Vercel deep dive — the most underrated infrastructure company of the AI era

If you only think of Vercel as "a tool for deploying frontend projects," you're probably seeing about 10% of the picture. Vercel today is valued at $9.3B with $340M GAAP annualised revenue, growing 84% YoY. That's not the loudest number in 2026 AI, but it might be the most improbable — because the starting point wasn't AI. It was "deployment."

My read: Vercel is one of the most underrated infrastructure-shape companies of this era. They spent ten years making themselves the water and electricity of the frontend world, then used that infrastructure to catch the new traffic when the AI wave arrived.

A founder's ten-year arc

From Lanús to San Francisco

Guillermo Rauch: a ten-year road from Buenos Aires to a $9.3B valuation.

Guillermo Rauch was born in 1990 in Lanús, Argentina. His father was an engineer. A Windows 95 machine came home, he taught himself to code. By 11 he was on Linux. By 13 he was listening to Richard Stallman talks and getting deep into open source. At 16 he became a MooTools core developer, leading MooTools Forge — the package manager. MooTools eventually lost to jQuery, but the experience gave him a deep understanding of how open-source communities work and how they fail.

He worked in Switzerland at 17. At 18, without finishing school, he emigrated to San Francisco on an O-1 extraordinary ability visa. To prove ability, he wrote Smashing Node.js — which became one of the most important early Node.js introductions.

LearnBoost — the Node.js military academy (2009–2013)

In SF, Rauch co-founded LearnBoost (2009) with Rafael Corrales and Thianh Lu — a digital gradebook for teachers. Raised ~$5M from CRV, Bessemer, RRE. The product didn't take off. The byproducts shaped the entire Node.js ecosystem.

Why? LearnBoost was one of the earliest companies to go all-in on Node.js. The ecosystem was so young that most of the basic infrastructure didn't exist, so the team had to build it. Rauch's own line: "Sometimes we'd use very early software, find it wasn't good enough, then write and open-source our own." Socket.IO and Mongoose came out of this.

He later distilled the pattern — startup byproducts sometimes become the main product. The rule replayed perfectly at Vercel — Next.js was originally a companion tool to the platform, and ended up the company's biggest asset.

Worth pausing on another LearnBoost figure: TJ Holowaychuk.

TJ Holowaychuk — the Node.js ecosystem's one-man power plant

TJ is a Canadian developer who worked at LearnBoost remotely. His output is so absurd that people sometimes wonder if he's real — from 2009 to 2014, one person, 600+ npm packages. He effectively built half of Node.js's early ecosystem by himself. His projects include:

• Express.js — the most popular Node.js web framework, inspired by Ruby's Sinatra. Used by PayPal, Uber, IBM.
• Connect — Express's middleware foundation.
• Mocha — the most popular JavaScript test framework.
• Jade (later Pug) — template engine.
• Commander.js — CLI argument parsing.
• Stylus — CSS preprocessor.
• Koa — his next-gen framework after Express.
• Superagent, Co, n (Node version manager), and many more.

You almost cannot use Node.js without touching code TJ wrote. All of it during his LearnBoost years. In some sense, LearnBoost's edtech SaaS was the side project. The open-source output was the main job.

In 2014 TJ posted Farewell Node.js on Medium, announcing a switch to Go. He left, founded Apex (a Serverless tool), then went quiet. He's never publicly presented at any major conference — one of open source's most mysterious figures. He's been clear about his disillusionment too: "Open source doesn't pay the bills."

Cloudup, Automattic, and the open-source commercialisation lesson (2013–2015)

In 2013, the LearnBoost team built Cloudup — drag-and-drop upload-and-share that gave you a shareable link before upload finished. It caught Matt Mullenweg's attention. Automattic (WordPress's parent) acquired the whole team — including LearnBoost and all related open-source libraries (Socket.IO, Mongoose).

Rauch spent two years at Automattic. The influence was deeper than people credit. He watched first-hand how WordPress built a business empire on open source — WordPress free and open, Automattic monetising through WordPress.com hosting. The "open-source framework pulls users, cloud service monetises" pattern transferred almost verbatim to Vercel.

ZEIT / Vercel — the core idea (2015)

In 2015 Rauch founded ZEIT (renamed Vercel in 2020). Starting premise: deploying a website should be as easy as sharing a file. Back then deploying meant nginx config, CI/CD setup, SSL, DNS. ZEIT collapsed it to one command — now.

