CloudZero’s own AI Economics Pulse for September found the 75th percentile of its 430-company customer panel crossed 10% of its cloud bill on AI for the first time in August.
Gartner’s latest survey found only 22% of organizations have scaled AI successfully and 11% don’t know what their own function spent on it last year.
CJ Gustafson showed what that gap looks like on an actual invoice this week. The throughline: spend is outrunning the ability to explain it, at every altitude from a single bill to a public earnings call.
Every CFO who added a generative AI line to the P&L this year has run into the same problem: the cost model doesn’t behave like anything finance has managed before.
Gustafson, who writes the newsletter Mostly Metrics, captured the disorientation from the planning side this week. Budgeting for AI, he wrote in the latest “Annual Planning Bible” edition, means “embarking on a world of continuous forecasting derived from a foundation of disparate data sources… something that was never possible before. It’s scary, and it’s beautiful.”
Scary and beautiful is one way to put it. The blunter version, from the same edition: “For the first time in a long time, Finance feels like they don’t have the answers.” That’s the mood behind every stat in this roundup: a bill that never tells you anything useful.
The bill nobody can explain
The provider bills for what it can meter (tokens by model, by day, etc.) because that’s what its own system tracks. It’s a usage log that only looks like a cost report.
However, it has no way to know, and no reason to say, which of those tokens produced a shipped feature and which ones were an old workflow nobody updated after a cheaper model came out. That distinction is the entire job of finance, and it’s a distinction the invoice was never built to make. Every AI bill arrives pre-stripped of the one thing a CFO actually needs from it.
That gap is exactly what shows up when someone finally sits down and reconciles the bill against the usage behind it.
Gustafson did that this week, walking through a sample invoice for a fictional company he calls Northwind. The bill: $61,656 for the month, four lines, no detail on what drove it. He says he’s heard the same complaint from “multiple CFOs and FP&A pros”: “I have no idea what I’m looking at when I get an Anthropic invoice. I just know it’s going up.”
Once he reconciled the usage export against the invoice, the waste sorted into a handful of familiar shapes: stale model choices, unmanaged caching, mispriced processing speed, and spend concentrated in a few jobs nobody had scrutinized.
None of that is new. The same cost-management failure modes predate generative AI by decades. What’s new is the cadence: a model lineup that changes monthly, a usage mix that shifts week to week, and a bill that only shows up once the quarter has already absorbed the waste.
So, the current challenge that has CFOs up against the wall is catching the spend before the invoice arrives. Not after.
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.
What CloudZero’s own panel shows
CloudZero’s own data, published September 8 in the monthly AI Economics Pulse, shows how fast that gap is widening. Across a same-store panel of 430 customer organizations, the median company’s AI spend reached 2.66% of its total cloud bill in August, up from 0.67% a year earlier. The top of the distribution is pulling away even faster: the 75th percentile crossed 10% of the cloud bill for the first time, hitting 11.12%, up from 9.63% in July.
The share of the panel with AI at 10% or more of cloud spend jumped to 28.2% from 23.9% in a single month. But the one tier that fell was the top one: the share of companies at 25% or more of cloud spend on AI dropped to 8.2% from 9.1%, its first decline since January.
CloudZero’s own Pulse offers two explanations, and can’t yet tell which one is doing the work: either the biggest spenders are getting questioned on their AI budgets and pulling back, or their overall cloud spend is simply growing faster than their AI spend and diluting the ratio. A company can drop out of the 25% tier either way.
The measurement gap, by the numbers
The panel data shows the symptom: spend climbing unevenly. Gartner’s own research shows the cause: finance can’t see what any of it is for. Surveying 1,303 organizations in a study published September 1, Gartner found only 22% have successfully scaled AI across multiple business units or adopted an AI-first approach, and roughly 11% are entirely unaware of what their own function spent on AI in 2025.
Investment isn’t the problem: 85% of functional leaders plan to increase AI spending in 2026, after already committing an average of 12% of their functional budgets to it last year.
“This lack of financial visibility heightens risk as spending accelerates,” said Tina Nunno, Distinguished Vice President & Gartner Fellow. “Without disciplined measurement tied directly to business outcomes, organizations risk wasted resources and unmet expectations.”
