Someone at Microsoft spent $28,000 on AI tools in 28 days. That number is not a rounding error or a misread invoice. It’s the single data point that put the entire company’s AI spending culture under a microscope, and it’s the kind of number that gets a memo written.
As reported by Gadget Review, Microsoft added a new column to internal compensation spreadsheets in 2026: ‘AI $ Usage Per Month.’ Roughly 350 US employees, out of more than 223,000 on the global payroll, voluntarily submitted data. Self-selected sample, yes. But the variance alone is worth paying attention to. The company-wide median sat around $300 per 28-day period. Some team medians told a different story entirely.
- CoreAI: ~$975 median
- Security: ~$526 median
- Microsoft AI: ~$490 median
- Cloud + AI: ~$325 median
- Company-wide median: ~$300
Individual maximums inside several departments cleared $10,000. The floor, inside those same teams, was sometimes tens of dollars. That spread is the real story.
When dashboards become leaderboards
Visible metrics have a way of becoming competitions. Inside Microsoft, that instinct quietly ran up the bill through a behavior employees started calling ‘tokenmaxxing’: deliberately burning AI tokens on low-value or useless queries just to rank higher on internal usage dashboards. Copilot made individual consumption visible to teams. So some people played the score instead of the job. It’s the enterprise equivalent of grinding side quests when shipping product is what actually matters.
In early August 2026, CoreAI EVP Jay Parikh sent an internal memo that made the company’s position explicit. ‘Tokenmaxxing is not what we are optimizing for. I want all of us focused on maximising outcomes that move the needle for our customers and our business.’ Division-level token budget targets are now active. Personal spending is tracked on internal dashboards. No hard per-engineer cap has been confirmed publicly. But the monitoring is real.
The model consolidation is the bigger competitive signal
The memo is only part of the story. Microsoft has shifted internal workloads to OpenAI’s GPT-5.6 Sol as the default model inside GitHub Copilot and related workflows, citing better value per token. Anthropic’s Claude for coding has reportedly been discouraged internally. That’s a quiet but meaningful reallocation of demand away from a direct competitor, and it’s the kind of default-setting decision that compounds over time across an organization this size.
This is the AI industry’s FinOps moment. The same reckoning cloud computing hit when ‘spin up whatever you need’ collided with six-figure AWS invoices. Someone built a dashboard. Someone wrote a policy. And suddenly efficiency mattered more than volume. Parikh’s framing is pointed: not fewer tokens, more impact per token. That’s a maturity story. Every enterprise running AI tools at scale right now is about to live the same one.
