CloudZero had a full team on the ground at Ai4 in Las Vegas during the first week of August 2026. The team included CTO Erik Peterson, who spoke on a panel about AI cost economics. The same problem surfaced everywhere we went: teams can see what they’re spending, but not whether it’s working. DIY cost tooling that fails time and time again, agent sprawl, and a widening gap between finance and engineering kept coming up throughout the week.
Ai4 draws the kind of Vegas crowd that fills a keynote hall for a fireside chat on token pricing. There was time for exactly two audience questions on Erik’s panel. One of them is still worth answering properly.
The question from the floor
“Say you pick a vendor, they drop an agent into a workflow, and your AI infrastructure cost jumps 10 to 20x. How do you approach that change?”
Erik’s answer, compressed by the clock, still holds up: Separate the build-versus-buy decision from the cost decision. Get honest about data hygiene, since that’s usually where the real multiplier hides. Figure out who’s actually doing the front-end analysis on the data you’re paying to generate.
Name who’s accountable when the number is wrong, too, because someone always is, and it’s rarely whoever picked the vendor.
The panel ran out of time before Erik could answer the harder half of the same question: how do you know your own team’s token usage is actually worth what it costs? He and the person who asked connected after the session. That question, of whether this spend is paying off and not just how much it is, is the one that followed CloudZero around the show floor all week.
Rachel Maness, an attendee we caught on camera, put it almost the same way unprompted: “You can prove that it makes you more productive, it makes you more efficient, but what are you doing with that additional productivity and efficiency? How is it actually impacting the bottom line?”
Report
Finance needs to prove AI’s return: CloudZero report
260 senior finance leaders (more than half CFOs) told us why the speed of seeing AI spend, not the size of it, separates who pulls ahead on AI from who gets burned.
Every decision is now a buying decision
Erik opened his side of the panel with a simple test: “What is it that we do around here? What does it cost to actually do that? We build homes. Okay, what does it cost to do that, right? We process these conversations. Okay, what does it cost to do that?”
The harder follow-up, he said, is whether “those costs at a unit level [are] going down over time, or are they becoming more expensive because we’re making the systems more complex along the way.” His reason the discipline never lets up: “All software is a lifetime commitment. It’s never finished, it’s abandoned.”
On stage, Erik kept coming back to one idea: unit economics, not total spend. “You can’t think about total spend,” he said. “You need to think about the cost per unit of value.”
Asked what breaks when cloud-era instincts meet generative AI, his answer cuts against a decade of traditional FinOps habit. “I could pursue 100% optimization,” Erik said. “I could have my entire estate of all the compute, all the cloud resources I have, at 100% max utilization and it could be all waste, right? It has to be tied to value. It has to be tied to a business outcome. That’s the big shift.”
He added that the old playbook doesn’t transfer: “You have to combine the costs with the usage… CloudZero thinks about this in terms of how much money you’re spending and then what it is being used for. That’s telemetry that drives that.”
The shift he described isn’t really about AI. It’s about who’s spending.
“It used to be in the cloud era, every engineering decision was a buying decision,” he said. “In the AI era, it’s really every decision is a buying decision. You sit down to write an email, you’re probably spending a few pennies.”
He noted that even CloudZero’s own head of people and culture asked for a GitHub account this year.
“What it means to build software has completely changed in its entirety,” Erik said. Everyone’s an engineer now, in the sense that matters to a finance team: everyone’s spending.
As he put it, the tradeoff cuts both ways: “what you’re able to accomplish often exceeds what you could have dreamed of doing in years. But you have to measure that against the outcome that you’re shooting for and use unit economics.”
That instinct wasn’t confined to the CloudZero booth. Fiona Victoria, who comes from an AI research background, told us the same expectation is spreading into teams that never had to think about cost before: “Now even engineering teams need to be aware… of what certain things cost. Can they actually compare different solutions and do more cost optimizations too?”
Erik traced the fix back to CloudZero’s own founding problem, a decade before anyone was talking about AI cost: the company needed to forecast what a system would cost before it had a single customer on it, so it built the forecast off a unit cost times an expected volume instead of an infrastructure budget.
“That was wildly more accurate than anything else,” he said. It’s the same math he’d apply to a company that spent $100,000 on AI last year and $10 million this year: “the pace of change exceeds our ability to observe it,” unless the forecast is built on a unit instead of a guess.
When the optimization costs more than the thing it optimized
Erik told a story worth repeating whole. CloudZero built a system to automatically pick the cheapest capable model for a given task, calling out to check prices and run the math before every request.
“On a single request scale, well, that was really efficient,” he said. “But when we actually started to use it, we realized the whole calculation around understanding which model to use cost more than just using a model.”
The lesson he draws from it is the same one that came up at every booth this week: “It’s very easy to start these experiments and you don’t think about the math, the multiplicative effect of that… I found a thing that costs 20 cents, but how often do I expect that to be used? Oh, it’s going to be called a million times. Well, now it starts to become expensive.”
Those kinds of things, he said, “are easy to miss if you’re not looking at the individual total cost of that transaction.”
What the booth heard, six times over
Walk a booth for four days and patterns start repeating word for word. Six patterns surfaced this week at Ai4 at the CloudZero booth.
The most common conversation started the same way every time. Someone had built their own AI cost dashboard, coded a rightsizing model internally, and gotten partway to visibility; and then they admitted, often in the same breath, that they couldn’t show ROI with it and would rather just buy something.
The lesson learned: DIY AI cost tooling isn’t failing because engineers can’t build it. It’s failing because the dashboard was always the easy 80%.
The harder question came up just as often, and nobody had a clean answer: is our AI spend actually worth it? Not how big the bill is, but whether it’s paying off. That’s a sharper question than cost-cutting, and it surfaced whether the person asking sat in finance or engineering.
More than one team described genuine zero visibility before the invoice lands: no forecast, no warning, just a number at the end of the month. One group had only just formed a working group to deal with it.
Agent sprawl showed up as its own version of the same problem. One prospect even mentioned, almost in passing, that they’d have 450 agents running by year end, with no plan yet for tracking what happens to cost and accountability once agents multiply faster than anyone’s counting.
Increasingly, the AI bill and the cloud bill are the same conversation. Multi-cloud sprawl into GCP and Azure came up constantly alongside AI cost, because treating the two as separate line items is exactly how finance loses the thread. Underneath all of it sits an accountability gap: when nobody owns the number, “not technical, so wasn’t sure if it was useful” becomes the default answer to “is this working?”. That’s not a technology problem. It’s an ownership one.
None of this should come as a surprise if you’ve read CloudZero’s own research. Finding the ROI of AI: The Finance Perspective, CloudZero’s June 2026 survey of 260 finance leaders, found 87% feel pressure to tie AI spend to results within the next year, and only 22% can do it today. Ai4 was that stat with a face on it, booth after booth.
Everything here, whether it’s the question from the floor, the booth confession that started to feel scripted by the third or fourth repeat, or Erik’s own story about optimizing away the savings he was chasing, all loops back to the same gap: Measuring AI spend is no longer the hard part. Most of the floor at Ai4 either has a dashboard or is building one.
But proving what that spend is worth is the part few have cracked, and CloudZero is one of them: unit economics is the discipline we’ve been building since before “AI cost” was a category, and it’s why we can tell customers not just what they spent, but whether it worked.
If the gap between what you’re spending and what it’s worth sounds familiar, see where you land against the 260 finance leaders from our survey with CloudZero’s AI Benchmark Tool.
