The bill nobody sees: what open source really costs in the age of AI

For twenty years, the cost of open source was measured in one currency: a maintainer's unpaid time. That bill never went away. But a second one has quietly arrived, and it isn't paid in evenings and weekends. It's paid in dollars — the compute a modern project burns just to keep moving.

The old bill: time, still unpaid

Start with the part everyone already knows is broken. By the numbers reported across the community in 2025, roughly 60% of maintainers are unpaid, a similar share have quit or seriously considered quitting, and 44% name burnout as the reason they walked. This isn't abstract. In November 2025, the Kubernetes project retired Ingress NGINX — one of its most-deployed components — not because it was obsolete, but because the people holding it up on nights and weekends could not continue. Other critical projects have frozen updates when a handful of maintainers hit the wall at once.

The uncomfortable truth underneath the burnout: the economics were always lopsided. Open source is the substrate the entire software industry is built on, and almost none of the value flows back to the people who maintain it.

Open source didn't get more expensive to build. It got more expensive to run.

The new bill: compute

Here's what's changed, and it's worth being precise about. Continuous integration — the classic "cost of running a project" — is largely not the issue for public repositories: GitHub Actions is free and unlimited for public repos, and in January 2026 the hosted-runner rates for everyone else were cut by up to 39%. Credit where it's due.

The new cost is AI, and it shows up in two places.

First, the tools maintainers now use to keep up. AI coding agents — Claude Code, Cursor, Codex — have become standard equipment, and they bill by the token. Real-world spend lands around $150–250 per developer per month in typical use; a developer running an agent all day on Sonnet-class pricing can burn closer to $594 a month, and heavy automation pushes into the $500–2,000 per engineer range. For a solo maintainer, that is a real line item that did not exist three years ago.

Second — and this is the structural shift — projects increasingly are AI. The demo in your README, the "explain this error" feature in your CLI, the natural-language layer on your API: every one of those is an inference call, and every call costs money, forever, per request. A project that ships an AI feature has signed up for a bill that scales with its own success.

The funding model is stuck in the coffee era

Now hold that new bill up against how open source is actually funded, and the mismatch is stark.

The dominant primitive is still "buy me a coffee" — a metaphor minted in an era when the cost of open source was a maintainer's attention. A coffee is a lovely gesture. It does not pay an inference bill. GitHub Sponsors, the most serious attempt at recurring support, reportedly counts on the order of a few thousand corporate sponsors against the hundreds of millions of companies that depend on open source — a rounding error away from a 100% freeloading rate. And the safety nets are thinning: Microsoft wound down its Azure Sponsored Subscriptions for open source in September 2025.

There are bright spots — grant funds and corporate pledges asking companies to pay a fixed sum per developer who benefits from open source. They matter. But they are cash, routed through a maintainer's personal finances, and then converted — by hand — into the compute the project actually needed. The loop is open.

Close the loop: fund compute with compute

If the cost is denominated in compute, the most direct way to fund it is compute. That is the entire idea behind TokenTip: a supporter's tip becomes spend-limited AI inference credit the maintainer can spend on any model. No conversion, no withdrawal, no detour through a bank account — the same currency the project burns, delivered straight to the work.

It isn't a replacement for a salary or a serious sponsorship. It's a primitive aimed precisely at the bill that the coffee metaphor was never built to cover — and one that any reader can act on in a single tap, from anywhere the project already reaches.

The software the world runs on is maintained by people who are tired and, increasingly, out of pocket. The least we can do is pay the bill in the currency it actually comes in.