Kushner
Venture Capital

How long can AI agents like Cursor keep burning cash? Investors are divided.

Power users of Cursor and Replit are racking up huge bills to Anthropic and OpenAI. How long can the game continue?
gettyimages-2224508763CROP.jpg

Josh Kushner’s Thrive Capital led Anysphere’s latest financing at a $9.9 billion valuation.

Kevin Dietsch/Getty Images

The latest high-flying, cash-burning venture-backed business is the AI coding agent.

Tools like Anysphere‘s Cursor, Devin by Cognition and Claude Code provide AI-powered software engineering assistants that can execute entire coding tasks. But running them doesn’t come cheap: The prices per token that foundational models like OpenAI and Anthropic charge have stalled, while usage is ramping up.

And that’s a problem for the “freemium” business model, whose free product tier is often its most popular.

“For the first time in SaaS history, the marginal cost of adding on a user is not close to zero,” said Amanda Huang, who invests in AI infrastructure and developer tools at Bain Capital Ventures. The magical “software margins” that historically made business software tools so lucrative don’t really apply here.

AI applications all face a vexing accounting challenge: They have to pay both a cloud provider and an LLM, in order to run API calls.

Anysphere, which recently raised at a $9.9 billion valuation and runs the AI coding agent to beat, is currently operating with negative gross margins, according to several investors familiar with its business model. Cursor did not respond to PitchBook’s request for comment.

Several of Cursor’s rivals are in the same boat. And even tools like Anthropic’s Claude, which doesn’t have to pay an external LLM provider, still face the problem of AI code generation requiring massive computing resources.

The ‘power user’ problem

AI coding agents are victims of their own success: They’re bringing on board more “power users” and existing customers are using the apps more, not less. Some providers, such as Claude Code, have already implemented usage limits to address the issue.

For providers of such AI apps, each API call also requires GPU compute time from a cloud provider that is either pre-purchased or priced on demand.

“The $20 you’re spending as a power user does not nearly account for the cost that those startups are spending on your usage,” Huang said.

Some investors don’t see the negative margins as a concern. Andreessen Horowitz GPs Sarah Wang and Martin Casado called criticism of them “a boring cliché” in a blog post, arguing that “lower gross margins at a moment in time are not a long term indicator of a lack of a sustainable business model.” A16z led Anysphere’s Series A and co-led its Series B round with Thrive Capital.

Others are treading more cautiously. “The winners will be those who treat margin control as a core product feature from day one,” said Navin Chaddha, managing partner at Mayfield Fund.

Right now, coding agents in particular are “inference whales"—tasks require much more compute than other fast-growing AI applications like legal questions or balance sheet analysis. That could get cheaper over time as startups optimize those costs better.

AI-native companies in ICONIQ Capital‘s portfolio with over $100 million in annualized recurring revenue are reporting median free cash flow of minus-10%, according to a survey done by the growth investor. The $100 million ARR mark is typically when software companies hit cash flow breakeven, said Vivian Guo, a principal at ICONIQ Growth.

“My perspective is that [the negative gross margins] are a short-term phenomenon,” Guo said. She also pointed out that many AI-native companies are hitting the $100 million revenue mark in some cases years faster than software companies have in the past.

The price is right

Household names like Uber and Airbnb famously relied on venture capital instead of positive gross margins to sustain their businesses for years while they grew.

The AI boom is no different—model training is being heavily subsidized by VC cash, while coding agents (and other apps) are growing from huge demand for their products.

“They have the best problem in the world: They’re getting ripped off the shelves, they can’t hire fast enough, they have global brand recognition,” said Sangeen Zeb, partner at Google Ventures. “I don’t think I’ve ever seen anything like it at the application layer,” Zeb said.

The crucial question is whether costs will decrease fast enough to catch up to growing user demand.

“I do think that the cost per token is going to continue to go down,” Greg Sands, partner at Costanoa Ventures predicted.

But will it go down quickly enough? Huang says maybe not. “I imagine something needs to change on pricing.”

Learn about our editorial standards.