For two years, "tokenmaxxing" was the rage inside enterprise tech companies. Employees competed to burn the most AI tokens, internal leaderboards tracked usage, and managers rewarded high spenders as innovation champions. That era just ended—hard.
Microsoft canceled Claude Code licenses across key product divisions. Meta deleted its internal tokenmaxxing leaderboard. Uber admitted it burned through its entire 2026 AI budget in just four months. And one unnamed company reportedly spent $500 million in a single month after failing to set usage limits.
The problem: Token spending isn't translating to ROI. MIT research found that 95% of generative AI pilots delivered no measurable profit-and-loss impact despite billions invested. Now the bill is coming due, and enterprises are scrambling to implement governance they should have built from day one.
Update — August 5, 2026: The governance this article said was coming has now been written down, at one of the largest engineering organizations on the planet. Microsoft executive vice president Jay Parikh emailed engineers that "Tokenmaxxing is not what we are optimizing for," and that token spend would be managed "with the same discipline we apply to every other critical resource" — first reported by 404 Media and picked up by The Register on August 5. As of July 2026 Microsoft divisions run against a formal AI token budget target. The article below stands, but the open question it ended on — what does the governance actually look like? — now has a concrete, copyable answer. See Microsoft Wrote the Policy Down.
What Is Tokenmaxxing and Why Did It Fail?
Tokenmaxxing started as a proxy for innovation. The logic seemed sound: if you want to know which employees are pushing AI boundaries, track their token usage. More tokens meant more AI-powered work, more experimentation, more innovation.
Meta, Amazon, OpenAI, and others built formal or informal leaderboards encouraging engineers to compete for the highest token counts. It was gamification applied to productivity measurement.
The reality was predictably perverse. At Amazon, the Financial Times reported employees spinning up AI agents to complete wholly meaningless tasks just to keep their token stats high—stats that managers were now using to assess performance.
Then came the bills. Tokens aren't free. Every API call, every generated response, every agentic workflow costs money. And when you're running unlimited usage across thousands of employees with no governance, those costs scale exponentially.
Salesforce CEO Marc Benioff said his company's Anthropic bill will hit $300 million in 2026. Uber's COO told a podcast that AI token spending was getting "harder to justify" without a direct line to shipped features and functionality. And Microsoft—one of the biggest AI investors on the planet—started canceling Claude Code subscriptions after realizing the costs weren't producing measurable returns.
The Sticker Shock: Real Numbers from Real Companies
Here's what the AI spending reckoning looks like in practice:
Microsoft: Canceled Claude Code licenses for employees in several key product divisions, according to reporting from The Verge. The reason? Cost versus measurable impact didn't align. Later reporting pinned that down: most Claude Code licenses in the Experiences and Devices group, cancelled in May 2026.
Meta: Took down the informal tokenmaxxing leaderboard its employees had created. The company quietly shifted from encouraging unlimited usage to implementing hard governance on who can access expensive models.
Uber: Burned through its entire 2026 "token budget" in the first four months of the year, driven in part by high Claude Code usage. COO Andrew Macdonald said the company struggled to connect individual productivity boosts to any company-wide impact.
Salesforce: On track for a $300 million Anthropic bill in 2026. CEO Marc Benioff publicly wished for a "smart router" that could determine which queries actually need the most capable (and expensive) models versus cheaper alternatives.
Anonymous Fortune 500 company: One firm reportedly spent $500 million in a single month after failing to implement usage caps. This wasn't a planned investment—it was runaway spending with no governance guardrails.
Average enterprise AI spend jumped 65%: From roughly $7 million in 2025 to $11.6 million in 2026, even as most firms cannot prove a return.
Microsoft Wrote the Policy Down
Two months after this article published, the actual control regime became public — and it is more specific than anything a consultant will sell you.
Parikh was careful to say the goal is not austerity. "We are not optimizing for fewer tokens. We are optimizing for more impact per token," he wrote, alongside the line that everyone quoted: "Tokenmaxxing is not what we are optimizing for. I want all of us focused on maximizing outcomes that move the needle for our customers and our business." Four mechanics came with it, and the mechanics are the part worth copying.
1. The budget sits at the division, not the individual. As of July 2026, Microsoft divisions operate against an "AI token budget target", with restrictions on the table for divisions that overshoot. No engineer gets a personal quota that cuts them off mid-task at 4pm on a Thursday. The accountability lands on the leader who owns the P&L — which is where every other infrastructure line already sits.
