AI is the Industrial Utility of Today

This week, we heard from multiple prospects all looking for one thing: track and control what AI costs, down to tokens, users, and projects, and put everything else on the maybe-later pile.

A year ago, those same conversations would have been a mix of offboarding, license use monitoring, and NHI detection. But today, it’s all AI.

One CTO said that tracking their monthly LLM costs was like “open brain surgery.” Another put it more simply: they had been early adopters of AI, and now those bills were becoming a massive liability.

AI is not a feature, it’s a utility

AI consumption is the only number that matters, and it’s also the only number nobody can see.

Software used to be a fixed cost. You negotiated it once a year, signed, and the number sat still until renewal. You bought 500 seats, used 180, and argued about the difference twelve months later. Painful, but predictable. The decision that set the bill was made at signature.

A utility doesn’t work that way. You don’t buy it, you draw on it. It meters by consumption, it scales with usage rather than headcount, and it runs whether or not anyone is watching. The bill isn’t decided at signature. It’s decided every day, by behaviour, by thousands of small choices nobody logged.

That’s the shift. The buying unit used to be seats. Now it’s tokens. And the control point moved with it: procurement governed software, but metering is the only thing that governs a utility. You cannot negotiate your way out of a consumption curve.

Which is the problem, because a monthly bill for tokens tells you almost nothing. You don’t know who the heavy users are. You don’t know which teams or use cases are driving it. You don’t know whether the spend was worth it. You don’t even know what next month looks like. You just know the bill is growing.

The steam engine was never the expensive part

Nobody remembers what a steam engine cost. Ask an industrialist in 1870 about the price of the engine and you’d get a shrug. The engine was a one-time line item that stopped mattering the day it was installed. The fight that decided who won was over coal. Priced per ton, burned forever.

And the operators who pulled ahead weren’t the ones who negotiated the cheapest ton. They were the ones who learned to read what a ton of coal actually produced, output per ton, before their competitors could read it, and before the instruments that measured it were standard equipment.

That’s the part worth borrowing. Not “tokens are the new coal,” which is a fine line and a taken one. Palantir’s Shyam Sankar said it almost verbatim on the Q1 2026 earnings call, and every FinOps vendor has said it since. The narrower lesson is the one that mattered in 1870: the price of the input is the easy half, and the operators who win are the ones who read the output half first.

Tokens are further along that curve than coal ever was on day one, which is exactly the trap. The input meter is already precise. You can look up published list rates to the million tokens, Claude Sonnet 5 at $2 in and $10 out, Opus 5 at $5 in and $25 out. Compared to the old seat math, a per-token price reads clean. It reads like the problem got simpler.

It didn’t. It moved.

The input meter is nearly built. The output meter isn’t.

Two questions get blurred together, and the difference between them is the whole thing.

The first is what did we spend and who spent it. That’s cost attribution, and it’s getting solved. Not finished, but close enough that you can see the finish line from here. FinOps Foundation’s State of FinOps 2026 found that 98% of practitioners now manage AI spend, up from 31% two years ago. The same report names granular AI-spend monitoring, tokens and requests and GPU, as the top-cited missing capability.

Read that carefully, because it’s easy to misuse. The gap they’re describing is finer metering on the cost side. Even the input meter isn’t fully installed yet. That tells you the easy half is nearly done, not that the hard half has started.

The second question is what did the spend buy. And that one has no meter at all.

It shows up the moment the CFO question changes, which it already has. It used to be “what are we spending on AI.” Now it’s “what did we get for it,” and the answer is mostly silence: only about 7% of senior leaders report established AI ROI. Ninety-three out of a hundred can increasingly tell you the bill and cannot tell you the return.

That gap isn’t a pricing problem. It’s a measurement problem, and the instrument for it hasn’t been built.

The spend moves fast enough to matter

Look at the spread rather than the average. Ramp’s 2026 business spend data puts median monthly AI spend at $2,246 and average monthly AI spend at $140,842.

Sit with that for a second. The median company spends about the price of a decent laptop each month. The mean is dragged two orders of magnitude higher by companies running orchestrated, agentic workloads. That spread is the entire risk profile of the category. One unchecked agentic workflow is how you travel from the median to the mean without anyone filing a purchase order.

The mix underneath it is moving the same direction. On the same Ramp data, premium models account for 45.8% of tokens but 55.9% of AI cost, up from a rounding error a year earlier. And the cost of a single interaction has climbed with the shape of the work: roughly $0.04 for a single 2023 linear call, roughly $1.20 for a 2026 orchestrated agentic task. About 30x.

The mechanism isn’t exotic. Output tokens cost exactly five times input tokens across the entire Claude lineup, and an agentic task consumes 5 to 30 times the tokens of a single chatbot turn. A tool that talks back and forth to itself twenty times before answering is not twenty questions. It’s one question with a bill that fans out.

Gartner has put a stake in the ground on where this heads. In June 2026 it predicted AI coding token costs will surpass the average developer’s salary by 2028. Two months later it went further, predicting AI inference costs per agentic workflow will rise more than fivefold through 2028, with a warning worth keeping in one line: product leaders cannot rely on more efficient token economics to rationalise AI costs.

The cheap tokens get cheaper. The expensive work gets more expensive and less predictable. The average hides both.

Two instruments, one wall

The input meter reads price times volume. It’s nearly finished, it’s improving, and it’s the half every vendor is racing to perfect.

The output meter reads what the spend produced. Per token, per user, per project, across every AI tool actually in use. That instrument doesn’t exist as a product yet.

And the place it’s darkest is the place IT never approved. The AI tools people adopted on their own, the shadow AI stack, where the spend is real and the attribution is nothing.

That’s the ground worth naming, because it’s the ground the current tools don’t cover. Cloud FinOps platforms like CloudZero and Vantage do genuinely deep token attribution, but they face engineering and infrastructure and stop at the infrastructure layer rather than spanning the application layer or the tools nobody filed a ticket for. SaaS management peers work the other side of the house, the procurement and approved-spend layer. Zylo’s AI Consumption Cost Management shipped in 2026, alongside Vertice and Productiv. Flexera calls itself the only complete AI cost management platform, and it’s infrastructure-first.

Line them up and the coverage map has a hole in the exact shape of the question buyers are walking in with: token and user and project, across many AI tools, including the ones nobody approved, read from an IT or ITAM seat. That’s where an output meter would have to live, and it’s where none of them live yet.

So what did the token buy?

Nobody can tell you yet. That’s the point.

Every AI tool can already tell you what a token cost. None of them can tell you what it bought, and the second number is the only one that was ever worth having.

The industrial analogy makes one more prediction, and I don’t have the answer to it. Coal got its output metrics eventually. Thermal efficiency, cost per horsepower-hour, all of it became standard, and the edge that came from reading output first went away once everyone could read it.

So the real question isn’t whether someone builds the output meter for tokens across the whole stack, shadow AI included. Someone will. It’s whether the years before they do are long enough, and the edge wide enough, that reading output first is worth as much now as it was for the operators who learned to read a ton of coal before their rivals could.

I don’t know. But the people who dropped everything else to ask are betting it is.