The Meter’s Running

THE NUMBER: $50 → $3,000 — what a metered AI seat cost a month ago versus what it can cost now. Annualize the new number and you get $36,000 a year, and remember what that price doesn’t carry: no payroll tax, no health plan, no PTO, no chair, no manager. If that bill is doing an employee’s job, the employee now has to cost less than $36,000, all in, to win the comparison. That’s not a layoff memo. It’s a price tag, and the market just printed it.

In Moneyball, the best scene isn’t on a field. It’s in a draft room, where gray-haired scouts argue about a kid’s swing, his girlfriend, his “good face,” and Peter Brand — the Yale economics major Billy Beane hired off a computer — says the quiet thing: you’re all trying to buy players. You should be buying wins. Wins are runs, runs come from getting on base, and getting on base is a number nobody in the room was bothering to read. The scouts hate him. Of course they do. His one number just made fifty years of gut obsolete. The 2002 A’s ran a $41 million payroll against the Yankees’ $126 million and won twenty straight, an American League record, because Beane stopped paying for reputation and started paying for measured output.

That’s the whole story this week, and it has nothing to do with baseball.

For about a century we’ve paid most knowledge workers a salary for one reason, and it isn’t generosity: we couldn’t measure what any single person actually produced. So we guessed, called the guess a wage, and moved on. Sales got commissions because closes are countable. Traders got PnL. Hedge funds got two-and-twenty. Everyone whose output was legible got paid on it; everyone else got a salary by default. That default was never a law of nature. It was a measurement problem.

This week the measurement problem got solved, and the thing that solved it is the same thing everyone’s panicking about: the meter.

💸 Somebody Turned the Meter On

Start with the receipts, because they’re ugly. GitHub Copilot moved every customer to token-based billing this week, and the screaming was immediate — a $19 seat ballooning into a $3,000 bill overnight. Uber’s own people admitted the company torched its annual AI token budget in four months. One company spent $500 million in a single month because nobody set a spend limit. Marcelo Lebre, who runs a global payroll outfit and watches the price of human work for a living, clocked it from the other side: one Anthropic user ran up $150,000 of Claude Code in thirty days. The most expensive coworker in that company is already billed exclusively on what it produces.

Ed Zitron published the angry version the same day, under the headline “AI Doesn’t Have ROI.” He’s half right, and it’s the important half. The problem was never really the return. It’s that nobody could measure the cost, because every AI subscription you’ve ever used was deliberately subsidized to hide it. You paid $20 or $200 a month and burned ten times that in real compute, and the labs ate the difference to buy your habit. Then, quietly, in the first quarter of this year, they moved enterprise customers to true token pricing. The subsidy didn’t fade. It got switched off, on a billing cycle, and a few million people met the real number at once.

Here’s where I part ways with Zitron. He reads the meter and sees a death certificate. I read it and see a price discovering itself — a different thing depending entirely on where you sit. The lab selling tokens sees an overdue repricing. The over-levered startup that built its margins on subsidized inference sees a funeral. The patient buyer with cash sees a sale. Same event, three chairs. There’s no neutral account of this week, only the seat you read it from, and your entry point and time horizon decide which story is true for you.

The strategic read: Before you decide whether this is collapse or opportunity, figure out which chair you’re in. If you sell intelligence, the meter is your friend. If you bought a business model that only worked while intelligence was free, the meter ends you. Most people haven’t asked which seat they occupy — which is exactly why they’re about to be surprised.

🚕 The Uber Playbook, Run at Ten Times the Speed

We’ve seen this movie. Subsidize the product below cost, let the sheer capital intensity drag you all the way in, starve the alternatives, then charge what it actually costs once the customer has nowhere else to go. That’s the Uber playbook, and the AI labs are running it page for page. But the analogy gets interesting exactly where people get it wrong, so let me fix the two pieces everyone fumbles.

First, the capital isn’t optional, and it isn’t confidence. It’s a commitment device. Once you’ve sunk eighty billion dollars into data centers, half-measures are death — the balance sheet forecloses the exit. That’s what the capital “race” actually is. Look at the window: Google selling ~$80 billion in stock, SpaceX on the same order, Anthropic and OpenAI each reaching for a hundred billion, before the data-center leases nobody’s announced yet. litcapital called it a “cataclysmic exit liquidity avalanche” to 800,000 views. Round it how you like — it’s roughly half a trillion dollars trying to get raised at once, and that’s only the part we can see. You don’t raise that and then choose to be a niche player. You go all the way in because the money already decided for you.

Second — the piece I had backwards until I worked through it — the winner doesn’t have to kill its rivals. Uber never killed Lyft. Waymo’s growing. The black cars never disappeared. Uber won by ending up on top and expanding the battlefield into Eats. The labs are doing the same thing, and their target was never each other. They’re coming for search, for SaaS, for the consultants, for the whole sediment of middle management that exists to coordinate work an agent now coordinates for forty cents. That prize is so large that Anthropic, OpenAI, Google, and Microsoft can all win without taking it out of each other’s hides.

