This week, three companies admitted the same thing: GitHub paused new Copilot Pro sign-ups. Anthropic pulled Claude Code from its $20/month plan. xAI put Grok 4.3 behind a $300/month subscription.
The message is identical: AI’s all-you-can-eat era is over. The real cost of compute exceeds what flat-rate plans can sustain, especially when AI stops doing autocomplete and starts working on its own for hours.
And yet Microsoft keeps selling Copilot for Microsoft 365 at $30 per user per month like nothing’s happening. How?
I tried to figure it out. The answer is more interesting than it looks.
The dirty secret: nobody shows up to the gym
The first reason is embarrassingly simple: most people who have a Copilot license, simply don’t use it. Only 35.8% of employees with Copilot access actively use it. The rest — nearly two-thirds — have the license assigned and never touch it.
Out of 415 million commercial Microsoft 365 users, roughly 16 million pay for Copilot. Penetration is under 4%, and a significant chunk of those 16 million generates almost no server load.
Companies that roll out Copilot across the board without a structured adoption plan end up with 40–70% of licenses unused in the first year.
It’s the gym-in-January effect. Everyone signs up, few show up. Microsoft collects the subscription fee without bearing the corresponding compute cost. Margins hold up as long as people don’t walk through the door.
E-mail summaries and agents coding through the night are not the same thing
The second reason is structural, and it explains why Copilot 365 and GitHub Copilot face very different fates — despite being in the same company.
Typical Copilot usage in Microsoft 365 is lightweight: an e-mail summary in Outlook, a draft in Word, a meeting recap in Teams. A few hundred tokens per interaction: the user asks, Copilot answers, done.
But GitHub Copilot is a different world: a coding agent can run in parallel across a codebase for hours, open dozens of files, generate and rewrite code in sessions that consume millions of tokens. GitHub itself admitted that “agentic workflows have fundamentally changed Copilot’s compute demands” and that “it’s now common for a handful of requests to incur costs that exceed the plan price.”
It’s the difference between sending a text message and streaming a movie in 4K. Same carrier, same network, radically different costs.
The productivity Copilot can afford flat-rate pricing because it does small things; the coding Copilot can’t, because it does big things.
The hidden second layer
The third reason is the most subtle, perhaps the most important.
Microsoft has already built the infrastructure for the post-flat-rate era. It just tucked it under the hood.
The $30/month license covers basic usage: Copilot in Office apps, summaries, drafts, chatting with company data. But the moment a company wants to build custom agents with Copilot Studio, that usage gets metered in “Copilot Credits” — a consumption unit billed separately, typically through Azure.
An agent processing 500 emails a day in a mailbox can burn 18,500 credits per month — roughly $185 — on top of the flat license. An employee who builds an HR agent and shares it company-wide generates Azure consumption for every colleague who uses it without a license.
Flat-rate is the front door. The real pricing scheme is already consumption-based. Most users just don’t know it yet because they haven’t reached that level of usage.
Copilot doesn’t need to make money: it needs to sell everything else.
The fourth reason is more strategic: Copilot at $30/month doesn’t need to be profitable on its own. It’s the Trojan horse that pushes companies toward more expensive E5 and E7 plans, drives Purview adoption for data governance, and generates Azure consumption through extensions.
Twenty years ago, Microsoft’s average revenue per user was $200–300 a year. Today, with Copilot, security, and the full 365 ecosystem, it’s reached $1,500 per user per year. Copilot is the multiplier, not the product.
Amazon doesn’t make money on Alexa. It makes money on everything you buy because Alexa is in your house. Microsoft is playing the same game: Copilot doesn’t need to be a profit center — it needs to be the reason you never leave the ecosystem.
But how long can this last?
The pressure signals are there for Copilot 365, too.
Since April 2026, free Copilot Chat in Office apps has been restricted for organizations with more than 2,000 employees. Price increases across several M365 suites take effect in July 2026. And GitHub Copilot — same ecosystem, same company — is shifting to token-based billing after weekly costs doubled since the start of the year.
Meanwhile, Google chose the opposite path: Gemini included in Workspace plans at no additional cost. The competitive pressure is real.
The truth is that the Copilot 365 flat-rate formula holds thanks to an unstable equilibrium: lightweight usage, low adoption, contained compute costs, and an ecosystem that monetizes elsewhere. That equilibrium works today. But as companies start actually using what they’re paying for, and as workflows become more autonomous and compute-intensive, the same forces that broke flat-rate pricing in coding will arrive here too.
The question is not whether the formula will change. It’s when.
In AI as in the gym, the best business model is one where the customer pays and never shows up. The difference is that at the gym, nobody’s working around the clock to convince you to go every day.
(Actually I should go to the gym, but I didn’t go yet – I think I’ll start in January 😂).
Donald Mucci 