In February, AI coding startup Windsurf was in talks to raise a significant new funding round at a $2.85 billion valuation, led by Kleiner Perkins. This valuation was double what the company had achieved just six months earlier, according to sources who spoke with TechCrunch at the time. However, the deal did not materialize. Instead, news emerged in April that the startup planned to sell itself to OpenAI for approximately $3 billion. While that deal ultimately collapsed, a key question remains: If Windsurf was growing rapidly and attracting venture capital interest, why would it consider selling at all?
Insiders reveal that despite the popularity and hype surrounding AI coding assistants, these businesses can be highly unprofitable. The high costs associated with running such platforms often result in “very negative” gross margins, as one source close to Windsurf explained. Essentially, the expenses of operating the product exceeded the revenue it generated.
A major factor behind these financial challenges is the reliance on large language models (LLMs). AI coding assistants face pressure to use the latest, most advanced—and most expensive—LLMs, as model developers continuously refine their offerings for coding and debugging tasks. This issue is further complicated by intense competition in the market, with rivals like Anysphere’s Cursor and GitHub Copilot already commanding large customer bases.
One potential solution to improve margins is for startups to develop their own models, eliminating the need to pay third-party providers like Anthropic and OpenAI. However, this approach carries its own risks. Windsurf’s co-founder and CEO, Varun Mohan, ultimately chose not to pursue building an in-house model due to the high costs involved. Meanwhile, model providers themselves are increasingly entering the AI coding space, with offerings like Anthropic’s Claude Code and OpenAI’s Codex.
Selling the business was seen as a strategic move to secure a high return before competition from model providers could erode Windsurf’s position. Industry observers believe similar margin pressures may be affecting other players in the space, including Anysphere, Lovable, and Replit.
Nicholas Charriere, founder of Mocha, a vibe-coding startup, noted that margins across code-generation products are either neutral or negative, describing them as “absolutely abysmal.” He added that variable costs for companies in this sector are likely within 10% to 15% of each other.
Unlike Windsurf, Anysphere has opted to remain independent, having reportedly turned down acquisition offers, including one from OpenAI. The company announced plans in January to build its own model, which could help control costs. However, its efforts faced setbacks when two key hires from Anthropic’s Claude Code team returned to their former employer shortly after joining Anysphere.
Some investors, like Erik Nordlander of Google Ventures, believe LLM costs will decrease over time. However, this expectation is not guaranteed, as the latest AI models have become more expensive due to their increased computational demands. OpenAI’s recent release of GPT-5, priced lower than competing models, has provided some relief, with Anysphere quickly integrating it into Cursor.
Anysphere has also adjusted its pricing to account for rising costs, particularly for heavy users, a move that surprised some customers. CEO Michael Truell later apologized for the lack of clear communication around these changes. Despite Cursor’s success—reaching $500 million in annual recurring revenue—investors caution that user loyalty may waver if a superior alternative emerges.
Given these challenges, Windsurf’s decision to exit the market appears justified. After the OpenAI deal fell through, key employees and founders joined Google in a deal that resulted in a $2.4 billion payout to shareholders. The remaining assets were later acquired by Cognition. While some criticized the move for leaving employees without roles at Google, a source familiar with the deal argued that it maximized outcomes for all stakeholders.
Beyond Cursor, other AI coding tools like Replit, Lovable, and Bolt continue to grow rapidly, though they too rely on external model providers. If even this high-revenue sector struggles to maintain profitability while depending on third-party models, it raises questions about the viability of other emerging industries built on similar foundations.