Cloudflare’s $1,100 Next.js API Rebuild Tests Vercel

A Cloudflare engineer reportedly recreated 94% of the Next.js API in one week for about $1,100 in Claude tokens. Vercel built its deployment business around Next.js. API coverage does not prove the copy could support a production service.

Vercel helped make publishing a web application easy. Coding agents are making the application itself easier to produce. As creation becomes abundant, AI app deployment platforms have a harder job to sell: governing which applications receive secure runtime capacity, identity, model access, observability and a durable route to paid use. The first deploy now introduces a longer operational relationship.

Vercel’s first advantage was a faster deploy

In 2020, coverage described Vercel’s $21 million Series A as financing a front-end developer suite built around the open-source Next.js framework. Netlify was raising capital for web-development automation. Cloudflare introduced Pages as another easy route from a project to a website. Developers still had to build much of what they published; reducing the work between finished code and a live page mattered.

That starting point explains why the Cloudflare rebuild matters, and what it does not establish. Reproducing framework interfaces weakens any assumption that familiarity with one framework alone protects its surrounding business. It says little about migrating an existing customer, matching every behavior, or supporting a service through failures. Agents have also become daily tools for highly paid professionals. Agents can make code cheaper to write without making it cheaper to operate.

Vercel appeared in one article in the earlier comparison period and 15 in the later one. With just one earlier article, that comparison cannot establish a change in customers or revenue mix. It does show why the 2020 description of a front-end suite no longer captures every question the company faces.

Generated apps still need someone to secure them

OpenAI added native sandboxing and a testing harness to its Agents SDK for long-horizon tasks. Those additions address a different problem from writing the first version of an agent: deciding what it can do, containing it while it acts, and checking its behavior before release. Amazon and OpenAI have also outlined a planned stateful runtime on AWS so agents can carry context across ongoing work. Persistent state makes permissions and recovery harder, not optional.

Researchers examined more than 5,000 web apps built with AI coding tools and reported that many had little or no authentication; about 40% exposed sensitive data. The finding does not show that hand-written apps are safe. It does put a price on treating “deployed” as a synonym for “ready for users.” Someone still has to assign credentials, limit access, monitor failures and respond when those controls fail.

Vercel has encountered that boundary in its own operations. It disclosed compromised customer accounts and data associated with an incident discovered in early April. The disclosure does not attribute the breach to an AI-generated application. It shows why account security belongs in an operator’s responsibilities, whatever produced the code. Vercel later joined IBM, Nvidia and others in Artificial Analysis’s Cyber Index Alliance to evaluate how agents find and fix vulnerabilities. Testing an agent’s security work does not transfer deployment accountability to it.

OpenAI brings the customer’s allowance across the boundary

In late September, OpenAI said ChatGPT Plus and Pro subscribers could use their existing plan allowances with 16 launch partners, including Vercel and Devin, through Sign in with ChatGPT. Cognition said OpenAI-model use in Devin Cloud, Desktop and CLI would draw down from the user’s ChatGPT quota. A subscriber need not buy a separate allotment of that model usage merely to begin using a participating product.

OpenAI keeps the account and the allowance against which eligible model use is counted, even as a partner removes a purchase step. The announcement does not establish that a subscriber’s allowance pays for hosting, every model, or every service a partner supplies. Nor does it disclose how much production demand the integration creates. For a deployment platform, knowing whose identity authorizes a workload is now distinct from owning the entitlement that pays for its inference.

OpenAI, Amazon, Cursor, Microsoft and Vercel also sit on the steering committee for Agent Plugins, an open format for packaging skills and MCP server configurations across compatible clients. Developers may be able to carry a plugin elsewhere, though client-specific extensions may still limit portability. The format alone does not settle who holds the customer. The related workflow-control question is which system can authorize an agent’s actions after a portable tool has been installed.

Model choice makes token share an incomplete measure

Open models accounted for 56% of Vercel tokens in August. The figure measures tokens, not customers, applications, revenue or margin. Vercel did not disclose a comparable prior-period share or customer population in the cited reporting, so its August figure cannot establish a new platform-wide baseline. It does establish that OpenAI is not the only model supplier relevant to workloads counted in that measure.

Other denominators produce other answers. An analysis of more than 100 trillion tokens on OpenRouter found steadily growing open-weight use, but a routing service’s traffic is not a census of enterprise applications. AT&T said open models power about 25% of its AI use today, across an average of 45 billion tokens a day; its expectation of a much higher share remains an expectation. Vercel’s 56% and AT&T’s 25% describe different populations. Dividing one by the other would manufacture precision, not a market share.

Model choice also has quality constraints. One analysis found that recent open-weight models lagged frontier closed models’ cyber capabilities by four to seven months. Ramp has launched a model router it says it used internally for three years; Runway has introduced one for generative media. Ramp’s internal operating history is evidence of a use case. Ramp and Runway have not disclosed comparable external-customer volumes or production cost-and-quality results in the cited coverage. An operator must decide which tasks can tolerate substitution before it can claim savings from it.

Capacity commitments outlast experiments

Vercel said its run-rate GAAP revenue reached $340 million at the end of February, up 86% year over year. That is substantial reported commercial demand, not evidence that every AI-assisted project is a disposable prototype. Render raised $100 million at a $1.5 billion valuation and said more than 4.5 million developers use its tools. Render’s developer count, unlike Vercel’s revenue figure, does not say how many users pay. The two companies’ reported measures cannot be placed on one adoption scale.

There is reason to resist extrapolating every new deploy into top-line growth. One assessment found companies using agents primarily to improve efficiency and cut costs, rather than to create new revenue. Meanwhile, OpenAI’s Guaranteed Capacity offers discounted compute against one- to three-year spending commitments. OpenAI’s offer is not evidence that Vercel has signed such a contract. It illustrates the timing mismatch facing anyone who commits to capacity while customer applications are still being tested. Big Tech’s AI data-center plans were estimated to need another 44 gigawatts by 2028; that is a planned requirement, not capacity already delivered.

For Vercel, the test is how many applications keep users, pay their running costs and justify the capacity reserved for them. Ninety-four percent measured the interface one engineer could reproduce in a week; it did not measure the service anyone was prepared to keep running.