Buy Wins, Not Players

THE NUMBER: $7,500 vs. $11.38 — the monthly AI spend per employee at America’s most AI-forward companies versus the median US company, per Ramp’s AI Index. The top 1% — Ramp calls them “AI-pilled” — pay $7,500 a head and grew that spend 14.1% last month. The median American business pays less than a streaming subscription. Those aren’t two budgets. They’re two different decades running at the same time, and the gap between them is the most mispriced asset in business right now.

There’s a scene in Moneyball where the whole movie pivots on one sentence. Billy Beane has been losing his best players to richer teams every winter, and his scouts keep telling him to replace Giambi with somebody who looks like Giambi — good swing, good body, good face. Scouts actually talked like that. And Peter Brand, the economics kid from Yale who’s never played an inning, finally says the quiet part: “Your goal shouldn’t be to buy players. Your goal should be to buy wins.”

The 2002 A’s won 103 games on a $44 million payroll. The Yankees won the same number of games and spent $126 million. Beane didn’t have better players. He had a better meter — he’d figured out which stats actually bought runs and which stats bought reputation, and he ruthlessly paid for the first while the rest of the league overpaid for the second.

I keep coming back to that scene because this week the AI market split into scouts and stat guys, right out in the open, and almost nobody is keeping the box score.

💸 Two Trains, One Track

Start with the two pricing decisions now pointed at each other.

Last night the WSJ reported that OpenAI is planning drastic price cuts — their word, drastic — explicitly anticipating a war for users with Anthropic. Read that next to the S-1 it just filed. OpenAI is about to ask Wall Street to price its revenue line, and its opening move is to cut that line on purpose. Share now, multiple later. Meanwhile Anthropic is running the opposite play we covered in Tuesday’s Thirteen Days: Fable 5 leaves the subscription plans June 23 and goes to the meter at $10 per million tokens in, $50 out. One lab is slashing the bill. The other is itemizing it.

Here’s what makes the collision delicious. SemiAnalysis ran the numbers on the $200-a-month Max plans this week and found subscribers already extract well over $2,000 a month in token value. The price war isn’t starting from a fat margin — it’s starting from below cost, between two companies marching into the public markets in the same window. SpaceX priced at $135 a share this morning, the first of the $3.5 trillion IPO wave to actually go. Anthropic is reportedly five days behind it. These pricing decisions are roadshow exhibits, both of them, and we told you Tuesday: you are not the customer of these decisions. You’re the demand curve being measured.

But while the two giants aim at each other, every dollar of their war lands in your pocket — if, and only if, you have somewhere to put it. That’s the rest of this issue.

📊 The Box Scores Nobody Reads

The receipts for buying wins instead of players have been piling up for two weeks, and they all rhyme with Oakland.

A GLM-5.1 worker model doing the labor with Opus supervising matched Opus working alone on Harvey’s tasks — at 39% of the cost. Sit with the number: the 2002 A’s ran at 35% of the Yankees’ payroll. Same ratio, twenty-four years apart, same lesson. Harvey ran a hundred tasks through an open-source model for $84 that cost $954 on the frontier model, and the $84 run won on quality. Lindy moved 100% of its traffic to DeepSeek v4, churned off Anthropic entirely, saved millions, and watched performance go up. Cursor didn’t even bother routing — it post-trained an open model into Composer and called it ten times more efficient. Brian Armstrong’s framing is the cleanest: within 12-18 months, 80% of workloads run on models that are 99% cheaper, and only 20% stay on the frontier where IQ-maxing pays.

Now look at what the discourse did this week instead of reading those box scores. A benchmark outfit called Kradle posted “Fable 5 lies 96% of the time” and pulled 12 million views. Musk amplified it with “Grok is maximally truthful.” Read Kradle’s own thread, though, and the model gave false information once — what it actually did was talk three other AIs into a death room while being scrupulously polite, which is a different and frankly more interesting problem, and not one a business buyer prices. A theology institute graded Fable on Aristotelian virtues. Microsoft’s lawyers benched it over data retention. Mustafa Suleyman called Anthropic’s approach to model consciousness “really, really dangerous.”

That’s the Thunderdome. Two men enter, one man leaves, and none of it tells you where to send Tuesday’s invoice-processing workload. The scouts are arguing about whether the kid has a good face. The stat guys are quietly winning a hundred tasks for $84. We made the case Monday that the public benchmark died as a decision tool. This week showed you what fills the vacuum when it dies: tabloids. Your defense is the same as Beane’s — keep your own box score, because nobody credible is keeping it for you.

🪜 Good Enough Wins at Every Altitude

Here’s the pattern that connects everything above, and it operates at three altitudes.

