In February, AI coding startup Windsurf was in talks to raise a major funding round at a $2.85 billion valuation led by Kleiner Perkins — double its valuation from just six months earlier, according to sources. But that deal never closed. Instead, by April, news emerged that Windsurf planned to sell to OpenAI for roughly $3 billion.

That acquisition famously collapsed — but it raised a bigger question: If Windsurf was growing so fast, why sell at all?

Industry insiders say the reality is that AI coding assistants, despite their hype, can be money pits. One source close to the company told TechCrunch that Windsurf’s gross margins were “very negative” — meaning it cost more to operate the product than it could charge customers.

The main culprit? The high cost of using large language models (LLMs). AI coding tools face constant pressure to integrate the latest, most advanced — and most expensive — models, since improvements in coding, debugging, and related tasks are moving fast.

Add to that fierce competition from players like GitHub Copilot and Anysphere’s Cursor, and the pressure on margins gets even worse.

Why Building Their Own Models Matters

The clearest path to better margins would be for startups to build their own LLMs instead of paying suppliers like OpenAI or Anthropic. But that’s risky and expensive.

Windsurf’s co-founder and CEO Varun Mohan ultimately decided not to take that route. Compounding the challenge, model makers like Anthropic (Claude Code) and OpenAI (Codex) are now direct competitors in the AI coding space.

Selling, then, was a strategic move — a way to lock in a strong return before being undercut by their own suppliers.

The Broader Pressure on AI Coding Startups

Windsurf isn’t alone. Industry figures say margins for most code-generation tools are “neutral or negative”. Nicholas Charriere, founder of Mocha, estimates that variable costs across the sector differ by no more than 10–15%.

Cursor maker Anysphere has so far resisted acquisition offers — reportedly including one from OpenAI — and announced in January it would build its own model. But even that move has faced hurdles: two leaders from Anthropic’s Claude Code team joined Anysphere in July, only to return to Anthropic two weeks later.

In theory, falling LLM costs could relieve some of this pressure. But in practice, some of the newest, most capable models are actually getting more expensive due to higher compute demands.

Cursor’s Balancing Act

OpenAI’s newly launched GPT-5 is priced below Anthropic’s Claude Opus 4.1, and Anysphere quickly added it as an option for Cursor users. At the same time, Anysphere raised prices for customers using Anthropic’s newest model — a change that surprised some Pro subscribers, leading CEO Michael Truell to issue a public apology for unclear communication.

Despite Cursor hitting $500 million ARR in June, investors warn that loyalty in this market is fragile. A better product from a competitor could quickly lure users away.

Windsurf’s Exit and What Comes Next

After the OpenAI sale fell apart, Windsurf’s founders and key employees joined Google in a deal worth $2.4 billion to major shareholders. The rest of the company was sold to Cognition. While some criticized CEO Mohan for leaving roughly 200 employees without roles at Google, a source close to the deal said it maximized outcomes for all staff.

The takeaway? Even some of the fastest-growing startups in the LLM era — like Replit, Lovable, and Bolt — remain heavily dependent on model makers, making profitability a constant uphill battle.

If AI coding, one of the most lucrative and in-demand use cases today, struggles to sustain margins, the implications for other emerging AI-driven sectors could be even more challenging.

Source