The proof it’s aimed at the services economy and not the other labs is in Microsoft’s own keynote. The company tuned its new in-house model on McKinsey’s workflows and bragged it beat GPT-5.5 by ten times on cost. McKinsey helped train the machine that bills by the outcome McKinsey used to bill by the slide. The call is coming from inside the house.

But here’s the catch. Uber got to gouge at the end because it had removed the alternatives. AI hasn’t — the substitutes are multiplying, not dying. So the real contest isn’t creating the surplus, it’s capturing it. When AI makes a job a hundred times cheaper, that surplus can land in the vendor’s pocket or evaporate into the customer’s, and competition pushes toward evaporation. There’s enough destroyed for everyone. Whether there’s enough captured is the whole game.

Why it matters: The compression is the part to sit with. Uber’s full arc — subsidize, dominate, revert to true price — took a decade. AI ran the loop in six months: miracle, jobs apocalypse, never mind, basically free, here’s your $3,000 bill. Whatever timetable you think you’re on, AI just put everyone on the same one. Plan for the cycle to finish in quarters, not years.

🧾 The Bill That Becomes a Benchmark

Now connect the two halves, because separately they’re just news and together they’re the story. The meter that exposed the true cost of the machine is the same instrument that finally meters the human sitting next to it. That’s Lebre’s point, and it’s the one nobody wants to say into a microphone: once “forty cents in tokens” and “four hours of an engineer” sit in the same ledger, you can finally compare them. And once you can compare them, the salary-by-default era is over.

This is the Moneyball draft room again, except the players are your staff. The obstacle to paying people on output was always attribution — you couldn’t isolate one person’s runs. That obstacle just collapsed. Lebre walks the history: piece-rate craftsmen, sharecroppers on a share of the harvest, factory workers paid per item until the assembly line scrambled attribution and hourly wages took over. Outcome pay never died; it just hid wherever output was hard to count. The meter drags it back into the open.

So run the math the way a CFO will. A $3,000-a-month AI seat is $36,000 a year, with no payroll tax, no benefits, no PTO, no manager, no real estate. If that seat does the work of a person, the person now has to come in under $36,000 fully loaded to be the rational choice. That’s a brutal number for most of white-collar America, and it doesn’t require anyone to be fired to start mattering. It just has to become visible — and the meter is what makes it visible.

What gets rewarded in that world isn’t hours. It’s what you can multiply. The engineer who deploys $50,000 in tokens and ships a product becomes worth more than the one who burns $5,000 and ships nothing — people get paid for what they orchestrate, not the time they log. Bounties on tickets. Revenue share on features. The hedge-fund comp model pushed down into jobs that never had it. Lebre flags the one real brake: labor law is built around hours and overtime, and outcome-based work needs new contracts and compliance rails before it scales across borders. That friction buys time. It doesn’t change the direction.

What this means for your business: Run the $36,000 test on one real workflow this week — not to fire anyone, but because your competitor is already running it. Price what the AI costs to do a repeatable task, then the fully loaded human cost beside it. If the gap is ugly, be the one who saw it first and moved the person toward the work that orchestrates, not the one who finds out when a leaner competitor underprices you.

🛠️ The Hedge Is Real, and It’s Still Hacker Work

Here’s the good news for everyone not selling tokens: the escape routes are real, and improving fast. Open models like MiniMax’s M3 claim frontier-class output at five to ten percent of the cost. Garry Tan, who runs Y Combinator, looked at Factory’s new model router — which cuts costs a quarter by sending each task to the cheapest model that can handle it — and called the shot: “the frontier labs will want their AI harness to be the moat, but the best case for consumers is that model capabilities flatten and commodify. Preview of the AI Harness Wars of 2027.” That’s the value-capture fight in a single tweet. And at the edge, unified-memory machines — a Mac Mini, one pool of memory feeding the chip instead of three discrete ones — run frontier-grade coding agents locally for the price of electricity, because the architecture is just cheaper on watts.

This is what Microsoft did at Build, which is why the week’s biggest product story is really a hedging story. Seven in-house MAI models, the Scout agent on OpenClaw, an NVIDIA-built stack from silicon to runtime — control “from chip to model to harness,” as one developer put it. The company that put $18 billion into OpenAI and Anthropic just unveiled the off-ramp from both, because the token meter made renting intelligence too expensive to stomach. Even the biggest investor in the labs is routing around the toll. So is Kirkland, building its own legal AI rather than paying frontier rates per query. When your own customers start building the escape hatch, the price got real.

