Varun Mohan on Windsurf and the Future of AI Coding

Varun Mohan on Windsurf and the Future of AI Coding

transcript windsurf codeium ai-coding ide agentic enterprise lenny-podcast

Varun Mohan on Windsurf and the Future of AI Coding

Guest: Varun Mohan — Co-founder and CEO of Windsurf (formerly Codeium); previously GPU virtualisation and compiler software. Background in autonomous vehicles; co-founder met at middle school; 1M+ users within four months of Windsurf launch.
Host: Lenny Rachitsky
Source: Lenny’s Podcast. Recorded ~early 2025 (~6 months after Windsurf launch).

Overview

Varun Mohan tells the origin story of Windsurf via two major pivots — GPU virtualisation → Codeium (AI autocomplete) → Windsurf (agentic IDE) — and shares his frameworks for startup decision-making, hiring, and product cannibalisation. The AI coding product content complements Cursor and Devin perspectives.

Key ideas

  1. Two pivots. Started as a GPU virtualisation company (eight people, profitable, managing 10,000+ GPUs). Recognised transformers were going to dominate and infrastructure differentiation would erode → pivoted overnight to Codeium (AI autocomplete, free across all IDEs). Hit ceiling of IDE platform constraints → forked VSCode to build Windsurf (agentic IDE) with full control over AI experience.
  2. Cannibalize yourself every 6–12 months. “Every six to 12 months, it should make our existing product look silly.” The goal is to build future-state products inside the current company before a competitor does.
  3. Dehydrated entity hiring. Hire as little as possible; think of each hire as a unit of water. Only hire again when the company feels dehydrated. Purpose: maintain focus and force AI amplification of existing team before adding headcount.
  4. Value at the application layer. Once every ML workload converged on transformers, infrastructure commoditised. The application layer — better UX, better workflows for developers — has no ceiling. That is where Codeium/Windsurf chose to compete.
  5. Agency as the most important skill. AI increases the ROI of technology production; engineers and builders who use AI tools as force multipliers will compound. The constraint moves from “how long does coding take” to “what problems do you choose to solve.”

Origin story: three stages

Stage 1: GPU virtualisation (~2021–mid 2022). Deep learning application infrastructure: run complex ML workloads on computers without GPUs, handling heterogeneous architectures (CNNs, GNNs, RNNs). Revenue, profitability, 8-person team. Assumption: ML model architectures would remain heterogeneous.

Stage 2: Codeium (~mid 2022–2024). Transformers converged; infrastructure differentiation thesis broke. Decision: move to application layer; vertically integrate inference infrastructure to build AI coding tool. Provided autocomplete for free across all major IDEs (VSCode, JetBrains, Vim, Emacs, etc.). Enterprise motion added (Dell, JPMorgan Chase) with private, on-prem, codebase-aware deployment.

Stage 3: Windsurf (~late 2024+). IDE platform ceiling: VSCode’s architecture capped how much AI capability could be surfaced. Forked VSCode to build a dedicated agentic IDE with full control. 1M+ users in 4 months.

Decision-making frameworks

On pivoting

  • Start from the assumption you will get a lot of things wrong.
  • When your foundational hypothesis is falsified (e.g., “transformers will be one of many architectures” → false), act quickly. Do not run old and new in parallel — it guarantees failing at the new.
  • Fall in love with the problem, not the solution.

On hiring

  • “Dehydrated entity” principle: operate lean until you genuinely feel understaffed.
  • AI increases per-engineer output dramatically; re-examine the hire threshold before adding headcount.
  • Annual output is now larger than the company’s total lifetime-to-date output — each year is a new lease on life.

On product strategy

  • Build things that are 6–12 months ahead. If you are not cannibalising your current product, someone else is.

Windsurf product

Cascade (agentic feature). Windsurf’s differentiator: AI that writes code, runs it, tests it, and iterates — a full agent loop. Not just autocomplete.

Enterprise deployment model. Private model hosting on customer infrastructure; deep codebase indexing; org-wide personalisation. Enterprise customers care less about “can it write code” and more about “does it understand our entire private codebase and is it compliant?”

Agentic capabilities. The forked IDE allows full control over how the AI interacts with the code, file system, and execution environment — things the VSCode extension model does not permit.

See also