Windsurf dumped a $3 billion sale to OpenAI after its high-cost AI coding assistant proved a money sink. The startup was in talks with Kleiner Perkins for a $2.85 billion raise early this year, doubling its valuation in six months. The deal collapsed, and Windsurf then eyed an acquisition by OpenAI. That fell apart too.
The reason? AI coding assistants like Windsurf face brutal economics. Gross margins are reportedly “very negative,” sources told TechCrunch. Running the latest large language models (LLMs) for coding and debugging costs more than they earn.
“It’s a very expensive business to run if you’re not going to be in the model game,”
a source close to Windsurf said. Competitors like Anysphere’s Cursor and GitHub Copilot bring additional pressure with massive user bases.
Building proprietary LLMs is the most direct fix to cut costs. But Windsurf’s CEO Varun Mohan rejected that expensive route. Instead, the company chose to sell before the big AI model providers—OpenAI and Anthropic—could squeeze them out.
Other vibe-coding startups like Anysphere (Cursor), Lovable, and Replit face the same margin crunch. Nicholas Charriere, Mocha founder, called margins “abysmal” across the sector.
Anysphere is betting on growth and independence, turning down acquisition offers including one from OpenAI. It is pursuing its own model to control costs, but that comes with risks. The company recently absorbed rising fees from Anthropic’s Claude and passed costs to active users, sparking customer surprise and an apology from CEO Michael Truell.
“Margins on all of the ‘code gen’ products are either neutral or negative. They’re absolutely abysmal,”
Charriere added.
New model costs don’t always fall. Some latest AI models cost more due to complex, resource-heavy tasks. OpenAI’s freshly introduced GPT-5 undercuts Anthropic’s Claude Opus 4.1 on pricing, and Anysphere quickly added GPT-5 for Cursor users.
Windsurf’s founders exited to Google after the aborted OpenAI deal, netting shareholders $2.4 billion in a separate deal. The remaining business sold to Cognition.
If the fast-growing AI coding market can’t fix cost woes, the question looms: What about newer AI sectors relying on the same expensive LLM infrastructure?
TechCrunch event
San Francisco | October 27-29, 2025