2. Per-engineer usage is instrumented, and nobody is ranked on it. Individual engineers can now see their own token spend through updated internal Copilot guidance. That is the same telemetry Meta's leaderboard ran on, pointed the other way: visibility for the person spending the money instead of a scoreboard for their manager. Same data, opposite incentive. If you take one thing from the last two years, take that distinction.
3. There is finally a real per-seat number. Microsoft's own guidance concedes that many engineers spend "in the range of hundreds of dollars a month to a few thousand dollars in tokens." Run that across a 1,000-engineer organization and you get $6 million to $36 million a year in raw model consumption, on top of seat licenses. If your finance team hasn't modeled a range that wide, it will learn the answer from an invoice.
4. The cheap model is the default; the expensive one is an exception. Microsoft made OpenAI's GPT-5.6 the default internal model — GitHub Copilot had previously defaulted primarily to Anthropic's models — on the stated grounds that shifting more workloads to OpenAI models "helps us get greater value from our token investment". This is Benioff's smart router, implemented the blunt way: not routing intelligence, just a cheaper default and the friction of having to ask for anything else.
One piece of context this article was missing: GitHub itself moved to usage-based billing in June 2026, metering "AI Credits" rather than raw tokens — so even inside Microsoft, the cost of a Copilot user stopped being a fixed line item. Microsoft declined to comment on the email, telling The Register it had "nothing to add."
The uncomfortable part for buyers: the company telling its own engineers to burn fewer tokens on AI coding tools is the same company whose external pitch is that every developer you employ should be running Copilot. Both can be true — Microsoft pays for frontier inference and sells a seat — but it should change how you read the vendor's ROI slide.
Why Enterprise AI ROI Is Still Missing
The spending crisis isn't just about runaway costs. It's about the disconnect between AI investment and measurable business outcomes.
MIT's NANDA Initiative studied 300 public AI deployments and found that 95% failed to deliver measurable financial return. Supporting research backs this up: S&P Global reported 42% of companies abandoned most AI projects in 2025, IBM found only 25% of initiatives delivered expected ROI, and Morgan Stanley discovered just 21% of S&P 500 companies could cite any measurable AI benefit.
The problem isn't the technology—it's the implementation. Most companies are stuck in what AI analyst Azeem Azhar calls the "productivity J-curve": the period when spending on a new general purpose technology actually reduces productivity before the big gains arrive.
Think about how factories first adopted electricity. The initial move was replacing gas lighting with electric lights—a cost savings, but no productivity change. Next, they replaced central steam engines with large electric motors, but still ran machines off central drive shafts. Still no major gains.
It was only when companies redesigned entire factory layouts around individual electrified machines that productivity exploded. That took time, experimentation, and a willingness to rethink workflows from scratch.
Enterprise AI is in the same place. Most companies are replacing "gas lighting" (manual tasks with AI tools) but haven't redesigned workflows. Tokenmaxxing is the perfect example: gamifying token usage without connecting it to business outcomes.
The 5% Who Are Winning on AI ROI
The MIT research found that the small minority of firms generating real AI returns did three things differently:
1. They focused on back-office automation, not flashy use cases. More than half of generative AI budgets went to sales and marketing tools, but the biggest ROI came from automating high-volume, repetitive tasks: invoice processing, lease abstraction, data normalization, document review.
For CTOs and CIOs: The winning plays aren't the sexy ones. They're the workflows your team complains about—data cleanup, manual reconciliation, compliance documentation. That's where AI pays for itself in weeks, not years.
2. They implemented hard usage governance before scaling. No unlimited access. No competitive leaderboards. Every deployment had budget caps, usage limits, and pre-defined ROI metrics.
For CFOs: Treat AI spending like any other capital allocation decision. Set hard dollar limits per department, require ROI justification before expanding usage, and track cost per outcome (not cost per token).
3. They measured outcomes against specific workflows, not aggregate adoption. Instead of asking "Are we using AI?", they asked "Did this specific AI tool reduce processing time by X% in this specific workflow?"
For business leaders: Don't measure AI success by seats purchased or tokens consumed. Measure it by time saved, error rates reduced, revenue generated, or costs eliminated—tied to specific business processes.
