Uber and Amazon executives say AI costs are getting harder to justify

Welcome to Eye on AI. Here’s the part that changes the room when you’re hearing it in Paris at VivaTech: leaders are no longer talking like AI is “always on” and budgets can absorb anything. They are talking like the math has finally shown up. And according to Uber’s COO, the bill has moved from theoretical to urgent, with AI spend getting harder to justify after Uber burned through its entire 2026 AI budget in four months.

That cost shock is the backdrop for a broader message coming out of VivaTech: the free-for-all era may be ending. Amazon’s SVP Peter DeSantis framed it as a normal phase of new technology adoption. But even he acknowledged the lived experience behind the headlines: companies start fast, then wake up to the realization that they are spending a bunch of money and need a better way to budget and control usage.

Now add the regulatory and sovereignty layer, because this is not just a finance story. The briefing is happening as “sovereign AI” concerns are front and center in Europe, with executives focused on how much technical dependence matters when the supply chain can be yanked. The source points to a specific example: the U.S. government abruptly shut down foreign access to Anthropic’s Mythos-tier models last week. In conversations at the conference, that action was treated as a stark signal of what happens if the U.S. decides to pull the plug, leaving Europe to grapple with technological dependence on America.

Put differently: budgets are being questioned, and access might also be at risk. When both constraints tighten at once, boards and execs get more serious about the systems they build and the vendors they rely on. If you cannot count on stable access to particular frontier capabilities, you are forced to design for redundancy, cheaper alternatives, and governance that can survive a policy curveball.

Meanwhile, inside companies, the shift shows up in procurement and model strategy. Philippe Rambach, Schneider Electric’s chief AI officer, told Fortune that the focus internally has moved toward matching use cases to cheaper, fit-for-purpose models. His point is blunt: “On the solutions that we build, we are very cautious to use the right model; you don’t always need to use the latest frontier model.” In his view, the question of AI cost is becoming more and more important, and companies need to measure it and include it directly in business cases, business plans, and decisions.

This is a practical reversal of how many organizations rolled AI out. The source notes that many firms handed out licenses liberally and encouraged heavy experimentation. That rollout style created an expectation that employees should “just use it,” for everything from work tasks to trivial activities like checking the weather. The second-order effect is predictable: usage scales up, costs scale up, and value does not automatically scale with it. In other words, companies learned the hard way that more usage does not equal more ROI.

There is also a behavioral trap executives are now trying to unlearn. The source cites an Axios-consultant observation: people tend to automate what they dislike, not what creates value. That matters because it suggests the early AI deployments may have been optimized for convenience or relief, not for outcomes that improve revenue, reduce costs in a measurable way, or strengthen operations. When that becomes obvious to finance teams, the conversation moves from experimentation to optimization, from “are we adopting?” to “are we getting returns?”

And we are seeing that shift with numbers and anecdotes that sharpen the urgency. The source reports a consultant told Axios that a client burned through half a billion dollars in a single month after failing to cap AI usage for employees. It also notes Uber’s rapid budget burn. Those stories are a warning label for any company where AI access is frictionless. If usage is not capped, it will find its own growth path, and that growth can outrun your ability to evaluate benefits.

So what does “coming back to reality” look like in practice? The source frames the new posture as tighter controls and more selective buying. Amazon’s DeSantis described this as part of the technology lifecycle: initial agility, followed by a period of figuring out how to efficiently use it and budget for it. For executives, that implies governance that sets boundaries, procurement choices that match model capability to task, and measurement that connects usage to outcomes.

The strategic stakes for peers are straightforward: you either put AI into a system where costs are controlled and performance is tied to business cases, or you end up explaining to leadership why experimentation is consuming resources faster than it produces value. And if sovereignty and access risks remain on the table, the pressure increases to ensure your AI strategy is not built on a single dependency, single model tier, or single regulatory posture.