Article illustration 1

Image Credits: Yuichiro Chino / Getty Images

In February, AI coding startup Windsurf appeared unstoppable—poised to secure fresh funding at a $2.85 billion valuation, doubling its worth in just six months. By April, it aimed for a $3 billion acquisition by OpenAI. Both deals collapsed, revealing an uncomfortable truth beneath the sector’s hype: even the most promising AI coding tools are hemorrhaging money due to the crushing costs of large language models (LLMs).

The Margin Meltdown

Insiders confirm that despite soaring user growth, AI coding assistants operate with "very negative" gross margins. As one source close to Windsurf bluntly stated: "It cost more to run the product than the startup could charge for it." The core issue? Reliance on third-party LLMs like Anthropic’s Claude and OpenAI’s Codex, whose fees devour revenue. Startups face relentless pressure to integrate the latest—and most expensive—models, as coding and debugging capabilities are key battlegrounds in the AI arms race.

"It’s a very expensive business to run if you’re not going to be in the model game," emphasized a source familiar with Windsurf’s operations.

This cost crisis is compounded by ferocious competition. Giants like GitHub Copilot leverage existing user bases, while newcomers like Anysphere’s Cursor and Replit fight for market share—all while model providers morph into rivals. OpenAI and Anthropic now offer their own coding tools, squeezing startups from both sides.

Survival Strategies: Build, Buy, or Bail

Faced with unsustainable economics, startups see two paths:
1. Build proprietary models to control costs, as Anysphere is attempting.
2. Sell early to capitalize on valuations before margins erode further—Windsurf’s rationale for pursuing the OpenAI deal.

Anysphere’s gamble illustrates the volatility. After rejecting OpenAI’s acquisition overtures, it hired Claude Code engineers to build its own model—only to see them return to Anthropic weeks later. It also sparked user backlash by adding fees for heavy Claude users, forcing CEO Michael Truell to publicly apologize. Still, the startup bets on falling LLM costs, with GV’s Erik Nordlander noting:

"The inference cost today is the most expensive it’s ever going to be."

Yet that assumption is shaky. Newer models demand more computational resources, and while OpenAI’s GPT-5 recently undercut Claude on pricing, long-term trends remain uncertain.

The Ripple Effect

The crisis extends beyond Windsurf and Anysphere. Nicholas Charriere, founder of rival Mocha, reports "absolutely abysmal" margins industry-wide, with variable costs uniformly high. Despite Cursor’s $500M+ annual revenue, investors warn user loyalty is fragile if alternatives emerge. As model providers vertically integrate, even thriving startups face existential pressure.

Windsurf’s fate—a talent acquisition by Google and an asset sale to Cognition—may preview sector-wide consolidation. For Replit, Lovable, and others, the message is clear: Building atop another company’s AI foundation invites peril. If coding tools—among the most mature AI applications—struggle to monetize, nascent industries face even steeper odds. The AI gold rush is here, but the picks and shovels are bankrupting their makers.

Source: Marina Temkin, TechCrunch