On May 26, The Verge ran a one-line headline: "Uber's CEO says AI spending is 'getting harder to justify.'" It reads like a passing executive quote, but the implications run deeper than that.
Uber spent $3.4 billion on research and development in 2025 alone, up 9% from the year before. Then reports surfaced that just four months into 2026, the company had already burned through its entire annual AI budget. That triggered a bigger question inside the company: is this money actually creating value?
The key answer came from Uber CEO and COO Andrew Macdonald in an interview. "We don't have that link yet. There's probably an implicit sense that more is shipping, but drawing a line from a metric like token usage to 'we're now building 25% more useful consumer features' is really hard."
That comment matters because it signals the start of AI adoption's second round — the "prove it worked" phase.
Does more token consumption mean more value?
Macdonald's diagnosis, restated plainly: token usage on AI tools like Claude Code is climbing astronomically. But the direct link between that rising usage and the value users actually receive isn't visible.
This isn't just Uber's problem. It's the stage nearly every company that has seriously adopted AI eventually reaches. In the early days, simply "using AI" looked like an achievement in itself. Code got generated automatically, reports got compiled faster, design mockups came back in multiple variations at once. That alone felt like enough of a result.
Now it's different. A year or more into adoption, the invoices have become impossible to ignore. Looking at those invoices, leadership starts asking: How much did this spending actually grow our revenue? How much happier are our users? How much better did our product actually get?
Surprisingly few companies can answer that clearly — even data-savvy giants like Uber.
The structural dilemma Uber is caught in
There's another telling detail in Macdonald's remarks. Earlier this month, Uber CEO Dara Khosrowshahi said: "We are offsetting the increase in our AI investment with a reduction in headcount growth."
The implication is clear. Uber is cutting labor costs and funneling that money into AI tools — hiring fewer people while buying more tokens.
For that trade to make sense, one condition has to hold: AI actually has to replace the work people would have done, and that has to show up in measurable terms — revenue, user experience, something concrete.
Macdonald's own words capture the weight of that condition: "We're going to have to start talking about token consumption and the associated cost versus headcount. If we can't draw a direct line to how much useful functionality we're shipping to users, that trade becomes hard to justify."
A decision to hire fewer people is hard to reverse. If a company shrinks its workforce and then concludes "AI didn't deliver what we expected," the whole operation gets shaky. Some Big Tech companies have already been through this cycle — laying off staff citing AI, then having to resume hiring.
The measurement trap
Why is measuring the ROI of AI investment so difficult? There are several structural reasons.
First, gains in frontline efficiency don't translate cleanly into final outcomes. Say one developer's coding speed goes up 30%. That doesn't translate into the company shipping new products 30% faster. In between sit planning, decision-making, review, deployment, marketing, and sales — and each of those stages is its own bottleneck. If coding gets faster but every other stage stays the same, the final output barely changes.
Second, acceleration casts its own shadow. When AI generates code faster, there's more code to review. When AI produces more content, there's more content to check. The work doesn't disappear — it changes shape. And that new work is rarely measured well.
Third, the baseline for comparison vanishes. "How long would this have taken without AI?" is a hard question to answer, because you can't compare before-and-after under identical conditions. The market shifts, the product shifts, the team composition shifts.
Fourth, productivity gains aren't permanent. AI tools deliver the most value early on, and the marginal returns shrink over time. Many teams see a stunning transformation in the first month, only to find that six months later, the cost-to-benefit ratio has gotten murkier.
Put together, these factors make ROI on AI investment one of the hardest things to measure in business today.
What "getting harder to justify" really means
Macdonald's phrasing is worth noting. He said "getting harder to justify" — not "impossible to justify." That distinction matters.
This isn't a decision to stop investing in AI. It means the bar for accountability around that investment is rising. Last year, budgets sailed through on the pressure of "fall behind if you don't adopt AI" alone. Starting this year, companies have to answer: "So what did we actually get from it?" If they can't answer, budgets get cut, use cases get restricted, and someone ends up holding the bag.
What this signals is that the "sales round" of AI adoption is over. Companies have moved from deciding whether to adopt AI to justifying the adoption they've already made. This round calls for hard numbers, not flashy demos.
Why smaller companies have the edge
This shift is a burden for big companies, but it can actually be an opportunity for small businesses and solo operators.
Large companies decide on AI adoption at the enterprise level. Budgets are large, use cases are sprawling, and ROI measurement gets complicated fast. At Uber's scale, simply tracing the causal link between token usage and user value is a massive undertaking on its own.
Small companies are different. A single operator paying for a Claude Code subscription can track, in their own head, exactly what feature that tool helped build and what revenue that feature generated. Because the unit of measurement is small, the cause and effect stays clear.
"Did revenue go up after using this tool for a month, or not?" A small company can answer that question within a month. A large company often can't answer the same question even on a quarterly or annual basis.
That asymmetry is small companies' unexpected advantage. The harder ROI measurement gets in the AI era, the more the edge shifts to organizations that can run fast, small-scale experiments and check the results quickly.
A practical way to measure ROI
For anyone trying to measure the ROI of AI investment in their own business or team, here's a practical approach.
1. Keep the unit of measurement small. Don't try to measure "the overall effect of AI investment" — measure "the effect of AI adoption in this one workflow." One customer-service automation, one type of content generation, one stage of code review. The narrower the scope, the more clearly the effect shows up.
2. Don't count only the direct costs. Beyond the subscription fee, include the time spent learning the tool, system integration costs, the extra time spent on review, and recovery costs if something goes wrong. Add all of that up to get the real cost.
3. Measure value through changes in user behavior. "30% improvement in internal efficiency" has no value on its own. What matters is whether that led users to get answers faster, return more often, or buy more.
4. Reassess every three months. The usefulness of an AI tool changes over time. A tool that looked great at first can look far less cost-effective six months later. Rather than adopting a tool once and using it forever, build the habit of periodically re-examining its value.
5. Consciously ask, "What if we did this without AI?" Throwing AI at every task isn't the answer. For some tasks, a person doing it directly is faster and more accurate. Consciously distinguishing where AI belongs and where it doesn't is the most direct way to raise ROI.
The start of the second round
Here's where the AI market stands right now.
Big Tech keeps pouring enormous capital into infrastructure. Nvidia alone has spent $40 billion on AI equity stakes this year, and OpenAI and Anthropic have teamed up with private equity firms to build multi-billion-dollar consulting arms. The tool-builders are still playing offense.
The tool-users are different. Even a giant like Uber has started asking, "Is this tool actually creating value for us?" More precisely: the sales pitch for adoption is over, and the justification phase has begun.
The friction between these two currents is the most interesting part. Tool companies push "use more," while user companies question "is what we're using actually worth it." That tension will shape AI market pricing, packaging, and differentiation strategy for the next several quarters.
The first round of AI adoption ran on excitement. Fear of falling behind was the strongest marketing pitch there was. That round is now winding down.
The second round is verification. "How did it actually go?" becomes the central question. The tools that survive this round will be the ones that can prove clear value, and the companies that survive will be the ones that can measure that value within their own business.
Uber burning through its annual AI budget in four months is a result of the first round. And Macdonald's admission that it's "getting harder to justify" is the opening signal of the second.
Now is a good time to check which of these two rounds your own business or team is standing in. Are you still being swept along by the excitement, or have you moved into the calmer work of verification? Answering that question honestly will determine competitiveness in the next stage of the AI era.
The more powerful these tools become, the more the judgment of the person deciding where to use them matters. The question Uber is asking right now is, ultimately, the same one every business is about to face.
