Vercel raised $300M in a September 2025 Series F that valued the company at $9.3B, co-led by Accel and GIC with BlackRock, StepStone, and Khosla Ventures joining the cap table according to Vercel’s own announcement. A separate $300M secondary tender ran alongside it for early investors and employees. Nine months later, the product stack that justified that valuation has shipped: AI SDK 6, AI Gateway, and v0 are live. The category Vercel calls “the AI Cloud” is a deployable product now, not a funding narrative.
What Vercel Actually Bought With $300M
This was not a typical growth round. Secondary reporting from AICurator places Vercel’s ARR above $200M with 82% year-over-year growth and a developer base that doubled to 3.5M+ in twelve months. Those figures come from a single secondary source and may be approximate. But the direction is clear enough: Vercel did not need the primary capital to fund operations.
The primary capital bought time to build the AI middleware layer before competitors recognized it as middleware worth building. The secondary tender is the tell. When a company runs a $300M secondary alongside its primary, it rewards early believers and locks in a shareholder base that bought the AI Cloud thesis at $9.3B. Those shareholders now need that thesis to be correct. Vercel’s job is to make the AI Cloud real enough that the next round prices it as infrastructure, not hosting.
The AI Cloud Bundle Dissected
The bundle has four components worth distinguishing.
AI SDK 6. Shipped December 22, 2025. The core shift: the abstraction layer moved from individual model calls to agent boundaries. AI SDK 6 introduces ToolLoopAgent, human-in-the-loop approval flows, stable MCP (Model Context Protocol) support covering OAuth, resources, prompts, and elicitation, and structured outputs combined with tool calling. This is no longer a convenience wrapper over provider APIs. It is a runtime for agentic workflows that happens to deploy on Vercel’s infrastructure.
According to a community technical guide, the SDK’s agent abstraction makes MCP a first-class integration primitive rather than an optional plugin. Once a team builds tool policies, approval gates, and agent boundaries against AI SDK 6’s APIs, migrating away from Vercel means rewriting that agent layer, not just swapping a deploy target.
AI Gateway. A unified API endpoint routing requests to hundreds of models with automatic fallback and spend monitoring. Vercel’s documentation states zero markup on tokens across all plans. The economics are straightforward: Vercel does not need to profit on token resale when the gateway drives compute, bandwidth, and SDK adoption on Vercel-hosted deployments.
The gateway’s value is not the routing. Any competent engineering team can wire together OpenAI, Anthropic, and Google APIs in an afternoon. The value is the operational surface: fallback logic, spend monitoring, rate-limit handling, and model-switching without code changes. That surface accumulates institutional knowledge inside Vercel’s dashboard, and that knowledge does not export.
Next.js as distribution. Next.js was downloaded over 500M times in the twelve months preceding the Series F post, exceeding cumulative downloads from 2016 through 2024. Grok, Claude, and Cursor all use Next.js for their frontends. Vercel controls the framework that ships with React, and the AI Cloud stack integrates first with Next.js. The distribution advantage is structural.
v0. Vercel’s generative UI tool, now extending to a mobile app, pushes the stack toward non-developer users. The brief does not include independent usage figures for v0, so its contribution to the bundle’s defensibility remains an open question.
Why Zero-Markup Tokens Are the Cheapest Lock-In
The “no markup on tokens” claim is worth reading carefully. It is not generosity. It is a pricing strategy that commoditizes the thing Vercel does not control (model inference pricing, set by OpenAI, Anthropic, and Google) and concentrates value in the thing it does control (the gateway, the SDK, the deployment pipeline, the operational tooling).
Any competitor can match zero markup. It costs nothing when you were not marking up to begin with. The lock-in is not in the token price. The lock-in is in the SDK abstractions, the gateway’s accumulated operational configuration, and the deployment coupling between Next.js and Vercel’s runtime. A team that has built three agents against AI SDK 6’s tool policy system, routed through AI Gateway’s fallback chains, monitored through Vercel’s dashboard, and deployed on Vercel’s infrastructure has four layers of switching cost. The token price is the one layer that is portable.
The Competitive Ratchet
No public statements from Netlify, Cloudflare Pages, or AWS Amplify respond directly to Vercel’s AI Cloud positioning. What follows is structural analysis, not attributed competitive reactions.
Each of Vercel’s adjacent competitors faces a choice: build a comparable AI middleware stack, or accept being categorized as commodity static-hosting infrastructure.
Netlify has the closest audience overlap with Vercel’s frontend developer base. Its feature set emphasizes deploy workflows and edge functions. As of this writing, Netlify does not ship a first-party model gateway, an agent SDK, or structured MCP integration. To match Vercel’s bundle, Netlify would need to build or acquire all three. The engineering investment required to reach parity with AI SDK 6’s agent abstractions is not trivial, and Netlify does not control a framework with Next.js-level distribution to amortize that investment against.
