Enterprise AI token spend hits ROI reckoning as Meta, Uber, AWS blow through budgets

Enterprise AI Token Spend Hits ROI Reckoning

Major cloud companies and Fortune 500 firms are burning through AI budgets at an unsustainable rate, forcing a reckoning on whether token spending translates to actual productivity gains.

Meta ran "token maxing dashboards" that encouraged employees to rack up AI usage for leaderboard status rather than business value. Employees left processes running overnight to boost their numbers, driving up costs with minimal return. Uber blew through its annual token budget within months, and its chief operating officer acknowledged the company is still struggling to understand the financial impact and ROI of AI spending. AWS spent roughly $500 million in a single month on AI, according to Axios reporting.

The scale of misspending suggests a broader pattern: companies used 2026's dramatically improved AI capabilities to greenlight work that should have stayed in the backlog. They're now chasing "make work" projects—employees using AI agents to check the weather, managers asking agents to scan contact lists for recently added contacts—that have no measurable impact on the bottom line.

The bull case rests on cost deflation. Token expenses will drop sharply as capabilities move into open source, get distilled into cheaper models, and hardware depreciates. What costs $500 million today might cost $50 million next year. That argument holds if the same output becomes available for a fraction of the price. But it doesn't excuse current misallocation.

The deeper problem is budget psychology. When managers receive a token budget—say, $100,000—they face pressure to deploy it. Returning unspent capital signals failure, even if restraint would have been wiser. This mirrors the perennial problem of annual marketing budgets: spend it or lose it. Token maxing dashboards simply made the incentive visible and measurable, which is why Meta and others have reportedly scrapped them.

The real test comes in the next earnings cycle. Companies will have to show material changes in operating expenses and explain whether token spending produced revenue gains or merely substituted AI costs for headcount reductions with no net improvement to profitability. Spencer Raskoff from Match Group illustrates the more disciplined approach: his company maintains single-digit-millions in token spend while systematically tracking ROI the way it would for digital marketing dollars.

The corrective impulse is already visible. Uber, Meta, Microsoft, Salesforce, and DoorDash have all signaled efforts to ration AI usage and ensure it drives measurable productivity. This happened quickly enough—within months of the euphoric rollout phase—that it suggests healthy self-correction rather than structural dysfunction.

What remains unsettled is whether companies will actually change behavior once the initial shock wears off. Jevons Paradox predicts they won't: when something becomes cheaper to use, people use more of it, not less. Goodhart's Law compounds the problem: once a measure becomes a target, it stops being a good measure. Token spend is now both cheaper and more measurable, which means the temptation to chase volume will persist.