The AI Cloud

@rauchg|Oct 17, 2025 (1d ago)3,444 views

Vercel was born out of my frustration in 2015 that while the cloud enabled this seemingly infinite array of possibilities (and compute), the pixels on the web weren't getting significantly better. And neither was the experience of crafting them.

The cloud promised to remove the burden of maintaining physical data centers and hardware, but ultimately we inherited much of that burden in digital form. DevOps, K8s, VPCs, CI/CD, IAM, CDNs, IaC...

The time you should have spent on the pixels... was spent on the configs. The jobs you could have devoted to product and customers, were devoted to the setup of clusters and machines.

Over the past 10 years, we've been fortunate to play a role in reversing this story. We led with developer experience, starting with frameworks and tools that, upon deployment, generated the infrastructure. React pioneered the concept that the DOM is the output of your state and data.

We set out to make Vercel the React of the Cloud, where the hyperscaler primitives are outputs not to be directly manipulated, or jQueryed. And, in the process, we made a popular React framework too.

The results: Next.js's seen 500M downloads over the past 12 months, which is more than in the entire 2016 to 2024 period. Vercel is crossing 1 billion application deployments. We now power two of the top 20 highest traffic web properties on the planet, like OpenAI and The Weather Company, while also serving millions of applications of all sizes, including 1-person companies making millions in ARR.

But to win the next decade, Vercel must heed the advice of its hyperscaler predecessor and partner: it's always Day One.

AI is eating the Web, and with it, the Web Services we're used to. There are three massive trends that are shaking up the web, the cloud, and software as a whole:

• Pages to Agents
• Problems to Solutions
• Closed to Open

I believe this is a transcendental time for developers and users. We can either give up the web, proxy all our data and intent through one closed mega-app, or reinvent it together in the open.

#From pages to agents

The cloud was conceived as a democratization of Amazon.com. Developers demanded faster mechanisms to provision infrastructure than clunky data centers, and wanted to dream up new experiences of their own.

Amazon.com is a set of pages. Very fast, dynamic, personalized pages. Emphasis on fast... something that, despite the rise of AWS, remained elusive to most and today underpins a lot of the success of the Vercel business. Customers like Under Armour, PetSmart, DoorDash, Nintendo, and Porsche partner with us because we help them work backwards from the customer experience, delivering fast pages and great product experiences, while we automate the infrastructure away.

One of the key unlocks of this first wave of the web was that page speed correlated closely with conversion. Google rewards page speed with higher rankings, which creates a flywheel of more business, more conversion, more organic traffic.

Pages got us here, but agents will get us there.

When you Google something and you get 10 blue links, wouldn't it be great if "someone" opened 10 tabs, did a ton of reading, researched the facts, compared them, visualized them in a way that's relatable to you, considering your knowledge, memories, preferences?

Well, that's just the definition of an agent, and something Google is already doing via AI Overviews, and in the process re-shuffling the traffic makeup of the internet. An agent is "just" a piece of software that's able to perform a bunch of steps on our behalf, intelligently, and deliver an outcome.

Unlike pages, agents are not usually synchronous. They're never static. They can think for seconds, minutes, hours. Soon, days and weeks. Unlike pages, they don't follow a rigid set of data fetching and rendering instructions. They can write their own code and create tools on the fly to achieve their goals.

This is making us rethink the services that make up the cloud.

Agents get closer to why we invented software to begin with. We wanted to scale expertise, make users more productive, make people's lives better. I believe every organization and individual should have the power to create and deploy their own. Instead of clinging onto the old web, we should pick up its lessons. Everyone has the freedom to own a domain and deploy to it.

#From problems to solutions

To wrestle with the complexity of building and deploying software, we invented an enormous ecosystem of tools, frameworks, and services.

Frameworks have so far served two purposes. One, to abstract away the complexity of the underlying platform, codifying the best practices of expert developers and organizations that have gone before us.

On Vercel, frameworks also serve an additional, critical purpose. They automate infrastructure through what we call Framework-defined Infrastructure (FdI). CI/CD, compute clusters, CDNs, caching systems, and more are seamlessly provisioned. Like a "cloud compiler" that turns your code into optimal infrastructure.

The lifecycle of software is roughly "build and run". Build is the phase where you write code, test it, and prepare it for production. FdI has given us autonomy here.

Run is the phase where your code is executed, serving users, processing data, and performing tasks. This is also where things can go wrong, slow down, or altogether break.

The latter phase, run, is where we believe AI will allow us to attain the next level of cloud autonomy. An AI Cloud will be able to monitor, optimize, secure, and repair itself, without human intervention.

Earlier this year, we introduced our Firewall Agent, capable of responding to threats and providing high-quality investigations and mitigation strategies to human operators. Our Code Review Agent now analyzes every pull request by executing it in a secure sandbox and emulating the work of a dedicated, expert colleague who looks at your code with a healthy amount of skepticism. A (thorough and extensive) vibe check if you will.

We believe that an agentic cloud will repair and optimize, not merely inform. Every application at scale today incurs gigantic costs associated with collecting telemetry and observability data, surfacing it in dashboards... hoping that humans will open them, prioritize issues, and fix them by hand.

Fundamentally, we believe an AI Cloud shouldn't give you problem after problem (alerts, 5xx errors, latency spikes, traffic anomalies...). It should give you solutions: pull requests, recommendations, and automated actions.

#From closed to open

The promise of rethinking all software with AI is thrilling. I believe all applications and services will involve AI in some way, shape, or form. I believe the cloud itself will become agentic.

As we reimagine the web and web services, we must ensure they remain open and decentralized. The beauty of the web is that anyone can own a domain, deploy to it, and link to anyone else. There's no "app review process", no "platform fee," and every protocol and format is open and documented.

The first wave of AI has been dominated by closed models, proprietary SDKs, protocols, and centralized interfaces.

Instead of a single agentic interface, we should have a web of agents. Instead of a single model SDK, we should embrace model choice. Our contribution here is the AI SDK, which not only abstracts over providers, it makes the developer experience even more delightful than the vendor-specific SDKs.

Developers seem to agree, with the AI SDK package (ai) on npm now being the second-largest AI package by downloads, second only to openai:

`openai``ai``@anthropic-ai/sdk``@google/genai`

AI SDK is now the #2 largest SDK for AI in JS/TS, but, crucially, one that's model and provider agnostic

The early success of MCP as the protocol connecting agents is a promising foundation of this new web, which won't be made up solely of pages. And more examples of open AI protocols are emerging, such as for commerce (Agentic Commerce Protocol) and coding (Agent Client Protocol).

#Conclusion

The internet of tomorrow will look very different. Users expect (and deserve!) more intelligent experiences, while increasingly relying on agents to solve their problems.

This is true for developers, who offload problem-solving to agents. For framework and devtools builders, this means our job is no longer to empower and delight developers, but also their agents.

It's also true for the cloud. The "hello world" experience of yesterday was a page, today it's an agent. The interface was clicking through UIs, today it's prompting, or just doing work behind the scenes. Surfacing insights and solutions, not endless UI or problems.

For all of us, developers or not, this is an opportunity to reinvent and reimagine the web and continue building it on open foundations.