But the ambition was never "deploy tool." In an a16z conversation, he said it cleanly: "I've always treated open source as a shortcut to finding product-market fit — if people don't use your thing when it's free, you should probably do something else." Open source isn't charity. It's business validation.

Same year, he created Next.js. The relationship between these two things is the entire story — open-source framework pulls developers, cloud platform monetises. The same shape as WordPress / Automattic, Terraform / HashiCorp, Elasticsearch / Elastic. Vercel did it especially well because their open-source projects didn't just become tools — they became de facto industry standards, which made the conversion funnel exceptionally smooth.

The open-source map — from Socket.IO to skills.sh

The Vercel open-source ecosystem: 10+ years of accumulation.

Walking through Rauch's open source — personal era, LearnBoost, ZEIT, Vercel — makes one thing visible. Vercel's moat isn't a single product. It's the network of open-source nodes.

Phase 1 — Rauch personal + LearnBoost (2007–2013)

• MooTools core contributions. Joined the core team at 16, led MooTools Forge. Lost to jQuery; gave Rauch a deep model of how open-source communities operate.
• Socket.IO (2010). WebSocket real-time library, de facto standard. 61K+ stars. Poetic detail — Vercel's serverless architecture still doesn't support WebSockets, so the founder's most famous open-source project doesn't run on his own platform.
• Mongoose (2010). MongoDB Node.js ORM. 27K+ stars. Still the most popular MongoDB ODM. Same origin — LearnBoost needed it.

Plus, as covered, the entire TJ Holowaychuk pantheon — Express, Mocha, Commander, et al.

Phase 2 — ZEIT era (2015–2020)

Hyper terminal: a web-tech open-source terminal emulator, 43K GitHub stars.

• Now CLI (2015). ZEIT's first product. Type now in your terminal, your project is live. Evolved into Vercel CLI.
• Hyper (2016). Electron-based terminal emulator, 43K+ stars. "Web tech for the terminal" pioneer. Eventually surpassed by Ghostty and Warp, but it shows Rauch's terminal fixation goes back to childhood.
• micro, serve, pkg (2016). micro was a minimalist HTTP microservice framework whose "every file is a microservice" philosophy directly influenced the Serverless Functions design. serve was a single-command static server. pkg packaged Node projects into binaries. The "zero config" obsession is visible throughout.
• Next.js (2016). The main event. React was hot in 2016 but there was no good SSR story. Next.js filled the gap. Today over 200M weekly downloads. Walmart, Apple, Nike, Netflix, TikTok all on it. More below.

Phase 3 — Vercel era (2020 onwards)

• SWR (2019). React data-fetching hooks with stale-while-revalidate. Influenced React Query and successors.
• Turborepo (acquired 2021). Monorepo build tool. Founder Jared Palmer joined and became Vercel VP of AI. Core rewritten in Rust, combined with Remote Cache, dramatically speeds team builds. Stretched Vercel from deploy to build.
• SWC (sponsored / hired). Rust JS/TS compiler, 20–70× faster than Babel. Original author kdy1 hired full-time. Integrated into Next.js as default compiler. ByteDance, Shopify, Deno also use it.
• Svelte (hired Rich Harris, 2021). The React champion company paying the salary of a competing framework's creator. Underlying logic — Vercel sells frontend deployment, not any one framework.
• AI SDK (2023). Open-source TypeScript AI toolkit. 3M+ weekly downloads. Vercel's fastest-growing open-source project. More below.
• Workflow SDK / Flags SDK / Chat SDK (2025–2026). A series of small, sharp open-source SDKs that combined form "AI application infrastructure."

Acquisitions: Splitbee (2022, privacy-first analytics, folded into Vercel Analytics), ModelFusion (2024, TypeScript AI abstraction, folded into AI SDK 3.1).

Next.js — from byproduct to industry standard, and its controversies

Next.js's evolution is itself a piece of frontend history. 2016 — SSR. 2020 — ISR, solving large-scale static-site update pain. 2022–2023 — App Router + React Server Components, the most controversial era. 2025 — 200M+ weekly downloads, Next.js-powered sites grew from 35K to 4M+. March 2026 — Next.js 16.2 finally added a stable Adapter API.

