apps, I think you've called them in the past, of maybe internal tools that you're not using, or is it graduating beyond that?

Yeah, I think we've seen a really significant evolution in the space of vibe coding because What we've heard from customers is that, wait, I wanna put this applications into production. Yeah. They're not vibe coded, they're touching really sensitive data. We just announced the partnership between v0 and Snowflake where a lot of our investment went into making the boundaries between the systems extremely secure and reliable. Furthermore, customers who wanna go to production, they don't wanna eject from this amazing user experience of being able to prompt their way to success. Success. And so a lot of effort has gone into creating, quote unquote, teams of agents. The first agent that designs the architecture, the agent that sources design inspiration, the agent that code— writes the code, and the agent that does QA. So I think what we've been seeing is more of a push towards production-grade vibe coding, or agentic engineering, if you will. Very strong enterprise adoption. I think we've been seeing You know, there was a hypothesis of maybe Vibe Coding will completely erode or eliminate entire categories of SaaS applications. I think the one that seems to be most at risk is anything involving data visualization. That's why Snowflake came to us. You know, this sort of static components that you drag and drop and put into a kind of formulation for your business intelligence dashboard, you know, they pale in comparison to what Vibe Coding can do because it's more contextual. Because these internal applications are very much purpose-built. So that entire category of software seems to be, you know, ripe for disruption as well.

Can you also talk about this kind of proliferation of agent orchestration frameworks? Like we see a lot of these different visions of like drag and drop and sort of your AI SDK, which is open source framework. Like what are you envisioning, you know, if you're encouraging everyone here to go build agents, like what actually does this look like?

Yeah, so I went viral for my mattress heating up in the middle of the night recently and also for making the case that agents are workflows, which is kind of a tabs versus spaces kind of discussion online right now, or what is the definition of workflow? But clearly, you know, what a lot of people are interested in is that they already have well-understood business processes. So I'll give you some examples from Vercel. Whenever someone submits a contact sales form, we created a triaging agent. That surfaces the most important leads or reroutes, for example, some customers to the support channels. We used to have a huge problem. I mean, nice problem to have, but we have way more submissions to our contact sales form than, than we could handle. So we create an agent for that. So that's kind of like the traditional workflow problem where you get some input and then you have some business rules, you have some agents as well in the mix. Now, of course, you have classification tasks et cetera. And so these are things that you intuit can be well solved by visual agent builders. And so we've seen the resurfacing of a lot of this tooling. Now the problem is whenever these become a crit— they go into the critical path of your business, right? Like our contact sales form, we can't miss even one submission. And so I think the argument goes that you prototype really quickly with visual building tools,, but then when they become a part of your critical business, you have to treat them as any other kind of software. You have to give them the engineering treatment. And this means having CI/CD, having really good observability, having, you know, SLAs on performance and things like this. So what we've done with the AI SDK is kind of like if you think about pages to agents, Next.js has been massively successful because we gave you a very easy to use framework. That can scale from one page to every page on the internet, uh, to running OpenAI and weather.com. And so we, we wanna do the same with agents. So the AI SDK gives you the same thing, ease of use, but nearly infinite scale.

So in a world of pages, um, Next.js maybe, or Vercel generally, you're measuring, you know, the, the biggest websites that are using your technology.

That's right.

What does this look like? What is a measurement of success for the AI SDK?

Great. Uh, so we have three company metrics that we look at very closely. One is requests under management. This is, of course, Vercel is at the front door of every request on the internet. Many of these have historically been pixel interfaces. The other one is compute. So compute cycles are extremely important, right? You can't just have a GPU. You also need the CPU to be doing all of the token processing, UI rendering, you know, business logic transformation, et cetera. And the other one is really important to your question is tokens under management. Okay. How many tokens are flowing through our network?

What is the strategic value to you about being that router for tokens?

Yeah, so for context, we, we announced our AI Gateway is now one of the largest routers of AI tokens worldwide. It's growing exponentially fast. It's slated to 10x our total token volume even just within the next few weeks. And so what's really exciting about being the gateway of tokens, and again, using the metaphor of pages to agents, One of the core pieces of infrastructure that helped the web become what it is today is the CDN. The CDN gave us what? Reliability when pages would go down. The CDN is intricately, intrinsically connected to the WAF, or the Web Application Firewall. It protects you when you get attacked. It shields you from bad traffic. And it can provide failover, meaning server— this happened when us-east-1 goes down. The Vercel CDN rerouted all of traffic to US-East-2, Ohio. And we overwhelmed Ohio 'cause they didn't have enough capacity, but that's for another time. But basically AI gateways can do the same thing. Models can go down, models can go get rate limited, but actually we argue they're more interesting than CDNs because we've seen that developers are very willing to switch between model providers. This was not the case when you would pick S3 and EC2. Maybe it would switch once every 5 years. v0 has switched models 6 times since its inception. So gateways can be the great normalizers and enable this transition from closed AI to open AI in the fullness of time.

Jumping off that, are you worried about, you know, apps and webs or sort of, you know, interfaces to the web moving within to these closed ecosystems like ChatGPT or Claude, like, You know, what is the risk to the open world?

Yeah, I'm not worried because of the emergence of, well, a couple things. One is I believe all of us can create your own agents really easily today, obviously using Vercel AI Cloud or things like that. So there's gonna be millions or billions of agents. There's not gonna be one agent to rule them all. I'm pretty confident of that. The other thing is we've seen the success of vertical and domain-specific agents. One of our customers, GC AI, just, announced a big fundraise today. They're killing it being this vertical solution for law. And so the world has demanded domain-specific agents. I always start with the customer and not just philosophically, you know, married into a protocol or technology stack. But even on the protocol and technology stack, we've seen the rise of MCP. When ChatGPT apps came out, I was delighted to see that they're building on the MCP protocol, which means you can now deploy to Vercel and embed a UI into ChatGPT. So the agent-to-agent interactions are seeing a lot of traction in the real world in more of an open way than you might, might have expected. Way more open than I would have expected. I think I, I believed closed-walled gardens were going to develop and hasn't seen— hasn't materialized. MCP has been doing really well. There's Agentic Commerce protocol as well so that you can syndicate your products into agents that buy or sell things on the internet. There's x402 spearheaded by Coinbase to enable microtransactions between agents. They prophesied sort of alignment between crypto and AI. So there's a lot of really good stuff happening on the open protocol side of things. And then also the success of our own AI SDK, which is again normalizing the model layer. It's now the second largest AI module in the JavaScript and TypeScript ecosystem. And just for context, only second to OpenAI. And probably gonna catch up to it sometime soon. So the world has sort of coalesced around open solutions, and we've also seen the tremendous progress of Chinese models. v0 runs a fine-tune of a Chinese model. It's becoming easier to fine-tune and extend open-weight models. And so I'm very, very bullish on the open web continuing to thrive. It'll take a different shape. We all need to adapt, But also on the model side, it's looking very promising. Excellent.

Thanks so much for joining us today.

Thank you.