But none of this is a Monday plug-in. Spinning up a Mac Mini, loading open weights, wiring them into one reliable workflow is the work of engineers and hackers, not the corporate worker who met AI through a chat tab. The narrative clock is synchronized — everybody got the same whiplash from miracle to $3,000 bill. The access clock is not. Engineers can hedge this afternoon; everyone else waits for someone to productize the hedge the way OpenAI productized the model. That gap — between “technically possible” and “my accounting team can use it” — is the next great business, and it’s wide open.

The action item: Price the hedge now, even if you can’t deploy it yet. Get a real number for what your core AI workloads would cost on open or local models, and keep it in your back pocket. You may not switch this quarter — but next time a vendor “adjusts” pricing, a number beats a shrug.

⚖️ The Meter Is the Honest Part. Watch Who Tries to Break It.

Step back from the panic and there’s something almost reassuring underneath it. A meter is just transparency wearing a worse outfit, and transparency, left alone, drags any market to its true price and clears it. The subsidized years were the distortion. The meter turning on is the market finally working — the price of intelligence discovering itself, the price of labor discovering itself right alongside it. That’s not the threat. That’s the system doing its job.

The threat is the “left alone” part, because two groups have every reason not to. The labs want the harness to become the moat — Tan’s whole point — so the price stays hidden inside a stack you can’t unbundle. And Washington just handed itself a lever: the executive order Trump signed this week gives the government a 30-day review of frontier models before release, “voluntary,” with the state deciding which “trusted partners” get early access. Even the Cato Institute flagged it: selective access is a moat the government hands out. Both moves keep the price from clearing in the open.

So the real question for the next year isn’t whether AI is a bubble. It’s whether the labs and their friends in government can wall off the price before the open, local, routed alternatives mature enough to force transparency anyway. The half-trillion-dollar raise is the bet that they can. The hedging behavior is the market betting they can’t. You don’t have to know who wins to position for it. You just have to know which side you’re funding with every renewal you sign.

---

What This Means For You

The meter is running, on the machine and on the seat next to it. Whether that’s an ending or a beginning depends on the chair you’re in and how long you plan to sit there. Three moves, whichever chair it is.

Put a meter on your own spend before it puts one on you. Set hard token caps in every tool today. The default settings were built to hide the cost from you, the same way they hid it right up until the bill arrived. Uber found out four months late. You don’t have to.

Run the $36,000 test, and act on the answer. Price one workflow, AI versus fully loaded human, and look at the gap without flinching. Then do what the number tells you — not panic, not layoffs, but moving people toward the work that orchestrates rather than the work that logs hours. The companies that move first get to choose. The rest get chosen for.

Pick your seat on purpose. Decide whether you’re selling intelligence or buying it, and build accordingly. If you’re buying, price the hedge and keep your model swappable so no vendor owns you. If you’re selling, your edge isn’t the model — it’s whether you capture the value you create before the harness wars commoditize it away.

The meter was never just for the machine. It only ever felt that way because the price of everything else was still hidden.

Stop pricing the AI. Start pricing the seat next to it.

Three Questions We Think You Should Be Asking Yourself

Which chair am I in — and for how long? Collapse, opportunity, and repricing are the same event seen from different seats. If you can’t say whether you’re selling intelligence or buying it, and on what time horizon, you can’t know which story this week is for you.

If I ran the $36,000 test on my own team, what would it show? Price a real, repeatable workflow as an AI bill against the fully loaded cost of the human doing it now. You don’t have to like the answer. You do have to know it before a competitor uses it against you.

When the meter prices my people the way it just priced the machine, what makes a person worth more than their bill? The answer is what they orchestrate, not the hours they log. If your best people are valued today for time-in-seat rather than output, you’re paying on the metric that’s about to stop mattering.

---

It’s about getting things down to one number.”
— Peter Brand, Moneyball (2011)

— Harry and Anthony

Sources

• Ed Zitron — “AI Doesn’t Have ROI” (GitHub token billing, Uber’s budget, the $500M month)
• Marcelo Lebre — per-token vs. per-hour thread (the $150K bill, 1.3 quadrillion tokens, the metering-of-labor argument)
• Garry Tan — Factory model routing and the “AI Harness Wars of 2027”
• CNBC — Microsoft’s MAI models, tuned on McKinsey, 10x cost efficiency vs. GPT-5.5
• Computerworld — Microsoft Scout, the OpenClaw-based agent
• @litcapital — the “exit liquidity avalanche” (Google, SpaceX, Anthropic, OpenAI raising at once)
• AP / ABC — Trump signs the executive order vetting top AI models for national security
• Fortune — Anthropic’s office vending machine became AI-run stores and cafés
• Moneyball (2011), dir. Bennett Miller — the Oakland A’s, the draft room, and getting things down to one number