What Enterprises Are Doing Now
The tokenmaxxing era is over. Here's what companies are implementing instead:
Smart routing systems: Salesforce's Marc Benioff called for this publicly—a system that automatically routes simple queries to cheaper models and reserves expensive models (like Claude Opus or GPT-4) for complex tasks. Some enterprises are building this in-house; others are waiting for vendors to deliver it. Microsoft's answer, two months later, was cruder and shipped immediately: swap the default model for the cheap one and make the expensive one something an engineer has to go out of their way to select.
Usage caps and budget limits: Hard spending limits at the department level. If the department hits its monthly token budget, the constraint lands on the person who owns the budget—not on an engineer mid-task. Microsoft's division-level targets are the reference implementation; per-seat hard cutoffs are the version that generates workarounds.
ROI gates for model access: Want access to Claude Code or the latest GPT model? Submit a business case showing expected time savings or revenue impact. No ROI justification, no access.
Model downgrading: Moving workloads from expensive frontier models to cheaper alternatives whenever possible. One CTO reported catching employees using Claude Opus to check the weather—a $0.10 query that should have cost $0.001 with a simpler model.
Use-case audits: Regular reviews of who's using AI for what, with immediate shutdowns of low-value or meaningless tasks.
The Bottom Line for Technical Leaders
If you're a CIO, CTO, or VP of Engineering, here's your action plan:
Audit current spending immediately. Break down AI costs by department, use case, and individual user. Identify where money is going and whether those use cases justify the expense.
Kill tokenmaxxing culture if it exists—but keep the meter running. Remove leaderboards, stop rewarding high token usage, and make it clear that innovation is measured by shipped features—not API calls. Do not delete the telemetry along with the leaderboard: Microsoft's move was to keep per-engineer usage visible to that engineer while ranking nobody on it. The number is a budget instrument, not a performance metric.
Implement usage governance. Set hard budget caps per team, require ROI justification for expensive models, and build smart routing to cheaper alternatives when possible.
Focus on back-office wins. Data processing, document automation, compliance workflows—these are where AI pays for itself fastest. Save the experimental use cases for when you have budget headroom.
Measure real outcomes. Track time saved, errors reduced, processing speed improvements—metrics tied to specific workflows. Aggregate adoption numbers mean nothing if they don't connect to business results.
The Bottom Line for Business Leaders
If you're a CFO, COO, or business executive, here's what you need to know:
AI spending is an operating expense, not an innovation exemption. Every dollar spent on tokens, seats, and implementation labor hits your P&L. Treat it like any other cost line—with governance, ROI tracking, and budget discipline.
Uncontrolled AI spend destroys value. For real estate investors, $50,000 of wasteful annual AI spending at a 6.0% cap rate erases roughly $833,000 of asset value. For SaaS companies, it hurts unit economics. For any business, it reduces net operating income with no offsetting benefit.
The 95% failure rate is avoidable. The companies getting ROI from AI aren't doing anything magical—they're applying basic capital allocation discipline. They set budget limits, measure outcomes, and kill low-ROI projects fast.
Ask three questions before expanding AI spend: (1) What specific business outcome will improve? (2) How will we measure that improvement? (3) What's our cost-per-outcome breakeven threshold? If you can't answer all three, don't spend the money.
The era of subsidized AI is ending. Hyperscalers are spending $675 billion on AI infrastructure in 2026, up 63% year-over-year. That investment will be recouped through higher token and seat pricing. The cost side of every AI calculation is going up, which makes ROI discipline more critical than ever.
What Comes Next
The tokenmaxxing collapse is a healthy correction. For two years, enterprises treated AI as an unlimited "innovation" budget without connecting it to measurable business value. That's over.
The companies that survive this reckoning will be the ones that treat AI like any other investment: with clear ROI expectations, hard budget limits, and ruthless prioritization of high-value use cases.
The companies that don't—the ones still chasing tokenmaxxing culture or treating AI spend as exempt from governance—will burn through budgets without results and get left behind.
The question isn't whether your company uses AI. It's whether your AI usage produces more value than it costs. For 95% of enterprises right now, the answer is no. But it doesn't have to stay that way.
Want to avoid the tokenmaxxing trap? Focus on back-office automation, implement hard governance before scaling, and measure outcomes tied to specific workflows. The era of unlimited AI spending is over. The era of disciplined AI ROI is just beginning.
Want to calculate your own AI ROI? Try our AI ROI Calculator — takes 60 seconds and shows projected savings, payback period, and 3-year ROI.