Many new Next.js features have the best experience on Vercel. Other platforms have had to support them via undocumented APIs. Cloudflare and Netlify launched OpenNext specifically to solve that. The stable Adapter API took nearly three years from promise to ship.

Security incidents have hurt trust. The March 2025 middleware-bypass vulnerability (CVE-2025-29927) was handled in a way the security community called opaque. December 2025 was worse — a React Server Components RCE (CVE-2025-55182, CVSS 10.0) was actively exploited. Wiz's researchers reported nearly 100% success rate, with default configurations vulnerable. For a company sitting under millions of production sites, security trust is the company's neck.

Stringing all the open source together, the pattern is clear: Vercel isn't "build a project, sell the service." It's "build every node of a networked ecosystem." Framework, build tool, compiler, fonts, data fetching, AI toolkit — each one is a node in the network.

Platform and business model — the product is the moat

The model is simple on paper: open-source framework pulls developers, cloud platform monetises. The product line has long outgrown "deployment."

Infrastructure — Fluid Compute + Functions + CDN

Vercel's seven-layer platform: from infrastructure to agent ecosystem.

Fluid Compute: 'The power of servers, in serverless form.'

Fluid Compute (2025) is Vercel's new compute model — combines serverless elasticity with traditional-server concurrency. One instance handles multiple requests concurrently, dramatically reducing cold start. Supports Node.js, Python, Edge, Bun, Rust. Pricing also changed — Active CPU billing — you don't pay for I/O wait, which is enormously friendly to AI workloads. Vercel Functions support up to 800-second execution, 4GB memory.

Storage + security + observability

Storage. Vercel Blob (file storage), Edge Config (global KV, 99% reads <10ms). Vercel Postgres and KV are deprecated; now you connect via Marketplace to Neon, Upstash, Supabase, etc. Don't build a database — build a database store. Smart.

Security. Vercel Firewall (L7 WAF + DDoS protection), Bot Management, BotID (the "invisible captcha" partnership with Kasada — no image-clicking; identifies bots from deep-signal analysis).

Observability. Web Analytics (from Splitbee, cookie-less), Speed Insights (real-user perf), Observability Plus (custom queries), Drains (log export to Datadog et al.).

Developer experience — the real moat

Preview Deployments (every PR gets its own preview env), Rolling Releases (2025, progressive global deploys), Microfrontends (2025 GA, native), Vercel Queue (2025, native job queue), Image Optimization, ISR — each makes leaving Vercel a little harder.

That's the moat. Not any single feature being irreplaceable. The combined experience being irreplaceable. Any single competitor only covers a small piece.

The financials

Vercel ARR growth: $1M to $200M in six years.

Vercel funding history: $863M total, valuation 3x in 15 months.

Revenue curve: $1M (2019) → $5M (2020) → $21M (2021, +320%) → $51M (2022) → $86M (2023) → $144M (end 2024) → $200M (mid 2025). $100M → $200M in 15 months. Gross margin around 70%.

Funding: $863M total across 6 rounds. Latest is September 2025 — $300M Series F led by Accel + GIC at a $9.3B valuation. Includes ~$300M secondary buyback. Series F at $200M ARR = 45× revenue multiple. High even by AI standards.

Personnel: Stripe's former CBO as COO. Mitchell Hashimoto (HashiCorp co-founder) joined the board. Stripe's CFO as adviser. Headcount ~800–900.

Pricing controversy — the other side of the growth

Pricing is the developer community's most persistent complaint. A classic HN comment: "On first inspection Vercel's pricing is unbelievably expensive — bandwidth at $550/TB, a 128MB function fully loaded for a year is $60K." Multiple devs report migrating to a $6/mo DigitalOcean VPS without any user-facing change.

One indie hacker's case: $1,800/mo Vercel bill, forced to close new signups, "started hoping users would churn so I'd pay less." Migrated to Cloudflare Workers, bill collapsed. His read: "Vercel is delicious at small scale. Once you scale up you realise it wasn't designed for you."

The worst case is DDoS. A documented incident produced a $23,000 bill — malicious traffic still billed at $0.15/GB. Unlike AWS Shield or Cloudflare, which absorb attack traffic, Vercel passes the cost to the user. Spend Management was added later, but "charge first, govern later" damaged trust.

