Vercel’s pitch leans hard on the open-source lineage of Next.js and the AI SDK. The framework will run anywhere. The problem is that “the framework runs anywhere” and “the platform you are paying for runs anywhere” are two different claims, and Vercel’s marketing treats them as one. The billable value sits in the parts that do not travel with the code.
What Vercel actually calls open source
Vercel’s own marketing leans on the open-source framing of Next.js. The landing page introduces Next.js as “the open source React framework Vercel built together with Google and Facebook”, burnishing a lineage alongside two of React’s parent companies. The New Project page advertises an AI Gateway Demo described as “an open-source AI chatbot built with Next.js, the AI SDK, and the AI Gateway”, attaching the open-source label to a stack that includes the AI SDK.
None of that is false. Next.js and the AI SDK are real open-source projects, and a developer can clone them, read them, and run them without ever touching Vercel. The lineage claim is doing reputational work, though. “Built together with Google and Facebook” attaches Vercel’s commercial product to the credibility of two companies whose frameworks long predate Vercel’s hosting business. The open-source framing is accurate at the component level and silent at the platform level, which is precisely where the revenue lives.
The asymmetry is the point. Every sentence Vercel publishes about Next.js being open and portable is true, and every one of those sentences is about a thing Vercel does not charge for. The open-source story describes the leaf. The invoice prices the tree.
Framework portability is not platform portability
Code portability and platform portability are separate problems, and the open-source pitch benefits from keeping them blurred. Next.js will run on a self-hosted Node process, on another platform-as-a-service, inside a container, or as static output for the right app shape. That is code portability, and it is real.
Platform portability asks a different question. It asks whether the deploy-preview pipeline, the edge network, the build system, the managed cache invalidation, and the newer proprietary primitives reproduce when you move. An exit-cost estimate that prices only the code migration is wrong by the size of everything behind it. The conflation is not an accident in the marketing; it lets a sales conversation resolve the platform question by answering the framework question. A buyer who hears “Next.js is open source and portable” and concludes “therefore I am not locked in” has answered the cheap question and skipped the expensive one.
This distinction is the whole article. Vercel can truthfully say Next.js is open source and portable, because that sentence is about a framework. The sentence a buyer needs to evaluate is about a platform, and Vercel does not make that one nearly as loudly.
Where the billable value actually lives
The capabilities Vercel charges for are listed on its own homepage, and almost none of them ship with the framework. The homepage bundles the platform’s value into a set of managed primitives: Durable Orchestration, Sandboxed Environments, AI Model Gateway, Fluid Compute, Web Application Firewall, Tenant Isolation, Custom SSL Certificates, and Preview URLs. The landing page adds the edge-delivery footprint: 90 cities, 24 billion-plus requests served, 10 petabytes of data, and a 99.99% uptime guarantee, alongside zero-configuration deployment and managed cache invalidation.
The deploy-preview workflow is the clearest example of the split. Every Git branch or pull request gets a live, production-like URL through Vercel’s built-in continuous delivery, integrated with GitHub, GitLab, and Bitbucket. That workflow is a platform-side pipeline feature, not a Next.js feature. The capability exists only because Vercel’s build system, edge network, and DNS wiring produce it together, and none of those three layers is in the npm package.
The “Agentic Infrastructure” framing widens the gap
Vercel’s current homepage leads with “Agentic Infrastructure,” and it is an expansion of the proprietary surface, not a retreat from it. The page pitches agents that “reason, execute code in isolation, run for hours, and recover from failure,” which presumes a runtime, an orchestration layer, and sandboxing that only Vercel sells as a managed product. The reference customers are cited as running on this surface specifically: Notion on millions of daily agent conversations, Zapier on 100 million-plus monthly website visits, and Mintlify on documentation for 20,000-plus companies.
There is also a coding-agent plugin. Vercel ships npx plugins add vercel/vercel-plugin for Claude, Cursor, and Codex, tying agent authoring tooling to Vercel-specific deployment assumptions. That is a second-order lock-in move worth naming separately. The open-source story is about the runtime. This is about the authoring loop, the place where a developer’s coding agent learns to assume Vercel is the deployment target. Once an agent has been conditioned to emit Vercel-shaped config and Vercel-shaped deploy steps, the cost of retraining it for a different host is real, and it is not measured in dollars.
The pattern matters more than any single product. Each new primitive widens the gap the open-source pitch papers over. The “Agentic Infrastructure” framing is live on Vercel’s homepage now.
