Uber’s COO says AI cost is hard to justify

Uber’s COO Just Said the Quiet Part Out Loud: AI Spending Isn’t Paying Off Yet

For the past eighteen months, every major tech company has been playing the same game. Announce AI features. Hire AI talent. Burn tokens like there’s no tomorrow. And in internal meetings, tie performance reviews to how much AI you’re using.

It’s been called “tokenmaxxing” — the art of maximizing AI usage as a proxy for productivity and innovation. The assumption, rarely questioned aloud, has been that more tokens equals more value. More API calls means more features. More inference means more progress.

Andrew Macdonald, Uber’s COO, just questioned it.

In an interview with the Rapid Response podcast, Macdonald said something that probably made his peers at other companies wince. He said it’s getting harder to justify Uber’s AI spending. Not because the technology doesn’t work. Not because Uber is anti-AI. But because the link between higher token usage and actually shipping useful features for customers has become, in his words, “very hard to draw.”

Without that connection, Macdonald said, the costs are becoming hard to justify. Especially when Uber has been slowing hiring in other areas to fund its AI investments.

This is not a small comment. This is the COO of a $150 billion company, one that moves millions of people and deliveries every day, saying that the core assumption of the current AI boom might be flawed. At least for now.

Tokenmaxxing: The Productivity Illusion

Let’s unpack the term, because it’s important.

Tokenmaxxing is a piece of internal slang that has leaked out over the past year. It refers to the practice of maximizing the number of tokens (the basic units of text or code that LLMs process) that an employee or team consumes, often as a proxy for engagement with AI tools. If you’re not burning tokens, you’re not trying hard enough.

Uber’s own CTO recently made a comment that sparked an internal debate. He mentioned, presumably offhand, how fast the company was burning through its Claude Code budget. The implication was that high usage was a good thing—a sign that engineers were adopting AI, moving faster, writing more code.

But Macdonald is asking the follow-up question that nobody wanted to ask: moving faster toward what?

“It’s very hard to draw a direct link between higher token usage and shipping more useful consumer features,” he said. That’s COO-speak for: we’re spending a lot of money on this, and we can’t see the return.

This is not an anti-AI position. Uber clearly believes in AI. They use it for matching riders with drivers, for predicting demand, for optimizing routes. Those are mature applications. The question is about the new wave of generative AI—the chatbots, the copilots, the internal productivity tools. Are they worth what they cost?

The Hiring Slowdown Trade-Off

The most concrete detail in Macdonald’s interview is about trade-offs.

Uber has been slowing hiring in other parts of the business to fund its AI investments. That’s a choice. Every dollar spent on AI inference is a dollar not spent on a new product manager, a customer support agent, a marketing campaign, or a safety engineer.

For a while, that trade-off seemed obvious. AI was going to make everyone so much more productive that you could afford to hire fewer people and still get more done. The math worked on the back of a napkin.

But Macdonald is now saying the math is getting wobbly. If the token spending isn’t translating into clear customer value, then you’re just shifting headcount budget to cloud compute without a clear return. That’s not innovation. That’s arbitrage without a thesis.

This is the kind of conversation that happens quietly in CFO offices across Silicon Valley. Macdonald is unusual because he’s saying it out loud.

Duolingo Already Got There

Uber is not alone in this realization.

The article mentions Duolingo, which recently stopped evaluating employee performance based on AI usage. That’s a significant reversal. Duolingo was one of the earliest and most vocal adopters of generative AI, using it to generate language learning content, dialogue, and feedback. They leaned in hard.

And then they realized that measuring AI usage—tokens per employee, prompts per day, lines of AI-generated code—was creating perverse incentives. Engineers were generating more tokens, not better outcomes. The metric was eating the mission.

So they stopped. Performance reviews now look at actual shipped features, user engagement, learning outcomes. Not how many times you asked ChatGPT to refactor a function.

Duolingo’s shift and Macdonald’s comments are not coordinated. But they point in the same direction: a growing recognition that the raw consumption of AI is not the same as the creation of value. And that the conflation of the two has been a convenient fiction for a tech industry desperate to show it’s riding the next wave.

The Autonomy Elephant in the Room

Macdonald also talked about autonomy—self-driving cars and delivery robots—and his comments there are worth noting, even though they’re a separate topic.

He called autonomy “existential” for Uber. That’s strong language. He said it won’t take decades to arrive, which is a shift from the more pessimistic predictions of five years ago. But he also said it won’t happen in a couple of years.

The middle ground: autonomy is coming, it will matter enormously, but it’s not an immediate solution to anything. Uber has to run its business in the meantime.

