Windsurf CEO: how a GPU startup pivoted twice to build an AI coding tool

Executive overview

Most AI developer tools iterated on GitHub Copilot. Windsurf's founders scrapped their GPU infrastructure business in a weekend, bet on agentic coding before the models could support it, and shipped a forked VS Code IDE in under three months.

The core insight is that every competitive advantage is a depreciating asset — continuous, compounding insight execution is the only durable moat.

Startups die slowly when they stop generating and testing new hypotheses, not when individual bets fail.

The pivot: from GPU virtualisation to Codeium

• Company started in 2021 as ExaFunction — VMware-style GPU virtualisation for deep learning workloads
• By mid-2022: managing 10,000+ GPUs, $2M+ revenue, 8 people, Series A raised
• Saw that transformer models would commoditise custom ML pipelines — eliminating the rationale for their product
• Decision made in a single weekend; entire team told Monday; coding on Codeium started the same day
• Key conditions enabling the pivot: lean team, cash-flow positive, $28M raised, full team conviction on the new direction
• Shipped a VS Code extension within two months and posted on Hacker News

Building a better autocomplete: early technical edges

• First version used a free open-source model — materially worse than Copilot, but free
• Retained inference speed advantage from prior GPU runtime infrastructure
• Trained their own fill-in-the-middle model within months — outperformed Copilot on that specific capability by early 2023
• Expanded to all major IDEs (JetBrains, Eclipse, Vim) early, driven by enterprise need — JP Morgan, Dell were early customers
• Shared cross-IDE infrastructure meant low marginal cost to add each new editor

Why they built their own IDE (Windsurf)

• By mid-2024 enterprise revenue was well over eight figures from the Codeium extension product
• VS Code's ceiling limited the agentic UX they wanted to ship
• Forked VS Code; shipped across all operating systems in under three months with an engineering team of fewer than 25
• Core bet: agents, not chat+autocomplete, were the right paradigm — Windsurf was the first agentic editor
• Deliberately avoided over-configuration; invested instead in deep code-base understanding and fast in-place edits
• Maintained a unified timeline of developer and agent actions so the AI always has current context

Context retrieval: going beyond RAG

• Rejected pure vector-database RAG as insufficient for code — precision and recall need to be very high
• Built a multi-signal retrieval system: keyword search, vector search, abstract syntax tree parsing, and GPU-powered real-time reranking
• Motivation: a query like "upgrade all instances of this API" fails if embedding search misses even a few hits
• Prior GPU infrastructure enabled running reranking across large code chunks in under one second
• Design principle: strive for what works, not complexity; added AST parsing only after evals proved it necessary

Evaluation infrastructure

• Used a property unique to code: it can be run
• Core eval: take open-source commits with tests, delete implementation, ask the model to reproduce it; measure retrieval accuracy, intent accuracy, and test-pass rate
• Also masks partial changes to test intent prediction (similar to Google's autocomplete task)
• Evals drove investment decisions — complexity was added only when evals showed it helped
• Combination of eval-driven and vibe-driven improvement: evals suited to opaque retrieval systems; user data and intuition suited to UI-level changes

Using Windsurf in production

• Internal teams commit code changes frequently as the primary safety net when using agents
• Agents can change more than intended if intent is underspecified — surgical prompting and frequent commits reduce this
• Boilerplate and repetitive tasks (e.g. server deployments) now handled entirely within Windsurf workflows
• Non-technical employees build internal tools directly — removing the PM/engineer backlog bottleneck
• Non-technical users represent a meaningful share of active users; many never open the code editor, living entirely in the Cascade agent panel

Hiring and interviews in the AI coding era

• Still maintain a high technical bar — problem-solving skill remains the key proxy
• Interviews include AI-assisted sessions to screen out people who resist the tools
• Also include in-person sessions without AI — basic coding without AI assistance indicates reasoning ability
• Open-ended system design questions with trade-offs replacing pure algorithmic questions; no single correct answer
• Engineering headcount is growing, not shrinking — the problem ceiling is high and the 99% time-reduction mission requires many more capabilities (design, deploy, debug)

Opportunities for new AI coding startups

• Large legacy migration market: COBOL-to-Java, JVM version upgrades, Rails upgrades — billions spent annually, few specialised tools
• Automated alert and bug resolution: significant enterprise spend, no clearly dominant product yet
• Both are niches deep enough to support large companies, not just one winner
• General principle: pick one thing and do it exceptionally well rather than building another general coding assistant

Compounding advantage and the competition

• Every insight depreciates; Nvidia-style compounding requires continuous new bets
• Comfortable with a 50% failure rate on internal bets — 100% success signals insufficient ambition
• Competitor landscape has shifted repeatedly (Copilot → Devin → Cursor); long-term strategy + execution flexibility matters more than reacting to rivals
• Alpha over base models must grow proportionally as base models improve — the gap between foundation model output and 100% is the product opportunity
• Treat pivots as a badge of honour; most founders fail because they prefer consistent failure over the discomfort of changing course