AI Economics Pulse: Your AI line item is winning, but is it working?

This edition of the Pulse is shifting lanes. We’re calling it the AI Economics Pulse now, because the question on every finance leader’s mind is whether AI spend and the returns on it can be made to pair at all.

That question came to a head over the last few weeks. The bills came due, and they came due in public.

Uber burned through its entire 2026 AI budget in four months and capped employee spending on Claude Code and Cursor at $1,500 a month. 

Uber wasn’t alone for long: the caps spread, with T-Mobile setting a $2,000-a-month per-user limit and Brex holding its engineers to $500 a week in tokens and everyone else to $5. 

Microsoft moved engineers off Claude Code onto its own Copilot CLI — framed as a tooling preference, but sources told The Verge the real reason was Claude Code’s rising cost. It’s probably both.

Walmart throttled employee access to its internal “Code Puppy” assistant when demand ran past plan. 

And one (unidentified) company reportedly ran up a $500 million Anthropic bill in a single month after forgetting to set spend limits.

Over-investors had the opposite problem. Amazon and Meta leaned in so hard their own people started gaming the metric. Amazon set targets requiring most of its developers to use AI tools weekly and tracked the results on internal leaderboards. 

Meta built one it called “Claudeonomics,” which ranked its 85,000-plus employees by token consumption, handing out titles like “Token Legend” to the heaviest users. So, to the surprise of very few, people “tokenmaxxed.” They ran junk prompts to pad their numbers, work be damned. The dashboard met its demise two days after it became public.

Two opposite reactions, same root cause: every one of these companies measured AI by what it consumes. None of them measured the outcomes. 

And cost, as it turns out, is the easy part.

When Uber’s COO explained the pullback, the real issue wasn’t the size of the bill. It was that it was “very hard to draw a line” from the spend to anything customers actually got (in other words, the outcome). 

That’s an attribution problem, not a budgeting one. 

Gartner sees the same thing coming: it expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, weak risk controls, and the one that should worry every finance leader: unclear business value. 

Companies aren’t overspending on AI; they just can’t tell you what it bought. Which raises the uncomfortable question: does AI even have a return?

The fight nobody’s refereeing

Right now there are three loud answers to that question, and all three come from people with something to sell (full disclosure: there’s a fourth one, and that’s us. We sell the fix.)

PR-firm CEO and tech critic Ed Zitron says AI simply doesn’t have ROI. You can’t even establish the cost of a unit of work, he argues, so the whole conversation rests on sand. 

SemiAnalysis CEO Dylan Patel and Head of Macroeconomics Malcolm Spittler say the opposite: the value is real, it’s just “dark output” that hasn’t surfaced in the numbers yet. 

Meanwhile, Nvidia CEO Jensen Huang says stop asking. Let a thousand flowers bloom, he says; it’s too early to demand a return. Though by June he’d flipped again, telling a Taipei audience the returns were suddenly “insanely profitable.” When the GPU salesman swings from “don’t ask” to “insanely profitable” in four months, that’s not analysis.

But look at who’s making each case. Zitron sells a newsletter built on AI skepticism. SemiAnalysis covers the chip industry. Huang runs the company selling the GPUs. The bull case and the bear case are both being argued by people whose livelihoods and reputations depend on the answer.

Zitron put the bear case bluntly on Bloomberg last week, and the part finance leaders should hear isn’t the IPO theater. It’s the measurement problem underneath it.

The headline-grabbing part is his market skepticism: Zitron warns that “people are conflating a semiconductor rally with an underlying successful business.” 

But the part finance leaders should actually sit with is quieter, and it’s about measurement. “You’ve got a thing where you can’t measure the costs and you can’t measure the return on investment,” he said. “What do you call that? You call it a thing without an ROI.”

He also made a prediction: “We’re going to increasingly see stories about companies implementing usage caps on AI tools for staff to manage costs.” 

We’ve already watched it play out. And when he explains why the caps are coming, he points straight at Uber, whose leadership, he says, is “having trouble justifying the AI spend based on the actual return that one could actually measure.”

Here’s where we part ways with Zitron, and the distinction is the whole point of this edition: Zitron looks at unmeasurable cost and unmeasurable return and concludes the return isn’t there.

We look at the same evidence and see a measurement gap. That’s not a verdict on AI, but rather a verdict on the instrumentation. The companies capping and cutting haven’t shown that AI has no return. They’ve shown they can’t see it. 

The data supports this: in a KPMG survey reported this month by the Wall Street Journal, only 26% of companies say they have a comprehensive view of their AI costs, while 22% have no visibility at all, or none until the bill lands. That’s a solvable problem for CloudZero.

The people actually holding the bill are less philosophical. Bain studied 951 companies and concluded, in its own words, that “the technology worked. The value didn’t arrive.” 

Sam Altman, asked on CNBC where the money was going, conceded “there’s a ton of waste” and guessed the industry would sort it out in a year or two. When the person selling the tokens can’t say where the money goes, this isn’t a vendor talking point. It’s the whole industry’s.

And none of them quite name it: this is a measurement problem in ROI clothing. You cannot prove a return you cannot measure. So before you pick a side in the bubble debate, look at what the spend is actually doing.