The same survey also found a wide performance split tied directly to that discipline: organizations that constantly track AI ROI and reallocate away from underperforming projects reported positive returns on 81% of their initiatives, while low performers couldn’t even say what the return was on 29% of theirs.
CloudZero’s own research from June 2026 found the same gap from the finance perspective. In a survey of 260 finance leaders, 87% said they need to tie AI spend to business outcomes within the year, and only 22% can do it today. That same survey found three out of five finance leaders admitting they’re already spending more on AI than they can justify, and two-thirds of boards are conditioning further AI funding on proof of return.
While AI initiatives are driven outside of the finance function at 74% of companies, finance ends up owning the spend tracking in 60% of them; the same mismatch Nunno described, on finance’s side of the ledger.
SAP’s own controllership team ran into this directly and wrote about the fix earlier this month. After assigning token costs to the business areas actually generating them, the company reported the conversation shifted “from ‘How much are we spending?’ to ‘What are we getting for this?'”.
Two companies, same spend, opposite verdicts
The same gap shows up at three different altitudes this summer: what the market pays for spend it can’t yet judge, what Gartner sees across the whole industry, and what Ramp sees inside individual companies.
Start with the market’s verdict: Alphabet and Amazon both reported this summer, and got opposite reactions.
Alphabet’s Google Cloud revenue grew 82% to $24.77 billion in the second quarter. Capex guidance went up anyway, from $180–190 billion to $195–205 billion, and free cash flow turned negative for the first time in the company’s public history.
The reaction? The stock closed down 7.13% the next trading day. Some of that gap is being financed with debt; TMT Finance put it at roughly $100 billion, up from about $16 billion a year earlier. Plus, an $80 billion equity raise that included a $10 billion private placement with Berkshire Hathaway.
Eight days later, Amazon posted AWS revenue up 37% to $42.2 billion, its fastest growth in 18 quarters, on its own capex increase. Yet, the market didn’t flinch.
Amazon CEO Andy Jassy has a case for increased capex guidance: data centers “generate revenue for 30-plus years” once built, and servers and networking equipment “break even in less than three years” against a five-to-six-year useful life. Hard to argue against that kind of real value versus invoices.
The same skew shows up further down the market. In July, Gartner forecast worldwide IT spending to hit a staggering $6.37 trillion in 2026, up 14.2% from the year before, with data centers and IaaS core to that growth.
“Building the compute capacity required for AI is the largest infrastructure project ever attempted by humanity,” said John-David Lovelock, Distinguished VP Analyst at Gartner.
Lovelock was clear that this isn’t a uniform windfall: “Technology budgets are being strained by inflation, supply shortages, rising hardware and memory costs, AI funding initiatives and shifting priorities.”
At the usage layer, the picture is fragmenting rather than simply growing: TechCrunch reported this week that Ramp’s spending data across 70,000 companies showed AI spend per employee at the top 1% of AI-using firms falling nearly 10% to $7,205 in August, as average token costs dropped to $0.68 per million from a 2026 peak of $1.15 in March.
Aggregate spend climbing, per-unit costs falling, usage concentrating hard at the top: a finance team that can’t yet tell which of its own AI dollars is doing which of those three things is flying blind through a market moving on all three axes at once.
What a CFO does this quarter
So, what’s a CFO to do? Near-term, this starts with ownership. Assign a name to every denomination of token spend, and pair the cost figure with a value metric before cutting anything.
Speed matters just as much if not more. CloudZero found that finance teams who see AI spend within a day of it occurring are more than twice as likely to hold an “invest aggressively” posture, and far less likely to freeze investment out of uncertainty.
There’s a bigger allocation question sitting underneath all of this, and it’s structural rather than a reporting-cadence problem: tying AI spend to the outcome it produced requires attributing shared, variable costs across coding agents, content generation, and customer workflows that all draw on the same model subscription, which no vendor bill is built to do on its own.
CloudZero’s own Pulse data and Gustafson’s invoice scenario both surface this gap, albeit from two different directions.
For finance leaders wondering where their own organization’s AI spend governance lands next to the rest of the market, CloudZero’s AI Benchmark Tool walks through five questions on how AI initiatives get driven, tracked, and funded, and shows where you land against the same 260-leader survey.
