The Meter

The meter is what you build when you cannot see the output.

Model labs want you to believe that tokenmaxxing, the regime where tokens consumed stand in for intelligence deployed, is the only way to scale intelligence inside an enterprise, and that charging for tokens is only fair because of the enormous surplus those tokens generate inside your business. The pitch sounds reasonable until you ask a boring accounting question. Where, exactly, does that surplus show up?

Nobody can answer it. Every tool we use to measure economic throughput was defined at the level of an individual firm. EBITDA, gross margin, revenue per FTE, etc. None of them were built for the wedlock that now exists between an F500 enterprise, its employees, its software, and unfettered intelligence courtesy of a model lab. There is no line item for surplus generated jointly with a foundation model, so when a lab says it captures only a fraction of the value it creates, nobody can compute that fraction, not the CFO and not the lab. A number that flatters whoever announced it is called a pitch.

That gap is the whole story. When the value created across a boundary cannot be measured, the only thing left to price is the input flowing across that boundary. The meter is not a pricing strategy. The meter is what you build when you cannot see the output.

A lab that cannot participate in the upside of applied intelligence has two options. It can vertically integrate and shut off API access to third parties, or it can meter the input at the one boundary it controls. The labs chose the meter, and the choice was not innocent. The prophecy behind it is self-fulfilling, created by the Silicon Valley elite to serve their own cause. If all economic output eventually routes through metered intelligence, the market cap of the model labs becomes an index on global GDP, with regulatory capture in Washington, potential federal bailouts included, capping the downside. The catch is that the index only pays out if the meter comes with pricing power. Electricity routes through everything, too, and utilities trade at ten times earnings.

Metering inputs is not inherently dumb. Electricity is priced per kilowatt-hour, and civilization runs fine. But physics fixes how many kilowatt-hours a job needs, and your utility cannot redesign your refrigerator to draw more power. Nothing fixes how many tokens a job needs, and the lab ships both the agent and the meter. Token revenue measures activity, not value, and the seller holds the dial on the activity. Even the flat-rate seat is a meter in disguise, priced off expected tokens and capped so that the heaviest users overflow back onto per-token pricing.

The meter also explains what the labs will never build for you, which is the real transformation of your business. Real transformation would require AI to become the least sexy part of the business, and it runs on exactly the slow, unglamorous work no meter rewards. The tool-shaped objects the labs ship instead (Claude Code, Codex) are built to maximize consumption of the underlying unit, whoever ends up footing the bill, and token growth is the lever holding up an extremely generous multiple on lab enterprise value. Nothing in that structure pays the lab to transform you.

And the meter is not even safe for its owners. A meter only earns when tokens flow, so the labs need ubiquity, and ubiquity invites the clones. Open-source weights and small models have made the model call behind most enterprise tasks a commodity. These are the tasks simple enough to specify completely, low K complexity to give it a name, and they are 99% of what an enterprise asks of a model (nobody can measure that number either, which is rather the point). I would not even call that band an oligopoly anymore. The fire that warms you also burns you. Tokenmaxxing built the lab valuations, and it fostered a world where even the smallest competitors are equally or more competitive than the labs in most enterprise deals. So the labs are left choosing between an arbitrary utopia and a knife fight, either declare AGI and sell the promise that the model will do science itself, discoveries you can patent, or wade into the enterprise and compete with a thousand app-layer rodents who know the customer’s process better than the lab ever will.

So the meter stays profitable only while two things hold. Nobody can measure the surplus, and nobody undercuts the meter from below. The second is already failing in public. The first is more durable and stranger, because the strangest thing the meter ever did was convince its customers to manage themselves by it.

The Market

Manage by your own convexity instead of the lab’s meter.

Token metering exists because nobody can measure the value AI creates across a firm’s boundary, so the labs price the input instead, and token revenue ends up measuring activity rather than value. The stranger consequence is that the firm has started managing itself by the lab’s meter.

Becoming an AI-native firm has become memetic. You restructure the org (read fire employees), you wire a frontier API through every workflow, and you report tokens consumed to the board as evidence of transformation, right up until the token bill becomes unbearable and you quietly hire people back. Tokens consumed is the lab’s revenue line, repackaged as your productivity metric. You are reading your transformation off somebody else’s invoice.

It gets worse because the party that printed the invoice is in the business of hollowing you out. Meter your inputs, make tokenmaxxing the measure of productivity, drain your decision traces, sell the frontier model back to you at a markup. The contracts now mostly promise that your prompts will never train the next model, and the promise is mostly kept. The lab learns from the aggregate anyway, which workflows the economy routes through it, where the failures cluster, which evals matter, and the next model arrives fitted to your industry, whether or not it ever saw your data. You pay in cash, and then again in exhaustion and dependence.

The meter corrodes the people watching it, too. Tokenmaxxing has induced a state of bravado, where the five-figure monthly bill gets screenshotted like it proves something. The enterprise is filling with slop, drafted by one model and summarised by another, while everyone in between reports productivity gains. We have run this loop with vanity metrics before. Page views, then GMV, then ARR at any CAC. The metric stops describing the work and eventually replaces it, and the correction is the same as it was then. Touch the grass. Spend your time doing something scarce.

So what should a firm actually do with this technology? Two fashionable answers get in the way.

The first is to build a tech team and go deep. Post-train models, build harnesses, maintain evals, measure drift. Take a law firm, where the case for this is supposed to be strongest. The red queen race between the labs already makes marrying one model foolish within a quarter. Worse, all of that infrastructure work is commodifying. Every rival firm can buy the same models and run the same evals, and when both lawyers use AI, the advantage moves to everything non-AI in winning a case, which is to say the relationships, the judgment about which matters to take, and the willingness to stand accountable for an outcome. A law firm’s incentive is to figure out a business model that turns this unprecedented technology into a durable advantage for itself and its clients, and that work happens in pricing and trust, not in MLOps.