There's also a structural issue: AI workloads are particularly unfriendly to Vercel's billing. An AI chat session may take 60 seconds of compute and tens of MBs of memory. One developer reported a screenshot service used 494 GB-hours over 12 days of testing, projecting >1,200 GB-hours/month, $160+/month extra. Meaning Vercel is pushing AI while its own billing model is a drag on AI growth. Fluid Compute's Active CPU billing partially fixes this; the structural bandwidth pricing didn't change.

Competitive positioning — two flavours of lock-in

Vercel's three-front war: deploy platform / AI code generation / AI IDE.

The clearest way to read Vercel's position is to compare to Cloudflare. Cloudflare has 330+ edge locations (Vercel ~19 regions), cold start <5ms, no egress fees. Rauch has admitted on Twitter — Vercel seriously tried Edge Runtime, but migrated off because of CPU performance, latency variance, and slow connections to cloud services, then built Fluid Compute. He also said "no beef" and acknowledged that Cloudflare's CPU billing inspired Vercel.

But the lock-in shapes differ. Cloudflare is storage-shaped lock-in — once you use Durable Objects, R2, D1, your application logic is bound to Cloudflare; there's no portable Durable Objects alternative. Vercel is framework-shaped lock-in — once you use Next.js ISR, Server Actions, image optimisation, leaving means losing features.

AI pivot and agent ecosystem — why Vercel got it right

v0 — 3.5M users of "talk to make a website"

v0.app: 'What do you want to build?' — talk to generate an app.

Vercel's AI strategy: the v0 + AI SDK dual revenue flywheel.

Launched 2023, initially a component generator. By 2025 it had become a full AI dev platform — multi-file projects, GitHub sync, PR branches, connections to Snowflake/AWS databases, one-click deploy. January 2026 — domain switch to v0.app. February — token-based billing. ARR ~$42M, 21% of total revenue. Enterprise accounts make up more than half of v0 revenue.

The key point: v0's beauty isn't that it generates code — it's that it generates Next.js + React code that one-tap-deploys to Vercel. Generate → deploy → live, closed loop. Every v0 user is a potential Vercel platform customer. That dual revenue flywheel is something Bolt.new and Lovable don't have.

AI SDK — the utility layer for frontend AI

AI SDK: a unified TypeScript AI toolkit, npm install ai.

Open-source TypeScript library, unified API for 25+ AI providers. Handles streaming, tool calling, structured output, React hooks. 3M+ weekly downloads. AI SDK 6 (end 2025) adds Agent abstraction, human-in-the-loop, MCP. Same playbook as Next.js — open source captures mindshare, then converts to platform.

AI Gateway + Vercel Agent + Sandbox

Vercel Agent: 'The intelligence layer for shipping on Vercel.'

AI Gateway (alpha) — unified AI model routing layer, ~100 models, handles auth, usage, retry, failover. Vercel Agent (beta) — automated PR review and production diagnostics, validates fixes in Sandbox. Sandbox supports enterprise 32 vCPU / 64 GB.

skills CLI and skills.sh — the npm of the agent era

skills.sh: an open-source agent-skills ecosystem, already 91,623 skills.

Most interesting new move in early 2026. The skills CLI is an open-source command-line tool — npx skills add installs a skill pack into your AI agent. Supports 18+ agents — Claude Code, Cursor, GitHub Copilot, Gemini CLI, Windsurf. skills.sh is the online catalogue and leaderboard.

Vercel's official skills include React/Next.js best practices (40+ rules, 185K installs), Web Design guide (100+ rules), React Native, AI SDK, deploy skill. Stripe et al. published their own skills within hours. Snyk partnership for security scanning.

InfoQ summarised it: "MCP solved how agents talk to tools. Skills solves how developers share and discover agent capabilities." If this ecosystem holds the way npm did, Vercel has just claimed a critical position in the agent era.

Stringing this together: Vercel's official positioning has shifted from Frontend Cloud to Agentic Infrastructure.

Closing thoughts

Vercel's story is ten years on one sword. An Argentinian who dropped out of school, moved to SF at 18, spent ten years on open source, infra, and developer experience. By the time the AI wave arrived, his company was already the rails and the grid. SaaStr put it right — AI's "overnight successes" are usually a decade's road.

The risk is real too. Vercel's moat is composite experience, not a single technical fortress. As AI makes each single point more replaceable, can composite experience keep being the moat? That's the question Vercel has to answer over the next few years.