Cloudflare Pages has a global network footprint that Vercel cannot match. Cloudflare’s AI strategy runs through Workers AI, its own inference runtime. The overlap with Vercel’s audience is partial: Cloudflare targets infrastructure engineers and platform teams, while Vercel targets product-facing frontend developers. The risk for Cloudflare is not that Vercel out-infrastructures it. The risk is that Vercel redefines the category such that “frontend hosting” defaults to “AI-aware frontend hosting,” making Cloudflare’s infra advantage secondary to its AI-stack gap.
AWS Amplify has the deepest infrastructure moat of the three and the weakest developer-experience narrative. Amplify competes on AWS integration breadth, not on curation or opinion. The AI Cloud thesis is fundamentally about curation: a curated model set, a curated SDK, a curated deployment path. AWS’s strength in breadth is a liability when the buyer wants an opinionated stack.
The structural argument: Vercel’s bundle raises the minimum viable feature set for any vendor competing for the same developer cohort. Before the AI Cloud, that feature set was Git integration, preview deploys, edge functions, and a free tier. After the AI Cloud, it includes a model gateway, an agent runtime, MCP integration, and AI-specific observability. The cost of staying in the category went up. That is the ratchet.
The Risk Side: Vendor Concentration and the Single-Platform Bet
The AI Cloud bundle is coherent. It is also a single-vendor bet on a stack that is less than a year old in its current form.
AI SDK 6 shipped in December 2025. The agent abstractions, MCP integration, and tool policy system are new enough that production adoption data is limited. Teams building against AI SDK 6’s agent boundaries are adopting an abstraction that could shift significantly in v7. Vercel has no track record of maintaining stable agent-runtime APIs across major versions because there has not been a major version before this one to maintain. That is not a criticism. It is a fact about version 1 of anything.
The MCP dependency is worth separating from the general version risk. MCP is an open protocol, but AI SDK 6’s implementation of MCP is Vercel-specific. A team that builds tool integrations against AI SDK 6’s MCP layer is not portable to a generic MCP client without reimplementation. The protocol is open. The runtime is not.
Next.js distribution is a strength that carries a governance tension. Next.js is open source and maintained by Vercel, but the framework’s trajectory is set by Vercel’s commercial priorities. Features that differentiate Vercel’s hosting (server actions, partial prerendering, the App Router) have progressively deepened the coupling between Next.js and Vercel’s runtime. The AI Cloud accelerates that coupling. The framework’s independence from its host platform has been eroding for three years.
The $9.3B valuation is a bet that Vercel can own the AI middleware layer for the React frontend ecosystem. The product stack now exists to justify that bet. The moat is real but narrow: it depends on Next.js remaining the default React framework, on AI SDK maintaining its early lead over alternatives, and on Vercel executing on a runtime abstraction that is months old. The Series F bought time. The shipping gave the narrative substance. Whether the substance holds under production load and competitive pressure is the question the next eighteen months will answer.
Frequently Asked Questions
Can AI Gateway be used without AI SDK 6 or Vercel hosting?
AI Gateway is available on all Vercel plans and can route requests from applications hosted elsewhere, but the operational data it collects (spend patterns, fallback frequency, model usage distribution) accumulates inside Vercel’s dashboard. Teams using the gateway independently of AI SDK 6 gain the routing and fallback layer but lose the agent-boundary integration that makes the SDK’s tool policies and approval gates portable between models.
How does Cloudflare Workers AI differ from AI Gateway architecturally?
Workers AI runs model inference on Cloudflare’s own edge infrastructure, while AI Gateway brokers requests to third-party providers (OpenAI, Anthropic, Google, and others) without running inference itself. The distinction matters for cost modeling: Cloudflare pays for GPU capacity and prices inference accordingly, while Vercel’s zero-markup model works because it captures value at the SDK and hosting layers, not at the inference layer.
What adoption benchmarks exist for AI SDK 6’s agent abstractions?
AI SDK as a whole crossed 3M weekly downloads before the Series F announcement, but that figure includes all major versions and usage patterns, not just the v6 agent runtime. The ToolLoopAgent, MCP integration, and human-in-the-loop approval flows shipped in December 2025, giving them less than six months of production exposure. No public benchmarks exist for these abstractions under sustained load, which is typical for a first-generation runtime but relevant for teams with SLA commitments.
What triggers would make the zero-markup model unsustainable?
If major model providers compress their own inference pricing (Google has already set a pattern with aggressive Gemini pricing), the gap between provider cost and end-user expectation narrows, reducing the cross-subsidy room that makes zero markup affordable for Vercel. The model also depends on gateway-driven compute and bandwidth revenue growing faster than the operational cost of maintaining fallback chains, spend monitoring, and rate-limit handling across hundreds of models. A significant shift toward on-device inference or self-hosted open-weight models could reduce gateway traffic volume without reducing its maintenance cost.
What does MCP elicitation add that standard tool-calling does not?
Elicitation allows an agent to pause mid-execution and request structured information from a user before continuing, distinct from tool-calling (invoking an external capability) or human-in-the-loop approval (a human approving or denying a proposed action). In practice, elicitation covers the gray area between full autonomy and full approval, a workflow stage most competing agent runtimes do not handle as a first-class primitive. AI SDK 6 treats elicitation as a native MCP capability, but it is also the least documented and least tested of the MCP integration features in the current release.