Please welcome Tobias Lütke and Guillermo Rauch, CEO of Vercel. [music] Thank you so much for coming. How is [music] Stop death going? [applause and cheering]
It's really sweet to have you all here. I think this has been a really, really wonderful event. Let's thank the organizers here for the wonderful system here. Team has absolutely cooked and thank you so much for joining. It's great to be here. We had on some of those notes earlier than with Harley already, but again, I said this before: if the web would be invented today and someone would propose it to Apple and Google to try to get it launched in the app store, it would just be rejected, right? Like it's like this is not a thing. So we luckily all live in a counterfactual world where the web is open and it's available for everyone. I launched my business building applications within the web. You are building incredibly inspired infrastructure for helping other people launch their web businesses. And I think there's like two companies, our two companies, that are at the definitely in a set of companies that someone would name in the top few companies fighting for open web, making sure it's there and making sure it's good. So I was looking forward to this conversation because I think, first of all, this is under-documented and under-discussed, but also there's no better place to discuss where things are going and what's now possible.
Thank you. I've always felt like our companies are two sides of the same coin, which is helping people publish to the web. And what I've always admired about Shopify is that it's so ambitious in that you've helped people not just publish, but actually build their businesses, and grow and scale their businesses. We've sat on the infrastructure side building the frameworks, the tools, the vast networks that help the internet be really fast. And I'm sitting here at potentially the most exciting time of my life because we've always been about making it easy to create on the web. And with Next.js and React and things like that, I think we made it pretty damn easy. Like I was obsessed with can we make building and deploying one command. I started Vercel with that idea: just one CLI command, Next dev, you dev on the thing, and then Vercel, which was called 'now' at the time, now it's... I had this like German detour, but one command, you deploy on the web. And then can we make that as accessible as possible to a huge number of people? And I think we sometimes forget because the world has changed so much in the past two years that the stats for how many web developers Vercel could address when I would make investor presentations or whatever was like 10-15 million, something like that. 10 to 15 million, and it was a number of JavaScript-savvy developers because we were like, okay, can we bring it down to what is the lingua franca of the internet? It's not Ruby, right? It's JavaScript. And it's the thing that's common, is on every device, is on every web browser. It reaches the highest number of people possible. But that was like 10 to 15 million. Now fast forward to today and I keep having this epiphany every day with coding agents, with this, whether you call it vibe coding or agentic engineering. We have now the ability to create at the moment of inspiration, at the moment that we have an idea, and we're opening up the market and the total addressable audience to hundreds of millions or billions of people that can now build with natural language. And obviously we can get into the details of what does it really mean to build and how deep do you go with those tools, but I think it's the most bullish I've felt about the web since I started Vercel. Yeah, because again, the web started as an academic document exchange, right? And at some point you can find stories about the invention of image tags and so on, this is all documented, which is kind of fun, especially even real-time in the newsgroups. And so it's been in this slow conversion of, out of all the duads that are required to create the modern web application, it's been the most roundabout possible way. It's been surprising. But along the way, because it's not independent, obviously the web browsers, there's multiple ones, they had to figure out how to agree, even how to pass the markup files. But also, you know, regulations came in, unfortunately. What started out extremely simple with just some formatting tags, to make it actually work in a world of different device sizes, responsive, accessibility laws, and so on. It's just like, there's this beautiful thing that we all need to be able to create documents, stories, now artifacts for businesses, websites, and online stores, and everything. But it was kind of buried against this backdrop. React came out which helped a lot. It at least wrestled everything into a functional direction which was clearly the right direction to go. When did you encounter React? What's your moment of discovery? I was working at a company that was fully monolithically PHP, which has always had a special place in my heart.
Because PHP was the OG file system API. You create a folder, put a bunch of files in it.
It's funny how to think back of this, now we like yeah. You know what's really good? Give file systems. They like file system with folder structure, and you know what, opening tags in the middle and just saying what you want. It's kicking it old school, but it's really on point.