As in previous Pulses, our platform data tells the June story in three views: AI/ML spend, spend by service, and spend by provider. Let’s tackle them one at a time. Jump to the end for our methodology and how we measure the data.

View 1: AI/ML spend. The fastest-growing share of the enterprise cloud bill.

Across the CloudZero platform, AI and machine learning reached 6.67% of total cloud spend in May. That’s an all-time high, up from 2.50% in December, and more than quadruple the 1.60% of a year ago. AI’s share has now risen for eleven straight months, with each of the last nine setting a new record. (That’s the spend-weighted figure; you’ll see it in View 2, where AI/ML shows up as a service category.)

The chart above tells the more important story: what AI costs the typical company. When org-weighted, which is every customer counting equally, the median company spends 1.78% of its cloud bill on AI, while the average is 5.24%, nearly three times the median.

That’s where the real divergence begins. A small group of companies is going all-in on AI while most are still dipping a toe. If you’re sitting at the median, some of the firms you compete with are spending three, four, five times more. 

The instinct is to benchmark against those companies, which is the wrong approach. The one that matters: can any of you prove the spend is working yet? 

View 2: Spend by service. AI isn’t a line item, it’s a multiplier.

One layer down, by service, AI refuses to stay in its lane. 

Storage hit a record 11.94% in April and held near 11% in May, rising alongside AI. That’s no accident. Training sets, retrieval pipelines, vector stores, and the data exhaust of every model call all land in storage. 

Compute is still the largest category by a wide margin, but it’s losing share — down from roughly 51% at its 2025 peak to 46.5% as AI/ML, storage, and databases take ground. And AI/ML as a category has now passed Networking, Security, Analytics, and Management. A year ago it trailed most of them.

The lesson for finance: the AI line on your bill understates the AI in your bill. The rest is scattered across storage, databases, and compute, and you won’t find it by reading service categories. You find it by mapping spend to the feature, model, or workflow that caused it. 

Everything else is guessing, and finance leaders don’t like guesswork at all.

View 3: Spend by provider. Your AI spend is splitting across vendors

At the provider level, the hyperscalers look stable. AWS sits around 67% of spend, Azure around 9%, GCP around 7.5%. Steady as ever.

The movement is underneath them. AWS Marketplace roughly tripled its share of spend, from 1.79% to 4.67%. Marketplace is exactly where AI tools and third-party software get bought, frequently outside the normal procurement and finance review. It’s a blind spot, and it’s widening fast.

The model vendors tell the diversification story most clearly. A year ago, direct spend between OpenAI and Anthropic was roughly 99.8% OpenAI. Today Anthropic holds about a quarter of that split. Enterprises are spreading their bets across models. Smart move — and a measurement headache. The same AI feature now runs across multiple providers, each with its own unit economics, which means cost-to-serve no longer lives in one place. 

The spend is measurable. The return has to be too.

Put the three views together and the contradiction resolves. The spend is real and accelerating. We can see that to the decimal. No one can prove the return for one reason: almost no one measures AI where return actually lives, down at the level of the individual customer or feature. 

Token caps don’t fix that. Leaderboards make it worse. Runtime guardrails govern the infrastructure. Cloud spend dashboards report the ops bill, but none of those answers the question a CFO takes to the board: what did this customer, this feature, this agent cost us, and was it worth it?

Billing shows what you spent. Telemetry shows why. Join them at your own dimensions such as customer, product, feature, or agent, and AI spend stops being a number you defend and becomes one you direct. We call it cost-per-anything, and it runs on telemetry-anchored, dimensional allocation rather than the tag-and-account guesswork most teams are stuck with.

It works in practice. Duolingo, a CloudZero customer, broke its text-to-speech spend down by feature and found one feature missing a cache that a similar one already had. Closing that gap cut TTS costs 40%. Same spend data everyone has. A sharper question asked of it.

What this means for June

Your board is already asking. The finance leaders who can answer “what did AI return?” will keep their budgets and grow them with confidence. The ones who can’t will be the names in next year’s version of this story. They are capping, cutting, canceling. Not because AI failed, but because they could never see what it was doing.

The spend is the easy part. The return is the job. And it’s measurable, if you’re willing to measure it where it actually lives.

How we’re looking at data (and why it matters)

For the AI Economics Pulse, we track monthly cloud spend trends using anonymized, aggregated data from CloudZero’s network.

Together, these views show not just where AI and cloud spend goes, but how spending patterns shift as new technologies — especially AI — reshape the cost landscape. One note on methodology: You may find that the AI/ML percentage in the service category section differs from the average in the dedicated AI/ML section. Both are correct; just measured differently. The service category is spend-weighted; essentially, total AI/ML spend divided by total cloud spend across all customers. Meanwhile, the average/median figures are org-weighted — every customer counts equally regardless of size.

Thoughts, comments, disagreements? Reply to this Pulse or email [email protected] with “AI Pulse” in the subject heading. We’ll feature the best feedback in an upcoming issue. Watch for our next AI Economics Pulse on July 21, 2026.

Author Spotlight

Keith MacKenzie

Keith is CloudZero's Content Marketing Manager with a specialty in topics around AI & cloud spend. He also brings more than a decade's worth of background as an editor and writer in the mainstream media and content marketing industries. He's also been to Chernobyl twice – and (sometimes) still has all his marbles.