AI Cloud

From pages to agents

Guillermo Rauch's 2025 essay argues that AI is eating not just the web's pages but the web services underneath them, forcing a recut of every cloud primitive around agents rather than pages, while keeping the medium open.¹ Rauch frames Vercel's original move as making it "the React of the Cloud": frameworks whose deployment generates the infrastructure, so hyperscaler primitives become outputs a developer never hand manipulates.

The first shift is from pages to agents. The web's first wave rewarded fast pages: page speed drove conversion, which drove ranking, which drove traffic. Agents are the next unit, software that performs many steps on a person's behalf, asynchronously, thinking for seconds to weeks, writing its own code and tools as it goes rather than staying static. Each traditional cloud primitive gets an agent era counterpart in Vercel's own product map: static or semi-static UI becomes full dynamic, generative UI with streaming; the framework for pages (React and Next.js) becomes the framework for agents (the AI SDK); the CDN of pixels becomes a CDN of tokens; compute for human written code becomes compute for agent written code; foreground, CPU bound compute becomes background, long running compute.

From problems to solutions

The second shift is from problems to solutions. Framework defined infrastructure already automated the build phase, described as "a cloud compiler that turns your code into optimal infrastructure." AI extends that autonomy into the run phase: a cloud that monitors, optimizes, secures, and repairs itself, handing a developer solutions such as pull requests, recommendations, and automated actions, rather than problems such as alerts, server errors, and latency spikes. Vercel's Firewall Agent and Code Review Agent are the first instances of this pattern.

From closed to open

The third shift is from closed to open. The first AI wave leaned on closed models, proprietary SDKs, and centralized interfaces. Rauch argues for a web of agents over a single agentic interface, and for model choice over a single vendor SDK. Vercel's own contribution is the AI SDK, described as the second most downloaded AI package on npm behind OpenAI's, paired with support for open protocols including the Model Context Protocol, the Agentic Commerce Protocol, and the Agent Client Protocol.

The scale cited as evidence: 500 million Next.js downloads in a trailing year, more than 2016 through 2024 combined; Vercel crossing one billion application deployments; and the platform powering two top twenty traffic properties.

Why it matters

The essay continues a thesis Rauch stated in 2020 when describing Vercel's original bet: own the hard, undifferentiated layer, in that case the deploy pipeline, so builders can just describe intent, and keep that layer open so the web does not collapse into one closed mega app that proxies all data and intent.² The "solutions not problems" claim reframes the entire observability industry, in which telemetry feeds dashboards that a human hopes to notice, as a temporary state before self repairing infrastructure takes over.

Tensions

The mapping is also a vendor's map of its own product line: the five AI Cloud cells named in the essay are Vercel products. How much of the framing is category definition and how much is category capture is an open question. The claim that an AI Cloud repairs itself without human intervention also runs against the harder truth that prioritizing what deserves attention at production scale is still described as beyond current models, and a self maintaining system that cannot be fully audited is itself a risk surface. Open protocols such as the Model Context Protocol are early, and the vision of a web of agents depends on them winning out over a closed, single app path that has not yet been settled.