This matters because autonomy is, in some ways, the opposite of the current AI spending problem. Autonomy is a long-term bet with a clear payoff if it works. Generative AI, at least in the productivity-tool sense, is a short-term cost with a very fuzzy payoff. Macdonald seems comfortable with the first and skeptical of the second.

Why This Matters Beyond Uber

The reason Macdonald’s comments are getting attention is not because Uber is the most important AI company. It’s not. The reason is that Uber is a real business. It moves people and things. It has a P&L. It cannot afford to burn cash on technology that doesn’t deliver measurable results.

That’s different from OpenAI or Anthropic, whose entire existence is AI. That’s different from Google or Meta, which have near-infinite other revenue streams to subsidize AI experiments. Uber has to make every dollar count.

So when Uber’s COO says the link between token usage and customer value is “very hard to draw,” it’s a signal that the era of blank-check AI spending might be ending for the rest of corporate America. Not for the frontier labs. Not for the big cloud providers. But for every other company that has been told, for two years, that they need to spend more on AI or be left behind.

Some of those companies are going to start asking the same question: what are we actually getting for this?

The Performance Review Problem

The internal debate at Uber started with the CTO’s comment about burning Claude Code budget. That’s a revealing detail.

Claude Code is Anthropic’s developer-focused tool. Burning budget means using it a lot. The CTO mentioned it positively, as a sign of engagement. That sparked a debate: is high usage actually good, or is it just expensive?

This is a management problem, not a technical one. Companies have tied AI usage to performance reviews. Engineers are incentivized to use AI tools even when they don’t help, because using them looks good on a dashboard. Product managers are incentivized to request AI features even when a simple rules-based system would work better, because AI is the buzzword of the year.

Macdonald is essentially saying: we need to unwind that. Usage is not outcome. Tokens are not value.

It’s a simple point. But in a hype cycle, simple points get lost.

What Macdonald Didn’t Say

A fair reading of Macdonald’s interview is that Uber is not abandoning AI. They are recalibrating.

He didn’t say AI is useless. He didn’t say Uber is cutting its AI budget to zero. He said it’s getting harder to justify the spending, and that without a clear link to customer value, the costs are a problem.

That’s a call for discipline. Not retreat.

The open question is whether other executives will follow. Will we see a wave of companies quietly walking back their AI investments? Or will they double down, convinced that the payoff is just around the corner?

The answer probably depends on what kind of company you are. If you are a cloud provider selling AI inference, you have every reason to keep the hype alive. If you are a rideshare company trying to make quarterly numbers, you have every reason to ask hard questions.

The Larger Pattern

Macdonald’s comments fit into a pattern that has been emerging for about six months.

First, the hype: every company must have an AI strategy. AI spending is a sign of forward thinking. Token usage is a metric of progress.

Then, the reality: most companies have no idea if their AI spending is working. The productivity gains are hard to measure. The customer-facing features are often gimmicks. The internal tools are used inconsistently.

Now, the reckoning: a few leaders are starting to say, quietly, that the emperor has fewer clothes than advertised. Duolingo changed its performance reviews. Uber’s COO is questioning the spend. Others will follow, likely behind closed doors.

This does not mean AI is a fad. It means the first wave of corporate AI adoption was sloppy. Companies bought without measuring. They implemented without understanding. They celebrated token usage as if it were synonymous with progress.

The second wave, if there is one, will be more disciplined. It will be led by people like Macdonald: operators who want to see a direct line from cost to customer value. If you can’t draw that line, you don’t spend.

What to Watch For

In the next six months, watch for three things.

First, more executives will make comments like Macdonald’s. Not all will be as direct. But the tone will shift from “AI is the future” to “AI needs to earn its keep.”

Second, companies will start publishing metrics about AI ROI. Not the hand-wavy “we believe it improves productivity” kind. Actual numbers. Time saved. Features shipped. Revenue attributed. The lack of such metrics today is telling.

Third, the startup landscape will adjust. The “AI wrapper” companies that have thrived on the hype will face harder questions from investors. The ones that can show real customer value, measured in something other than token volume, will survive. The rest will quietly fade.

The One Sentence to Remember

Macdonald’s interview is worth reading in full, but if you take away one sentence, take this:

“Without that connection, the costs are becoming hard to justify.”

That sentence is going to be quoted in boardrooms across the world over the next year. Not because Uber is special. But because Uber just said what everyone else has been afraid to admit.

AI is not magic. It’s a tool. Tools have costs. And if you can’t see what you’re getting for your money, you’re not being innovative. You’re just being spendy.

The tokenmaxxing era had a good run. But the era of accountability is coming. And Macdonald just rang the bell.