Token consumption is exploding at Uber. Useful consumer features are not. Both of those things are reportedly true right now — and they cannot continue to coexist without someone being forced to make a very uncomfortable decision about capital allocation.
That’s not my framing. That’s essentially what Uber’s own president and COO, Andrew Macdonald, said in a recent interview. The company reportedly burned through its entire annual AI budget in the first four months of 2026. And when asked whether that spending translated into better products for riders and drivers, Macdonald’s answer was a carefully worded but damning admission: the link “is not there yet.”
I think this moment deserves more attention than it’s getting. Not because Uber is uniquely reckless — it almost certainly isn’t — but because it’s one of the few companies senior enough, and honest enough, to say the quiet part out loud. And if it’s true at Uber, one of the most data-rich, engineering-heavy companies on the planet, it’s probably true almost everywhere.
The Thesis
The dominant narrative around enterprise AI spending goes something like this: companies that invest aggressively in AI tooling today — copilots, code assistants, model APIs — will compound those productivity gains into a structural competitive advantage within two to three years. Lag behind now and you’ll spend the rest of the decade catching up. The urgency is real, the logic goes, even if near-term ROI is murky.
I believed a version of this narrative. I still believe AI will reshape how software gets built. But Uber’s admission forces a sharper question: what exactly are we waiting for, and how long is a reasonable waiting period before the capital markets — and the board rooms — stop accepting “it’ll be clearer in the coming quarters and years” as an answer?
Macdonald put the stakes plainly: “If you’re not actually able to draw a direct line to how much useful features and functionality you’re shipping to your users, that trade becomes harder to justify.” He’s talking about the trade-off between token costs and headcount. Uber’s CEO Dara Khosrowshahi has already confirmed the company is hiring fewer humans to offset rising AI costs. That’s a real, irreversible decision being made on the basis of productivity gains that the COO simultaneously admits haven’t been measured.
Supporting Argument: The Measurement Problem Is the Real Crisis
Here’s what I think most coverage of this story is missing. The issue isn’t that AI tools aren’t working at Uber. They may well be. The issue is that no one there — at a $150-billion-plus company with a world-class data infrastructure — has built the instrumentation to know whether they’re working.
Macdonald acknowledged that “maybe implicitly there is more that is getting shipped,” but that’s a guess, not a measurement. When your entire ROI argument rests on implicit assumptions, you don’t have an AI productivity story. You have an AI faith story.
Consider what that means in aggregate. Uber spent $3.4 billion on R&D in 2025 — 9 percent more than the year before — and is now blowing past even that elevated baseline in 2026. If the world’s ride-sharing giant, which processes millions of data points daily and employs armies of data scientists, can’t instrument the productivity impact of its AI tools, it’s hard to imagine which enterprise actually can. The measurement gap may be industry-wide, not company-specific — which would make the collective ROI narrative across enterprise AI a shared illusion, not just Uber’s problem.
This connects to a broader pattern that tech executives have been quietly obscuring: the gap between what AI tools promise in demos and what they actually deliver at scale in production environments. Uber is just the rare company that said it on the record.
Supporting Argument: Replacing Headcount Before Proving Productivity Is Backwards
The sequencing here should trouble investors more than the spending itself. The standard logic for technology-driven labor substitution is: prove the technology works at scale, measure the output per dollar, then make staffing decisions accordingly. Uber appears to be doing this in reverse — cutting human hiring first, justifying it with AI spend, and acknowledging afterward that the productivity link hasn’t been established.
That’s not a technology adoption strategy. That’s a cost-cutting exercise dressed in AI language. And the risk isn’t hypothetical: if the AI tools don’t deliver the feature velocity that justifies the headcount reduction, Uber doesn’t just lose the productivity gain — it loses the institutional knowledge and velocity that those un-hired engineers would have provided.
The disruption to hiring pipelines caused by AI adoption has already created structural gaps in many tech organizations. Uber may be accelerating that dynamic without the evidence to support it. There’s also a compounding risk: as AI coding tools proliferate, organizations are discovering that the maintenance cost of AI-generated code can quietly offset the development speed gains — a cost that rarely appears in the headline productivity metrics being used to justify headcount reductions.
Supporting Argument: Token Costs Are the New Cloud Bill — and Just as Misunderstood
Remember the early cloud era, when companies excitedly migrated workloads to AWS and Azure, only to receive their first bill six months later and realize no one had modeled the ongoing consumption costs? Token consumption for large language model APIs is following the exact same arc — just faster and at greater scale.
