The AI Spending Reckoning: When "More Tokens" Stops Making Sense

The Honeymoon Is Over

For the past two years, the tech industry has operated on a simple faith: spend more on AI, get more out of it. Pour tokens in, productivity comes out. The math seemed obvious — until it didn't.

In May 2026, Uber's COO Andrew Macdonald said something that cut through the noise: it's becoming harder and harder to justify the company's AI spending. Not because the tools are broken. But because no one can draw a straight line between the money going in and the value coming out. ^[1]

Uber's "Head-Exploding" Moment

The story starts a few weeks earlier, when Uber's CTO Praveen Neppalli Naga went viral after telling The Information that Uber had already blown through its entire 2026 Claude Code budget — in just four months. By March, Claude Code usage had jumped from 32% to 84% across Uber's roughly 5,000-strong engineering organisation. Individual engineers were spending between $500 and $2,000 a month on tokens alone. Around 70% of code committed at Uber now originates with AI. ^[3]

Macdonald described this as a "head-exploding moment" — one that sparked urgent internal debates about token consumption, headcount trade-offs, and whether any of it was actually moving the needle.

"That link is not there yet, right? I think maybe implicitly there is more that is getting shipped, but it's very hard to draw a line between one of those stats and, 'Okay, now we're actually producing 25% more useful consumer features.'"
— Andrew Macdonald, Uber COO ^[1]

Uber's CEO Dara Khosrowshahi had already signalled the pressure earlier in May, announcing a hiring slowdown to offset the company's surging AI investment costs. ^[1]

Microsoft Quietly Walks Away

Uber isn't alone. In one of the most telling signals of the moment, Microsoft — the world's largest software company, and a major investor in OpenAI — quietly began cancelling most of its internal Claude Code licences inside its Experiences and Devices group (the division behind Windows, Microsoft 365, Teams, and Surface).

Affected engineers were told to migrate to GitHub Copilot CLI by June 30, the last day of Microsoft's fiscal year. The official reason: toolchain unification. The unofficial reason, as The Next Web put it bluntly: the bill. ^[3]

The economics are brutal. Token-based pricing behaves nothing like the per-seat software licences that finance teams know how to model. Agentic coding sessions run for hours, spawn parallel threads, and generate enormous volumes of context. The cost scales with how much the model thinks — and modern AI agents think a lot. When the company with the most leverage in the room walks away from a vendor whose product its own engineers prefer, the signal is unmistakable. ^[3]

An Industry-Wide Repricing

This isn't just an Uber or Microsoft problem. It's a structural condition spreading across the industry:

Duolingo walked back its decision to include AI usage in employee performance reviews after staff questioned whether they were being asked to use AI for its own sake, not for outcomes. CEO Luis von Ahn admitted it felt like accountability for activity, not results. ^[1]
Target has expressed anxiety about usage-based pricing models for AI agents, weighing costs carefully before committing. ^[2]
Starbucks shut down an AI inventory management experiment after concluding the system simply couldn't be trusted. ^[2]
Nvidia VP Bryan Catanzaro told Axios that for his own team, the cost of compute now exceeds the cost of the employees using it — a remarkable admission from the company selling the chips. ^[3]

Meanwhile, Gartner has placed generative AI squarely in the trough of disillusionment, predicting that 25% of planned 2026 AI budgets will slip into 2027 as proofs of concept fail to clear procurement. A separate Gartner study found that only 28% of AI infrastructure projects fully deliver against their business case. ^[3]

The "Tokenmaxxing" Trap

Big Tech coined the term "tokenmaxxing" — the philosophy of using AI as aggressively as possible, measuring employees by their AI usage, and treating token consumption as a proxy for innovation. For a while, it felt like a competitive edge. ^[4]

But the cracks are showing. The problem isn't that AI tools don't work. Many of them work remarkably well — well enough that engineers use them constantly, and that constant use is precisely what breaks the financial model. A 2024 MIT analysis, widely circulated in finance circles, suggests that at current pricing, AI automation pencils out as cheaper than human labour for only roughly a quarter of the jobs people assumed it would replace. ^[3]

The deeper issue is that each new generation of agentic AI, by design, consumes more tokens per unit of work — because it reasons longer, plans more elaborately, and runs more parallel processes. Per-token costs are falling, yes — roughly tenfold every 18 months. But per-task token consumption is rising faster. The maths isn't improving. ^[3]

What This Means Going Forward

The AI spending reckoning isn't a sign that AI is failing. It's a sign that the industry is maturing — and that the easy, faith-based phase of AI investment is ending. The next phase will demand something more rigorous: actual proof that the spend produces outcomes.

Companies that survive this repricing won't be the ones that tokenmaxx the hardest. They'll be the ones that ask the harder question first: not "are we using AI?" but "is the AI working?"

That's a healthier question. And frankly, it's about time someone started asking it.

Sources

Stay in the loop

Keep up to date with the latest news and updates