Microsoft has started rationing one of its most visible internal resources: AI tokens. In an email obtained by 404 Media, executive vice president Jay Parikh told staff that the company is not optimizing for token volume, and that divisions would now carry AI token budget targets as Microsoft tries to squeeze more value from the models its own engineers use.
Parikh framed the change in the language of resource discipline. He wrote that the company is updating its internal rules and treating token consumption with the same discipline as any other critical resource, and that employees should be judged on outcomes for customers and the business rather than on how much they consume. Engineers are racking up bills measured in the hundreds, sometimes thousands, of dollars a month, and no target spend figure has been shared.
The company has also reached for a cheaper dial: OpenAI’s GPT-5.6 is now the default model inside Microsoft, chosen to get greater value from the token budget. That follows a report in May that Microsoft had quietly dropped most of its in-house Claude Code licenses and was pushing engineers onto GitHub Copilot CLI before the fiscal year ended, and GitHub’s June move to usage-based billing, which bills internal use in AI Credits instead of raw tokens.
The episode is a window into tokenmaxxing, the practice of treating the number of tokens an employee burns as proof of productivity. It flatters internal adoption dashboards but inflates cost. Per-token prices are down about 98 percent from late 2022, yet enterprise AI outlays have roughly tripled, in part because agentic software chews through far more tokens per job than the autocomplete-style interactions that set the original price points. Token-priced tools do not behave like seat-based software licenses, and finance teams are still learning to budget around that difference.
Microsoft is not the only company pulling this lever. The Next Web counts Walmart, AT&T, Amazon, Meta, and Uber among the employers that have capped or throttled employee AI spending, part of a broader shift from an everything-is-free experimentation phase toward procurement discipline. The timing is notable because Microsoft’s own finances are strong; its most recent quarter came in ahead of analyst estimates on revenue and operating income. The Register, which reported the email’s contents, said Microsoft had nothing to add when asked about it, and that the company is not seeking an overall reduction in AI use, merely more impact per token.
One employee who passed the email to 404 Media said the caps felt like an admission that the firm running the AI infrastructure could not afford to let its own people use the products freely. That reading is arguably too dramatic, since Microsoft’s balance sheet is not in question. The more mundane explanation is that the company is applying metered economics to itself before selling the same discipline to customers.
Sources: Microsoft tells engineers to curb their token-burning enthusiasm (The Register, Aug 5, 2026); Microsoft Tells Engineers ‘Tokenmaxxing Is Not What We Are Optimizing For’ (404 Media, Aug 4, 2026); ‘Tokenmaxxing is not what we are optimizing for’: Microsoft tells engineer to calm down on AI usage (TechRadar via Yahoo Finance, Aug 5, 2026); Microsoft tells employees to stop tokenmaxxing, sets division-level AI budgets (The Next Web, Aug 4, 2026)