Microsoft Limits How Much AI Its Own Teams Can Use

Twenty-eight cents on one side, $89.32 on the other. That's what the same unit of delivered work costs in tokens, depending on whether the developer behind it is a light user or a heavy one. A ratio of 1 to 319, for roughly twice the output.

A gap that size eventually produces a budget. At Microsoft, it already has: since July, every division has a cap on AI spending.

What the internal memo says

The story comes from 404 Media, which obtained an email on August 4 from Jay Parikh, executive vice president at Microsoft, along with an update to Copilot's internal usage rules. The message fits in one line: "Tokenmaxxing is not what we are optimizing for." Stacking up tokens isn't the goal. The word itself isn't a Microsoft coinage: it's been circulating in the industry for months, and Parikh picks it up only to wave it off.

Three things change in practice. Every division has carried an "AI token budget target" since July 2026. Every employee can track their own consumption on an internal dashboard. And the default internal model becomes GPT-5.6, cheaper than the alternatives, so the company can "get greater value from our token investment."

The individual dashboard is the most telling detail. It's the meter that goes up the day an unlimited plan turns into a line item.

One caveat, and it matters: Microsoft did not respond to 404 Media's request for comment. This is an internal document reported by a third party, not a policy the company has announced publicly. That said, the underlying reporting holds up: a signed email from a named executive, a dated internal policy, and pickup from Techmeme, Slashdot, and The Next Web with no denial to date.

The warning sign came in May

Microsoft had already pulled most Claude Code licenses that spring from its Experiences and Devices group, the division behind Windows, Microsoft 365, Outlook, Teams, and Surface. The internal memo, reported by The Verge on May 15, was signed by group EVP Rajesh Jha and pushed thousands of developers toward GitHub Copilot CLI ahead of the fiscal year close at the end of June.

July's rules fill in the scale that was missing back then. Per the document, many of the company's own engineers were spending anywhere from a few hundred to a few thousand dollars a month in tokens. Each.

The price is falling, the bill isn't

The unit price of a token has been dropping for two years, every vendor says so, and it's true. What changed is volume. According to Nicholas Arcolano, head of research at Jellyfish, per-developer consumption has multiplied by 18.6 in nine months.

Agents are the reason. A developer typing prompts by hand has a physical limit: the length of a workday. An agent kicked off in the morning and left running until evening has none. Nobody watches the meter while it runs, which is precisely the point of setting it loose in the first place.

What heavy usage actually costs

Jellyfish, which tracks engineering-team activity, published the sharpest numbers in this whole story: 12,000 developers, 200 companies, across the first quarter of 2026.

The median developer costs $52.38 in tokens a month. At the 75th percentile, that jumps to $226.58. At the 90th, $691.14. That's the benchmark most executives staring at their own bill don't have, with no way to tell whether the number is normal.

The payoff is the real story. The most frugal 20 percent of developers shipped 11 merged pull requests over the quarter for about $3 in tokens. The top-spending 20 percent shipped 23, for about $1,822. That's $0.28 versus $89.32 per unit delivered. Jellyfish sums it up: roughly ten times the tokens for about twice the throughput.

Doubling your spend to double your output would be a good trade. Multiplying your bill by more than three hundred to double it is a different one entirely.

Uber, Accenture, Amazon

Microsoft isn't alone here, and it wasn't even first. Uber burned through its entire 2026 AI budget by April, in four months, before capping access to Claude Code and Cursor on June 2. Amazon pulled an internal leaderboard that had been ranking its heaviest token spenders. At Priceline, a Cursor contract renewal came back four to five times more expensive, according to Chris Reed, director of IT financial management.

The strangest detail comes from Accenture. Justice Kwak, who leads agent strategy there, said in an internal meeting whose audio later leaked that the company's own data shows engineers aren't the ones driving token spend. It's everyone else, on routine tasks like turning a PDF into a slide deck. The same firm had told staff back in February that skipping AI tools could cost them a promotion.

The Linux Foundation launched a "Tokenomics Foundation" in July to set measurement and billing standards. When an industry stands up an entire organization to explain its own invoices, the topic has stopped being a footnote.

Why this isn't just about Microsoft

A week ago, we looked at what AI is costing Microsoft on the investment side: $35.8 billion in capital expenditure for the single quarter ended in June, and a committed bookings backlog of $678 billion, up 84%. That was the story of a company spending enormously because it has paying customers waiting at the end of the capacity it's building.

Six days later, the same company turns out to be counting its own employees' tokens. Both facts describe the same story from opposite ends. Microsoft owns the data centers, resells the service, and therefore pays for AI at close to the best cost basis available anywhere on the market. If the best-positioned player in the world thinks a cap is necessary, nobody else has much of a case for treating this cost as background noise.

A power company that installs timers in its own offices simply knows the price of a kilowatt better than its customers do.

What we still don't know

No cap amount is public, and neither is the number of divisions involved. We know the budgets exist, not where they're set.

The most widely repeated market figures of the past few weeks, things like "enterprise AI bills have tripled" or "the average AI budget went from $1.2 to $7 million," circulate without a named source. The only solid numbers in this whole story come from a tooling vendor measuring its own customers. That, on its own, says something about how immature this market still is.

And finally, a merged pull request measures throughput, not value. Arcolano says as much himself: whether extreme spending pays off depends on the business value of the code shipped, and most companies still can't measure that.

Which is exactly what Parikh is asking his teams to do in that same email: "We are not optimizing for fewer tokens. We are optimizing for more impact per token." The line is fair. It just happens that impact is the one of the two numbers nobody has figured out how to count yet.