Companies aren't cutting AI spending — they're tearing down the leaderboards and engagement metrics that rewarded token usage over outcomes, after those metrics proved trivially gameable at Meta, Amazon, and Uber. Enterprise GenAI spend actually tripled even as the incentive structures collapsed.
Mike Taylor spent late 2025 running multiple Claude Code terminal sessions at once, building agents, and starting his mornings with AI-generated meeting prep. So when Gusto published an internal leaderboard tracking employee AI engagement earlier this year, Taylor expected to see his name near the top. He didn't expect what came with it: he was the company's biggest "superspender," racking up token costs wildly disproportionate to the work he was producing. Taylor isn't a junior engineer who got carried away. He is Gusto's chief financial officer, according to Bloomberg Businessweek's July 31 profile of the reckoning now sweeping corporate AI budgets.
Taylor's story is one data point in a pattern that by summer 2026 had become impossible to ignore. Through the first half of the year, a string of major employers built gamified leaderboards to encourage what Silicon Valley started calling "tokenmaxxing" — treating raw AI token consumption as a proxy for AI-forward productivity. Amazon had KiroRank. Meta had an internal tracker nicknamed "Claudeonomics." Both are now dead, and the way they died tells you more about what went wrong with corporate AI adoption than any ROI survey.
The root of this cycle traces back to a specific, dateable moment: Nvidia CEO Jensen Huang's public suggestion, echoed approvingly inside several of these companies, that engineers should be spending roughly half their salary-equivalent on tokens to stay competitive. Cognition CEO Scott Wu later dated what he called "the tokenmaxxing era" to January through May of 2026 and declared it over — a timeline that lines up almost exactly with when Amazon, Meta, and Uber each independently hit their breaking point.
Meta's version is the most fully documented. An internal memo reviewed by The Information and shared with roughly 6,000 employees this year disclosed that staff had burned through 73.7 trillion tokens in a single 30-day window, with costs tracking toward billions of dollars annually. The tracker behind it, Claudeonomics — named, un-ironically, after a competitor's product — ranked the top 250 token consumers with titles like "Token Legend." Some employees reportedly left AI agents running idle for hours purely to climb the rankings.
Meta CTO Andrew Bosworth tried to slow this down in an April memo, writing plainly: "Nobody should be using AI tools just for the sake of using them. All motion is not progress and token usage alone is not a measure of impact of any kind." The company is now building a centralized "AI Gateway" dashboard to track spending in real time and is steering employees toward its internal coding assistant, MetaCode, instead of third-party tools. Formal token budgets arrive in 2027.
Amazon's KiroRank followed the same arc on a faster timeline. The leaderboard, built to encourage adoption of Amazon's internal Kiro tool, instead trained engineers to aim AI agents at busywork solely to boost their scores. Amazon quietly deprecated it at the end of May. A separate disclosure this week added a sharper data point to the same story: leaked internal documents describe a $1.8 million cost overrun on a single Claude-powered task at Amazon that went undetected for five months, because no one had built the spend alerts to catch it.
Uber's version didn't involve a leaderboard, but the underlying failure was the same — consumption without a governor. CTO Praveen Neppalli Naga disclosed in April that Uber had already exhausted its entire 2026 budget for Claude Code, four months into the year, with per-engineer costs running $500 to $2,000 a month. The company has since capped spending at $1,500 per employee per tool per month. Uber's president and COO, Andrew Macdonald, later told colleagues the company hadn't established a clear relationship between rising token usage and the number of useful features it was actually producing.
Every one of these programs was built on the same unexamined assumption: that AI engagement — measured in tokens, sessions, or agent-hours — was a reasonable stand-in for AI-driven output. Economist Charles Goodhart's famous observation, that a measure stops being a good measure once it becomes a target, has rarely gotten a cleaner corporate demonstration. Meta and Amazon didn't just fail to prevent gaming; they built the incentive structure that made gaming the rational move. An employee facing a performance review that rewards visible AI usage, watching colleagues climb a public leaderboard for the same behavior, is not misbehaving by running agents on pointless tasks. They're following instructions.