Macdonald explicitly flagged this: “We’re going to have to start talking about token consumption and the associated cost versus headcount.” That framing — token cost as a line item comparable to a salary — is actually the correct way to think about this. But it’s a framing that most enterprises adopting AI tools haven’t operationalized yet. They’re approving AI tooling licenses the way they once approved SaaS subscriptions: optimistically, without modeling marginal consumption costs.
For market participants watching AI infrastructure spend, this is a signal worth taking seriously. The skepticism Wall Street is developing around AI spending has so far focused on hyperscaler capex and model training costs. The enterprise consumption layer — companies like Uber paying per-token for Claude Code, Copilot, and their equivalents — represents a second wave of AI spending that’s only beginning to be scrutinized. Uber blowing its annual budget in four months is an early data point for what that scrutiny will eventually find.
The Strongest Counterargument
The most serious pushback to my thesis comes from people who have actually worked inside large-scale AI deployments — and it’s worth steel-manning properly.
The argument goes: of course you can’t measure AI productivity gains in the short term, and demanding that you can is a category error. Transformational technologies always have a J-curve of adoption where costs spike before productivity follows. We saw it with electrification in factories, with enterprise software in the 1990s, with cloud migration in the 2010s. In each case, the companies that held their nerve and continued investing through the measurement ambiguity came out ahead. Uber demanding ROI accountability after four months of a multi-year transition is the kind of short-termism that causes companies to abandon genuinely valuable investments too early.
There’s something to this. The J-curve argument is historically grounded, and Macdonald himself acknowledged that clarity “maybe will become clearer over the coming quarters and years.” Anthropic, the maker of Claude Code, would likely argue that the productivity compounding hasn’t started yet — that the tooling is still in the adoption phase. That’s a defensible position.
But here’s where I think it breaks down. The J-curve argument assumes that companies are investing in a technology with a proven eventual payoff — and that the only question is timing. That was a reasonable assumption for cloud migration, where the cost and capability case was already established. For AI coding assistants at the current frontier, the eventual productivity ceiling is genuinely unknown. And crucially, Uber isn’t just being patient — it’s simultaneously making irreversible headcount decisions based on productivity gains it admits it cannot measure. Patience is appropriate when you’re waiting to measure something. It’s reckless when you’re making structural org changes while you wait.
Moreover, critics like Mark Cuban have raised pointed questions about whether the underlying economics of current AI products can ever produce returns at the cost structures being built. That’s a different concern from mine, but it rhymes with it: the aggregate enterprise spend on AI consumption may be building a cost base that the productivity gains structurally cannot justify.
Why It Still Holds
Even granting the J-curve argument its full weight, Uber’s admission changes something important: the evidentiary standard that enterprises and their investors should require before accepting “the link isn’t there yet” as a satisfactory explanation for nine-figure spending overruns.
What Uber is describing isn’t a measurement gap that will naturally close as adoption matures. It’s an instrumentation failure — the company hasn’t built the systems to track whether AI-generated code is shipping more useful features per dollar than human engineers would have. That’s not a technology problem; it’s a management and measurement problem. And it’s fixable. But fixing it requires acknowledging it exists — which is exactly what Macdonald did, to his considerable credit.
The companies that will come out ahead in this cycle won’t be the ones that spent the most on AI tokens. They’ll be the ones that built rigorous feedback loops between AI spend and product output early enough to course-correct before the capital is irretrievably committed. Right now, those companies appear to be a minority. Uber, by going public with its uncertainty, has at least earned the chance to become one of them.
For investors, the takeaway isn’t to short AI — it’s to demand the same honesty Macdonald showed from every enterprise claiming AI productivity gains on their earnings calls. The real determinant of AI success at the enterprise level has always been measurement and data discipline, not model capability. That’s the overlooked variable in nearly every AI ROI conversation happening right now.
The Prediction
Within the next four quarters, I expect at least two or three other large enterprises to make similar admissions — either voluntarily or under analyst pressure during earnings calls. When that happens, the conversation will shift from “AI is productivity-enhancing” to “which organizations can actually prove it,” and companies with rigorous AI measurement frameworks will command a valuation premium over those running on AI faith. What would prove me wrong: a credible, methodology-transparent productivity study from a major enterprise showing measurable feature-velocity gains attributable specifically to AI tooling. That study doesn’t exist yet. When it does, the narrative changes.