I was just reading the docs on the liquid improvements. My first idea for Next.js was actually to create a liquid style syntax for React. I actually haven't shared as much. Breaking news, the earliest prototype, because I loved this idea of seamless dynamism that could be sprinkled over markup in documents. And what was missing at the time in the JavaScript community was that you needed to become almost like a web server expert to take advantage of JavaScript on the server side. The whole fast CGI and worker processes and daemons, and then if you go further, you start peeling that on and you get into cluster Kubernetes, etc. So it's like, wait a second, PHP only required for me as a young kid in Argentina to create a folder called webroot and throw a bunch of PHP files. Imagine if you're an agent for the agents in the audience. You're salivating because they love when this coding CLI, they just love operating on the file system. And it's so intuitive. It's also intuitive for humans. And so the idea for Next.js was like, can I bring that ease of use and friendliness to the React world, which React was also going down a little bit that path of like, oh my god, I need to take a PhD in order to create my first React website. It was kind of nuts.
Yeah. Yeah. And like, obviously we can dine out on the PHP story. It is really really funny because honestly, every time I see that deploying Shopify takes like 40 minutes in CI, and then there's an entire amount of work and waste heat emitted into the universe by just getting a new version up, is pretty remarkable. And like, back in the 90s, here's what we did.
I this is a topic for me because I'm so obsessed with deployment. And you know one of the things that has made Vercel such a good counterpart to agents is that we started the company with just a CLI. I mentioned one command.
Is incredibly inspired now. I mean, you couldn't have nailed the landing better on that. CLI tools are now the absolute meta. Yeah. And to your point about how the web emerged so much, another thing that gets me really excited is that we're going back to the fundamentals of Unix.
Um, the idea of processes, small programs that plug in together, open protocols, open standards, file systems. I mean even the terminal, right? A lot of us are using agents on the terminal, which is an invention like the VT100 terminal goes back to like 1970s. And so there's a return to fundamentals that is extremely healthy for the ecosystem. Obviously with Shopify pioneering or pushing a lot of the new standards for agentic commerce and things like that, MCP for example is a great success. I know it's somewhat controversial because CLI sometimes are seen as being in competition with that. But anything that helps the broad ecosystem click together and democratizes and decentralizes power, I think is extremely good for this new era that we're going into.
So the area we're going into here is now the browser is incredibly mature. The APIs it already has are extraordinarily underexposed partly due to this complexity. Like, how is what you're doing relating to again the accessibility laws and all these kinds of things? Various governments especially in Europe feel like they're web designers and tell us exactly what the checkout button needs to say. So we're managing. But there be dragons in some of these places. However, at the same time, this sort of calcified everything to a world where templates reign supreme, and we made very very configurable templates, and Shopify kind of bent this too. The thing we read which we had on stage yesterday is that we're going back to basics on like, hey, templates are good and easy to reason about for humans, and for agents, and can be modified. And the reason why this is important is because I think with our additions that we launched, we try to, they are meant to be a love letter to the web. We pay a lot of attention to make these engaging and fun, and also just delightful around new technologies. But then the next set of technologies that are coming right now that are becoming available, like HTML canvas and these kinds of things, this doesn't even, all these words don't compose to something terribly exciting, but when you actually figure out what HTML and canvas can do, it's mind-blowing to a degree that no one's prepared for. So we're just sort of getting these gifts from the platform. So this is an amazing time because you have an agent, your stack makes it tremendously easy for the agents to reason about. It's like etched into the depths of the neurons. There's so much React code out there and there's so much RLHF to make everything from their book.
