Corporate America’s AI Hangover: The Boom Meets the Bills

The great AI spending spree is hitting a wall. Billions flowed in, expectations skyrocketed, and then the invoices started piling up.

Uber capped employee use of AI coding tools this month after its annual budget for one system ran dry in just four months. A senior executive admitted the expense was getting tough to justify, with little proof it improved rider features. Similar stories are echoing through boardrooms across the tech sector and beyond.

Amazon scrapped an internal ranking of AI token usage after staffers gamed it with pointless queries. One executive sent a blunt email: “Please don’t use AI just for the sake of using AI.” GitHub switched its Copilot assistant to a usage-based pricing model, forcing developers to confront the real cost of heavy reliance. These aren’t isolated complaints—they signal a broader shift.

The AI investment wave that lifted markets through 2025 is now entering its payback period. Companies that rushed to embed the technology everywhere are discovering a stubborn gap. The models work well in narrow applications. But scale them across thousands of workers and workflows, and the economics fall apart. Savings miss forecasts. Costs climb faster than returns.

A Bain & Company survey of 951 large firms found AI-driven savings running below projections, even as most companies planned to spend more. “The technology worked. The value didn’t arrive,” the report stated. OpenAI CEO Sam Altman called the revenue-visibility issue “the most fair criticism” of the moment.

The discomfort runs deeper. A February 2026 study from the National Bureau of Economic Research found that 90% of firms reported no measurable productivity impact from AI. Yet executives projected gains of 1.4% in productivity and 0.8% in output. It echoes the productivity paradox of the 1980s and 1990s, when computers spread widely before their economic benefits became clear.

Spending continues at staggering levels. US mega-cap tech companies were expected to commit $1.1 trillion on AI infrastructure between 2026 and 2029, with total AI-related outlays surpassing $1.6 trillion. Morgan Stanley projected global data-center investment near $3 trillion from 2025 to 2028. OpenAI alone has talked of $1.4 trillion over eight years for new data centers while reporting just $13 billion in revenue, with anticipated annual losses through 2028.

Ray Dalio warned this month that the AI surge displays classic bubble traits. “All great technology changes produce bubbles,” the Bridgewater founder said. He expects wealth built on paper to convert into harder money, with painful consequences.

Gary Marcus declared in early June that 2026 would likely see the bubble unwind. Cory Doctorow made a sharper point: the early internet grew more profitable with each new user. Generative AI grows less profitable. “Every new AI user makes AI less profitable, as does every new use for AI, and each generation of AI loses more money than the last,” he wrote.

Market concentration tells its own story. In late 2025, the five largest companies accounted for 30% of the S&P 500, the highest level in half a century. AI-related stocks drove roughly 80% of US market gains that year. Yet cracks have appeared. When Broadcom held its longer-term outlook steady, the Nasdaq dropped 4.2% and the Philadelphia Semiconductor Index fell 10.3%.

Executives inside the builders see the tension. Sundar Pichai has spoken of “elements of irrationality” in the trillion-dollar investment boom. Sam Altman has conceded overexcitement.

Some voices push back. Fidelity Investments noted that several classic bubble signals had not yet materialized. Free cash flows had not shrunk despite heavy capital expenditure. Valuations, while elevated, sat below dot-com peaks on certain measures. JPMorgan and Goldman Sachs argue that AI ties to genuine enterprise revenue and profit growth.

Still, the internal pushback from adopters carries weight. Uber didn’t abandon AI—it stopped treating it as free and unlimited. Amazon didn’t reject the technology—it rejected mindless deployment. These adjustments suggest a maturing market. Precision over proliferation. Targeted gains rather than blanket transformation.

The bubble’s revenge isn’t a market crash. It’s the quiet realization inside conference rooms that the easy wins were never easy. The models impress in the lab. They demand discipline, curation, and realistic expectations when loosed on real operations.

Productivity may yet rise. Certain workers already achieve dramatic output lifts when tools are applied with care. The broader economy has started its slower absorption phase. History shows such technologies eventually pay off. They rarely do so on the timelines or at the multiples that fueled the initial mania.

So the reckoning continues. Companies trim wasteful usage. Vendors adopt usage-based pricing. Boards press for measurable returns. The technology isn’t disappearing. The fantasy that it could be sprayed everywhere and pay for itself just did.