What self-hosting Next.js does and does not reproduce
Self-hosting Next.js reproduces the application code and the framework’s rendering model. It does not reproduce the platform that makes Vercel’s pricing defensible. Per Vercel’s own product pages, what travels with the code is the component tree, the routing, the rendering modes, and the build output. What does not travel is the deploy-preview pipeline, the 90-city edge network, the managed cache invalidation, Fluid Compute, and the AI Model Gateway, because those are documented as platform capabilities rather than framework features.
Where the cached sources go quiet is on behavioral parity. Whether the platform behaviors that have self-hosted equivalents produce identical runtime semantics off-platform is not something Vercel’s marketing pages establish. The specific practitioner complaints that circulate in React-community discussion, ISR cache divergence between a Vercel deployment and a self-hosted Node process, edge middleware parity gaps, and vendor-tied observability, are not confirmed by the sources in this brief. They are plausible and worth testing before any migration. They are not established facts, and an article that asserted them as such would be overreading thin, vendor-only sourcing.
How to price the exit before you commit
The useful question is not “can I leave?” A buyer can almost always leave; Next.js runs elsewhere by design. The useful question is what leaving costs, in engineering time and in regressed behavior, and which of those costs someone is currently treating as zero.
A practical exit-cost audit, scoped strictly to the platform surface Vercel advertises:
- Inventory the primitives in your request path. Fluid Compute, the AI Model Gateway, the Web Application Firewall, and edge middleware each represent a replacement project if you move, not a configuration change. Count them before counting savings.
- Price the deploy-preview replacement. If your review workflow depends on per-branch preview URLs, leaving means standing up your own continuous delivery with preview environments. That is genuine infrastructure work, and it is usually underestimated.
- Map your cache and edge assumptions. Managed cache invalidation and the 90-city edge network are platform-side. A self-hosted or alternate-host deployment needs its own CDN strategy and its own invalidation mechanism, and the behavior under a real traffic spike may not match what the team is used to.
- Audit your agent surface early. The “Agentic Infrastructure” primitives and the coding-agent plugin are the newest and the stickiest layer. The earlier a team binds agent authoring and agent runtime to Vercel-specific assumptions, the larger the eventual exit, and this surface is still being built out.
- Separate “runs elsewhere” from “behaves the same elsewhere.” Run the application on a second host under realistic load before crediting the portability claim. The open-source pitch answers the first question. Only a test answers the second.
Vercel’s open-source framing is not a lie. Next.js and the AI SDK are open source and portable. The framing is incomplete in a direction that happens to serve the seller. The framework is portable. The platform is not, and the platform is what the invoice describes. A buyer who treats those as the same sentence will underprice the exit by exactly the width of that conflation.
Frequently Asked Questions
Does the lock-in argument apply to a static-export Next.js app?
A Next.js app configured for static export has the smallest exit cost, because it produces plain HTML, CSS, and JavaScript that any CDN or static host can serve. The lock-in grows with each Vercel-specific runtime feature the app uses, including edge middleware, on-demand ISR revalidation, Fluid Compute, and the AI Model Gateway. An app using none of those can move in days; one using all of them faces a multi-month replatform.
Do Turborepo and the AI SDK being open source lower the exit cost?
Turborepo is MIT-licensed and the AI SDK is open source, so neither creates exit friction on its own, and both run on any Node host. The friction sits one layer up, where the AI SDK is built to pair with the AI Model Gateway and the managed gateway is the proprietary part. The SDK travels with the code; the gateway behind it does not.
Which Vercel primitive is hardest to replace for a team building AI features?
The AI Model Gateway is the stickiest primitive for AI-heavy apps, because it sits in the request path between application code and the model providers, handling routing, rate limits, and fallbacks. Removing it means rewriting the provider-call layer, not just re-hosting the app. Deploy previews and edge delivery have direct replacements; a gateway woven into application logic does not.
What would force a rethink of Vercel’s open-source pitch?
The pitch weakens if an open-source project closes the parity gap on Vercel’s managed primitives, the way OpenNext already lets teams run Next.js on AWS with closer behavioral parity. Vercel’s mid-2026 pivot to Agentic Infrastructure, built around Fluid Compute and the AI Model Gateway, reads as a response to that pressure. The proprietary layer keeps moving so the open-source story never catches up.
What do the named reference customers reveal about Vercel’s target buyer?
Vercel cites Notion for millions of daily agent conversations, Zapier for over 100 million monthly visits, and Mintlify for documentation across 20,000 companies. Those three point at teams running agent workloads and high-traffic production apps, not static site publishers. The buyer most exposed to the lock-in is the one whose architecture already assumes Fluid Compute and the AI Model Gateway.