So you guys are adding Tailwind support to liquid templates, that's really exciting. We similarly have a standardized on Tailwind for all new create Next.js apps. You touched on the power of APIs as well. One of the exciting things about even before agents, the magic of Next.js was that, and this was super controversial because a lot of the people that were coming from Ruby on Rails were saying,
They were saying, why is Next not like Ruby on Rails? Why doesn't it have an opinion like Active Record? Why doesn't it have an opinion? Are you Vercel guys out of your mind? Do you miss half of the requirements? And it was hard to answer that because we were living in a different world of composability. I mean, headless or composable commerce was one of the things, actually one of the big growth engines of Vercel before AI happened. One of our greatest growth years, much like you guys, was COVID where the business 5xed year-over-year, and a lot of people discovered that they can move really fast if Next.js and React act as the front-end layer and then you compose a bunch of back-end services. And that's why we were saying, not that you couldn't build a full stack application with Next.js, you could at the time. Now we've made it a lot easier. But hey, if you plug in the Shopify API, now you've gone from trying to reinvent and boil the ocean to the best of both worlds, literally the best backend and the best front end together. And so fast forward to today, and now you have agents love to operate on this level of abstraction. I'm just so excited about what Next plus Vercel plus Shopify can do in terms of furthering the web and growing it. The other thing, since you touched on HTML, if you've been watching closely what's happening with what do agents love to output when you ask them a question? What do they do? We've been very fortunate that a lot of them love to create applications with Next.js, but they also love to write HTML. Simple HTML artifacts are, you could argue that the web browser is the world's most successful sandbox. People are excited about sandboxes as well today because agents get so much smarter if you give them a computer. But an agent plus a web browser is basically unstoppable. Now, it's hard to predict what the web is going to look like because things like HTML and canvas, WebGL, WebAssembly, and agents are becoming more and more competent. I do think the web will play a central role because of this popularity and widespread adoption. But it's evolving extremely fast.
Right now. It's like so much is happening. I'm sure you have that experience too, but what are some notable like you wake up, you see a thing, you go 'holy' moments that you remember in the last? It's basic for me, I don't know. I would love to hear as well, probably on a weekly basis. The one that was just selling Vanessa in the green room, cybersecurity massive topic. I mean news breaking left and right about agents' capabilities with cybersecurity. We run a lot of benchmarks to understand what are our agents really good at and how do they compare to others. So the most recent mind-blowing moment was seeing what models like Kimi and Sonnet can do in terms of finding vulnerabilities in code. When you launched the auto research example for optimizing the rendering performance of liquid or the virtual machine or whatever, those tasks that are so freaking hard for a human being. I just yesterday I reverse engineered a binary because I was asking one of our developers why is the binary so big?
That's a good rite of passage for public company CEOs as well. Yeah. And you could go to the source and you could look at all the linker compiler options, you could look at the options that it was configured with, because agents just understand everything. It can look at the binary, start dissecting it, it understands ELF, and it gave me like three big ideas for how this binary could get smaller. So those are the moments where I'm like 'holy crap'. The fact that me as a normie human, I find reading binary impossible.
I think that's a good thing. But an agent just has it for breakfast. And sometimes I think we underutilize agents every day. I'm asking myself how much am I carrying the inertia of the world where I grew up in and I lived in over the past few decades. Because one subtle one is they can ingest so much information, right? Like one million content, three Harry Potter books, whatever people say, but you can just paste vast amounts of content. And I think sometimes we still treat it a little, we've taken the metaphor of it's a digital coworker a little bit too hard. Because a digital coworker, I can't go into a Slack message and be like 'I paste three Harry Potters, give me an answer', right? So I'm just astonished by the level of capability that we now have at our disposal. And the other thing that we were talking backstage, imagine a world where your Next.js and Shopify applications and websites self-improve. One way in which I'm convinced we carry the inertia of the world has been is that we think that everything we must be prompting. We must be the ones that sit down, bust open our terminals. 'Hey, can you make the button a little bigger?' No. The agent can say, 'Hey, I made the button bigger. I configured a feature flag on Vercel. I ran two permutations of the test. I'm seeing checkouts per minute go up. What do you think? I think we should ship it.' And proactively makes improvements to your applications and infrastructure. I think that world is possible today. There are versions that we've already shipped. For example, our AI anomaly detection system. It can tell that your application is receiving anomalous traffic. A lot of you all are creating storefronts for businesses that are based in certain geos. There's no reason that overnight Staples Canada should have a massive spike of traffic from Jamaica. And so today, we rely on people noticing these things or hand-wired alerts. Vercel can tell you proactively that something anomalous is happening with an AI investigation. So the thing that I keep thinking about all day every day is how do we make these things more proactive so that it frees up our time to do the interesting parts of our business.
