The high costs and thin margins threatening AI coding startups

In February, Windsurf, an AI coding startup, was reportedly in discussions to secure a significant funding round at a valuation of $2.85 billion, led by Kleiner Perkins. This would have marked a doubling of its valuation from just six months prior. However, those talks fell through, and by April, the startup made headlines again when it announced plans to sell itself to OpenAI for a similar valuation of $3 billion. Unfortunately, that deal also collapsed, leaving many to wonder why a rapidly growing startup would consider selling at all. Insiders have indicated that despite the excitement surrounding AI coding assistants, many of these businesses are struggling to maintain profitability. A source close to Windsurf revealed that the startup's operational costs are so high they lead to significantly negative gross margins, meaning the expenses of running the product exceed the revenue generated. This issue stems from the exorbitant costs associated with utilizing large language models (LLMs). AI coding assistants face immense pressure to continually provide the latest and most sophisticated LLMs, as developers are constantly refining their models to improve coding performance and debugging capabilities. This competitive landscape includes established players like Anysphere’s Cursor and GitHub Copilot, which already boast extensive customer bases. For many startups, including Windsurf, the path to better margins lies in developing their own models, thereby reducing dependency on costly suppliers like OpenAI and Anthropic. However, this path is fraught with challenges, and Windsurf’s co-founder and CEO, Varun Mohan, ultimately chose not to pursue building an in-house model, recognizing the potential financial strain. Meanwhile, companies like Anysphere have been growing rapidly enough to resist acquisition offers, including one from OpenAI. Reports indicate that Anysphere is also working on developing its own model, which could enhance its control over operational costs. Industry observers like Eric Nordlander from Google Ventures believe that the cost of LLMs is expected to decrease in the future, although that trend has not yet materialized, with expenses for some models actually on the rise. The recent introduction of OpenAI's GPT-5 at lower fees compared to competitors has put further pressure on the market. Anysphere quickly adopted this model for its Cursor users, even adjusting pricing structures to accommodate rising operational costs, which took some customers by surprise. Despite Cursor's impressive performance, reaching $500 million in annual recurring revenue by June, there are concerns about user loyalty. Investors speculate that if a rival company develops a superior product, Cursor’s customer base may not remain steadfast. After the failed deal with OpenAI, Windsurf’s founders and key employees transitioned to Google, resulting in a $2.4 billion payout for significant shareholders. The remaining operations were then sold to Cognition. While some critics have pointed fingers at Mohan for leaving behind approximately 200 employees, sources argue that the acquisition ultimately benefited all parties involved. The competitive nature of the AI coding sector raises questions about the sustainability of startups reliant on model makers. As this rapidly growing industry continues to generate substantial revenue, the challenges faced by Windsurf may serve as a cautionary tale for others in the space.