Cursor’s use of Moonshot AI’s open-source Kimi model highlights how companies are turning to open-source AI to reduce dependence on costly third-party APIs and counter the growing impact of the AI “token tax.”
Open-source models are emerging as a critical strategy for AI companies seeking to escape the growing “token tax” that is eroding margins across the industry. The issue has gained attention following reports that Microsoft is cancelling most internal Claude Code licences after compute costs exceeded the cost of the employees the tools were meant to augment, while Uber reportedly exhausted its entire 2026 AI budget within four months.
The token tax refers to the cost premium paid by companies that build products on third-party AI models. While model providers operate at internal cost and can subsidise heavy users, API customers pay retail rates that include supplier profit margins, making inference a major operating expense.
The challenge is reflected in industry data. An ICONIQ Capital survey found AI-native product gross margins are projected to reach 52% in 2026, still well below the 75%–85% margins typically achieved by mature SaaS firms. Inference alone accounts for roughly 23% of revenue at scaling-stage AI companies.
Cursor offers one of the clearest examples of how open source may help address the problem. After facing heavy inference costs from Anthropic and OpenAI APIs, the company launched its Composer model and later disclosed that Composer 2 was built on Moonshot AI’s open-source Kimi model. Around 75% of the compute budget was directed towards Cursor’s own training efforts.
The shift from rented inference to internally controlled AI infrastructure reportedly helped Cursor achieve slight gross-margin profitability on large enterprise accounts, underscoring a broader lesson: owning, training, or customising open-source models may become essential for improving AI economics and reducing dependence on dominant model providers.