Yeah. Yeah, this is such a good way to think about it. Just because your example especially on observing the log files, it's clearly possible. Everyone on Shopify knows how to go add the data, but there's too much of it to watch. It's extremely toilsome. And this is really just automate the toil so that you can work on the things that are actually making a difference. We want the merchants on Shopify to make the next version of their products with more of their strengths and few of their weaknesses, so that they can find the best possible product market fit. And anything that can sort of array itself into the background, agents, partners, many of you here, and Shopify, and your infrastructure layers, intelligent agents, all these layers, it can make it so much more fun to build businesses. Because sometimes, I've talked with entrepreneurs who, when you ask them 'hey, what would 16-year-old you think about that?' you get back sometimes something which is actually surprisingly hedged. It often takes the form of '16-year-old myself would be unbelievably proud of what we have accomplished, but kind of dismayed for how I spend my day.' And I think we can hone this to a much better point and just focus on the stuff that made you fall in love with solving that problem initially. I do think there's something fundamental about push versus pull. We all kind of go and get information from the world. And I think logs is such a great example because if you think about when people think about AI and agents, the co-founder of Notion had a good zit today saying a lot of the world is still thinking about AI as search.
Like outside of the San Francisco bubble, when you go and talk to executives, they think of AI as some really cool search engine. And so we have to actually also realize that our own utilization of AI, wherever you sit, can probably be a lot greater. And I think this pull versus push thing is quite fundamental because we're all used to this idea that we go to systems and we pull information out of them. But some of the best experiences in the world are literally when you get something good sent to you, right? Like I think this is why people love buying stuff with my wife loves Shop Pay. We love getting things that are going to improve our lives. Sometimes we buy good things, sometimes we buy bad things and we waste money, whatever. But agents can give you this gifts.
I gave my Hermes a budget and the Shopify API and it buys gifts for me. It's awesome. I get random stuff, really dialed at this point. I always like it. It's really cool. I discover new merchants with it, too, which is even cooler. So, I'm not super representative here, because I probably don't have to necessarily look at the checkout screen so carefully. But it still is a delightful experience. So I think the amount of trillions of dollars of enterprise value from turning some things which are currently pulled into push.
Yeah. Yeah. I think also about it in the sense of if you think about the very definition of intelligence, a lot of people have asked what is the definition of intelligence. Is intelligence an IQ test? I think the one that I've been most persuaded by is that intelligence is all about prediction.
Exactly. So if you're Toby, young Toby, and you say 'ah, this web thing is going to be really big' and 'this commerce thing is also going to be really big', I'm going to place a bet here. It's like you're basically predicting the future. In the future, more and more of the commerce of the internet is going to shift to the web and to the internet, and more and more merchants are going to come online. And I think we now have to also have a humbling moment for ourselves that models like Fable are extraordinary and superhuman at certain kinds of prediction. One such prediction could be what is the next product that I should buy? What is the next action I should take on a website? In fact, one of the reasons that we created Next.js was the frustration over a static web. The web only scaled until recently, until Vercel and others have helped solve this problem because we cached the crap out of it, right? So going back to the beginning, we had amazing things like PHP. We were querying a database or a data source and then we were printing dynamic content out, and then we put it on the front page of Reddit or Hacker News or Digg, and then it shot the bed immediately. So we said okay, what should we do? Cache control. And so now we took this amazing gem that was by definition generative, right? We all get excited about generative AI, but the web as it went more dynamic, it basically became generative, but then we cached it. One of the reasons that both Shopify and Vercel got excited about React server components is that we do believe caching is important. You should be caching certain component roots within the page because it has to perform. But the orchestration layer has to be generative. It has to be predictive. It has to be intelligent. And if you think about the commerce at large today, it's not quite there yet. I'm curious your perspective. I don't think it is.