The part of this story that's easy to miss, buried under the "AI backlash" headlines, is that the money isn't actually going away. Enterprise spending on generative AI nearly tripled in a single year, from roughly $11.5 billion in 2024 to about $37 billion in 2025, according to Menlo Ventures' survey of enterprise decision-makers — even as inference prices fell sharply over the same period. Only a fraction of that widely cited figure was Big Tech capex on chips and data centers; this is spend on applications and models specifically. Companies aren't retreating from AI. They're dismantling the specific incentive structures — leaderboards, engagement-weighted performance reviews — that rewarded usage instead of results, because those structures turned out to be trivially gameable.
That's a narrower and less dramatic story than "companies are cutting AI spending because it disappointed them," and it's also the more accurate one. The evidence for genuine disappointment exists too — an Accenture survey cited by Fortune found only 23% of C-suite leaders report widespread, sustained business value from AI — but conflating that finding with the leaderboard-killing spree overstates how much companies are actually pulling back, and understates how specific the fix is. What's being torn down is a measurement error, not an investment thesis.
There's a real case that the "crackdown" framing oversells what's happening. AI infrastructure spending by the largest tech companies is still climbing sharply — Meta alone plans up to $135 billion in AI infrastructure spend through 2026 and has committed $600 billion to data centers through 2028. Against numbers like that, capping an individual engineer's monthly tool budget at $1,500, or quietly retiring an HR leaderboard, looks less like discipline and more like cost-control theater at the edges while the real money keeps flowing into GPUs.
That critique has some force, but it misunderstands what kind of failure this is. The leaderboards weren't a rounding error in the AI budget; they were the mechanism translating dollars into work, and that mechanism was broken. A company can be simultaneously right to keep building data centers and wrong about how it measured whether its own employees were using the tools productively. Those are different questions with different answers, and treating a fix to the second as evidence about the first is exactly the kind of flattening this piece is trying to avoid.
It's also worth being precise about scale. Meta's internal token spend, even at 73.7 trillion tokens a month, works out to a few billion dollars a year — real money, but a rounding error against $135 billion in planned infrastructure spend. If the leaderboards were purely cosmetic, companies wouldn't have bothered dismantling them; a rounding error doesn't usually get its own CTO memo. The fact that Bosworth, Naga, and Amazon's Dave Treadwell all felt compelled to intervene personally suggests the incentive problem was doing real damage to how teams allocated attention, not just money.
The organizations moving fastest past this moment aren't the ones cutting AI usage — they're the ones building the harder infrastructure to measure what the usage actually produces. Salesforce has reportedly shifted internal measurement toward tracking completed agent workload rather than raw token volume. Meta's AI Gateway dashboard is explicitly designed to pair spending visibility with budget allocation, not just consumption tracking. That's a meaningfully harder problem than counting tokens on a leaderboard, because it requires agreeing, function by function, on what "output" even means for AI-assisted work — a question a leaderboard never had to answer.
Mike Taylor's position is the tell. Gusto's CFO didn't get fired for being the company's top AI spender; Bloomberg's reporting frames him as an unusually candid case study rather than a cautionary tale. What changes now isn't whether he keeps using Claude Code — it's what he's asked to show for it. The companies retiring their leaderboards this year aren't measuring less. They're finally being forced to measure something harder than how much their employees typed into a chat window.
The headline was the 94% drawdown from a $3.44 billion peak. The more useful arithmetic is the cash sitting inside the thing being bought — and the fact that everyone except the angels was paid in another private company's paper.
The biography is the least interesting part. What the board actually did was unbundle the CEO job — keeping Tim Cook on with a defined policymaker-engagement remit — and that separation says more about Apple's next decade than the appointment does.
Every outlet reported the size of Meta's teen-safety settlement and the screen-time limits. Almost nobody read the payment structure — where roughly $5.3bn is released only if two competitors act.