Microsoft's GitHub will move every Copilot plan to usage-based billing on June 1, 2026, ending years of flat-rate subscriptions on a product the company has been losing money on since launch. The billing announcement, posted Monday, replaces premium request units with a token-priced credit system, and reopens a debate AI critics have been waiting to have.
What changed, and what didn't
The framing is corporate poetry. Copilot, the post says, "is not the same product it was a year ago," with agentic workflows now driving costs that flat pricing can't cover. Read it again: Microsoft is acknowledging it cannot continue subsidizing the compute bills of nearly two million developers.
Those bills have been ugly for a while. In October 2023, the WSJ reported Microsoft was losing more than $20 per user per month on a $10 product, with some users costing the company as much as $80. The figure circulated and was largely waved off at the time. Microsoft's response was that Copilot was "thriving." Two and a half years later, the bill came due.
Base subscription prices are not changing. Copilot Pro stays at $10 a month and now includes $10 worth of AI Credits, where one credit equals one cent of token spend. Heavy users pay extra. Annual subscribers also face new model multipliers, with Claude Opus 4.7 jumping from 7.5x to 27x.
The bubble argument, revived
Tech critic Ed Zitron, a familiar name in AI economics skepticism, used the GitHub news to anchor a sprawling long essay arguing monthly AI subscriptions were never economically viable. His core claim: AI products have been letting users burn $8 to $13.50 in tokens for every dollar of subscription, and the math eventually catches up.
The supporting data is hard to dismiss, even if the tone gets exhausting. Anthropic and OpenAI have already moved enterprise clients to token-based pricing. Cursor, the AI coding startup, operated with negative gross margins until very recently and only reached "slight" gross-margin profitability after introducing a proprietary model and routing to cheaper alternatives like Kimi, per TechCrunch. Anthropic's own developer docs now estimate Claude Code costs $150 to $250 per developer per month. The previous figure was $6 a day for the average user. Quietly raising your published cost estimate by 4x in a few weeks is its own kind of admission.
And then there's OpenAI. Leaked projections obtained by The Information show the company expects ChatGPT Plus subscribers to drop from 44 million to 9 million in 2026, an 80% collapse. The plan to compensate is to grow ChatGPT Go (the $8 ad-supported tier) by 36x to 112 million subscribers, mostly in price-sensitive markets like India and Brazil. That would be the largest paid-user acquisition target in software history, on a one-year deadline. Calling it "ambitious" undersells the problem.
Where the case overreaches
Zitron also argues that the data centers being built today will be obsolete before they pay back, citing NVIDIA's annual chip refresh cycle and a six-year GPU depreciation schedule. This is the part of the argument worth pushing back on.
The H100, the workhorse training chip for the current AI generation, is three years old. There is no visible drop in demand for it. The A100, released in 2020, is exactly six years old (Zitron's stated end-of-life mark) and remains in active deployment across cloud providers. The "GPUs become worthless after six years" claim assumes obsolescence happens on a hardware timeline rather than a workload one. So far, that has not been borne out.
None of which means the broader picture is fine. It means the strongest version of the bubble argument doesn't need the GPU-obsolescence prop, and probably weakens itself by leaning on it. The data center economics, the OpenAI revenue projections, the Oracle debt situation, the negative gross margins at code-gen startups: those facts stand without help. People wrote the same kind of bubble pieces about mobile carriers in the early 2000s, and not all of them were wrong, but none of them were exactly right either.
Who's next?
The GitHub preview bill experience launches in early May, giving subscribers a side-by-side estimate of their April spend under both pricing models before the June 1 cutover. The bigger watch is whether Anthropic or OpenAI follow Microsoft's lead and move consumer subscribers to token billing. Zitron calls that scenario "closing time." He has been calling something closing time for a while.