The second fashionable answer is rollup fatalism, the quiet belief that none of this matters because the firm itself is doomed. If generic intelligence plus scale beats encoded human specialization at every job (the bitter lesson applied to the economy), then the economy is one giant rollup waiting to happen, and every firm is a small speck inside it, so you may as well cut costs and wait. But the rollup needs a second condition: somebody has to own the intelligence, and that condition is failing in public. The input is commoditizing. A capability everyone can buy is a tide that lifts no one in particular.

But when the tide reaches the coasean boundary of the firm, it encounters something it cannot reach.

Take procurement of any good or service for an example (a super large market opportunity with all the FUD, full of Hayekian forces at play). The prices there are negotiated, local, relationship-dependent, and perishable. The knowledge that wins a deal does not exist until the negotiation creates it, and the people who hold it reveal it selectively, to counterparties they trust, at the moment it matters. You cannot train on what nobody will say to a machine, and by the time it has been said, the price has moved. A meter cannot see this knowledge. Neither, for what it’s worth, can the firm’s own ERP. This knowledge is the firm’s reason to exist, and it lives in exactly the people the restructuring memo fires first.

Which brings me to the answer I believe. When AI automates away the manual drudgery of being a ‘procurement/pricing specialist’ what does the firm do with a pricing specialist that now has 4x more capacity per hour? The lazy answer says you now need 4x fewer pricing specialists. But in nearly every procurement function I have looked at (a biased sample), the returns on sales and pricing capacity are convex. A good hit rate gets you the relationship, and the relationship eventually gets you better rates. That loop is the convexity, and the law firm runs the same curve on its relationships and its judgment. I don’t claim the curve rises forever; every loop saturates somewhere. I claim almost no firm I have met is anywhere near the top of it. A firm that frees up specialist capacity and throws the specialists out is selling a convex asset to save a linear cost. This is why enterprise AI keeps moving the needle only a few percentage points. The gains land, the firm cashes them out by cutting the people whose hours compound, and the compounding never shows up. The alpha is in repurposing the freed capacity toward the convex work, which in most firms means pointing it at sales (read signal collectors).

The reason almost nobody does this is that the firm’s own dashboard punishes it. Revenue per FTE is an average, and averages are silent about margins. Wherever relationships compound, the marginal hour is the most productive one, so optimizing the average means amputating the margin. And the trap is that firing people improves the ratio all the way down. The firm doing the right thing looks worse on this metric, quarter after quarter, than the firm bleeding out.

It comes down to one substitution. Manage by your own convexity instead of the lab’s meter. The problem is that no metric you currently run can see whether the redeployment happened, and someone still has to build that instrument.

The Instrument

Stop selling intelligence and start selling the measurement that the labs cannot do.

The labs measure tokens because nobody can measure the value AI creates across a firm’s boundary, and firms, managing themselves by that same meter, fire the very people their returns depend on instead of redeploying them toward the relationship work where returns compound. The layer in between is the application layer, where I work, so the incentive problems from here on are mine.

Selling intelligence cannot be the answer, because the commoditization race rolls downhill, and the app layer is standing in its path. The labs tried to commoditize the app layer. Open weights commoditized the labs instead. And the app layer keeps building smaller and smaller models to protect its own unit economics. Follow the race to its end, and the app layer’s last complement left to commoditize is the human at the edge.

There, the race stalls. The human at the edge is, in a lot of cases, just difficult to commoditize, even when the work is low K complexity and a commodity model handles the task fine. The task was always cheap to describe. The human was never just the task. They hold the dark context that never made it into any system, and they own the relationship. A customer can call them at 11 pm and yell. An app vendor that automates these people out is not finishing the race; it is killing the customer it feeds on (an autoimmune disease)

The durable business is the opposite one. Stop selling intelligence and start selling the measurement that the labs cannot do. The labs meter tokens because nobody can see outcomes across the firm boundary, but an app vendor sits inside the workflow, at the boundary itself, where it can anchor its price to numbers the customer already runs the business on. Quotes sent, response times, win rates, jobs closed. Audited how? Those numbers predate the vendor, so there is a baseline. Hold out a desk, a trade lane, a quarter, and the counterfactual stops being rhetorical. The customer can compute every one of them without the vendor in the room. That is the difference between a measurement and a pitch, and it is the answer to my pricing question. You charge for the percentage points by being the instrument that proves where they came from.

This is the slow, unglamorous work, no lab meter rewards: industry process knowledge, genuine product adoption that compounds the app layer efficacy, purpose-built harness that holds up in production, and a business model that runs without subsidies. Legibilize the process context, the part of the firm’s dark knowledge that can survive being written down, and leave the perishable market knowledge where it lives, with the humans you repurpose. The traces the harness produces stay the customer’s property, the opposite of the lab’s exhaust economics, where everything you reveal helps fit the next model someone sells back to you. And the outcome ledger doubles as the missing metric, the first dashboard that can show a firm whether its freed hours actually became convex work. The data in it is per-customer and per-relationship, which is the one kind of moat that symmetric AI adoption cannot wash away.

The meters will keep spinning either way, and consumption will keep getting reported as transformation. Which leaves the boring accounting question. Where does the surplus show up? The labs built a meter because they could not answer it. The firms fired the wrong people because their dashboards could not answer it either. Whoever builds the instrument that answers it gets to write the prices.