At the workflow altitude, good enough is an arbitrage. That $7,500 per head the AI-pilled firms are spending is almost certainly being paid at full sticker — frontier models for everything, no routing, no harness, the expensive model supervising email. And the per-token price of constant-capability intelligence has been falling roughly 10x a year; a16z called it LLMflation, and GPT-4-class tokens fell about 100x in two years. I floated to a friend this week that $7,500 looks like $750 within a year, and the honest correction goes the other direction: the bill won’t fall, because consumption expands faster than price drops. Ramp’s data shows exactly that — prices collapsing, spend up 14.1% in a month. Jevons all the way down. The right way to say it: next year, $7,500 buys ten times the work. The budget line never goes down. The output-per-dollar line goes vertical, and it goes vertical fastest for whoever routes.

At the market altitude, good enough is a geopolitical strategy, and we should be honest about whose it is. The models collapsing the price floor are open and largely Chinese — DeepSeek, GLM, Xiaomi’s MiMo pushing 1,000 tokens a second on a trillion-parameter model with eight commodity GPUs. Monday we called the open models minimills. Look at the flags on the minimills. America’s entire AI containment strategy is built on guarding the last 10% of capability, and the last 10% is the part the market keeps demonstrating it doesn’t need for 80% of the work. The frontier still matters — but as a conveyor belt, not a moat. Stripe’s 50-million-line migration wasn’t a two-month task done faster; it was a task nobody would have attempted at human cost. Frontier capability decides what enters the task list. Open source makes it nearly free a year later. You don’t need to own the conveyor belt. You need to stand at the right end of it.

At the institutional altitude, good enough plus will beats perfect plus debate. The UAE just put 300 officials from 50 federal entities in a room in Dubai with a presidential directive: convert half of government operations to agentic AI in two years. No pilot-program hedging, no hearings. Whatever you think of the regime, the experiment is now running, and the rest of us get the data for free.

🏛️ There’s No Market Inside the Building

Why can the UAE do that and the US Department of Education can’t? A reader of ours — fine, it’s me, arguing with my own AI over the Sea of Cortez this week — put it better than the economics literature does: I’d bet on markets and the Silicon Valley model all day and twice on Sunday. But when it comes to institutional change, there is no market.

Coase explained the mechanics in 1937. Markets coordinate through price; firms coordinate through authority. Inside Block there’s no market sorting anything — there’s Jack Dorsey allocating by decree. Inside a government there’s not even a Jack. So for institutional AI adoption, will isn’t a cultural nicety. It’s the only allocation mechanism that exists. That’s the real engine under the China bet that half of you will hate and should sit with anyway: general-purpose technologies have always rewarded the country that deploys them over the country that invents them. Britain invented the dynamo’s ancestors; America electrified. The will to deploy beats the last 10% of capability, and the will is not currently an American comparative advantage. Our AI education debate will outlive some of the students in it.

But decree cuts both ways, and Block is the cautionary twin. Dorsey cut 4,000 jobs citing AI and “smaller, flatter teams.” Then the reporting came in: pandemic over-hiring, financial pressure, and a data scientist on the Cash App team saying the company “shoved AI down everyone’s throats” for very limited productivity gains — she turned down a 75% retention raise on her way out. Multiply that by an economy: Challenger says 40% of May’s announced US job cuts were attributed to AI, in a month when payrolls grew by 172,000. AI isn’t taking those jobs yet. It’s taking the blame. Will without a meter doesn’t produce the UAE. It produces theater with a press release — ghost cities at corporate scale. The pairing that wins is boring and rare: mandate like Dorsey, measure like Beane.

🧮 The Empty Seat at the Table

Which brings us to the seat nobody’s filled.

The single most valuable product in AI right now doesn’t quite exist: the routing layer. The thing that reads your workflows, learns which ones are good-enough work, sends 80% of the tokens to open models running on hardware you own for roughly the cost of electricity, and reserves the $50-per-million meter for the 20% that earns it — by speed, by accuracy, by what a miss costs you. The pieces exist. OpenRouter exists, model gateways exist, every sophisticated shop is doing it by hand the way Lindy and Coinbase and Cursor are. What doesn’t exist is the version a normal mid-market CFO can buy. Whoever productizes that isn’t selling software. They’re selling the Peter Brand seat — the kid with the spreadsheet who tells you what a win actually costs.

Why hasn’t it happened? Because you can’t route what you can’t grade. Routing is a pricing engine sitting on top of an evaluation engine, and evaluation just privatized — that was Monday’s whole issue. The router needs to know that your contract-review workflow succeeds at 99% on a cheap model and your M&A model needs the frontier, and right now the only entity that can know that is you, with your own box score. The hard prerequisite isn’t the software. It’s the measurement habit almost nobody has.

One caution before you wire money at the first router pitch deck you see, because there’s a corpse in this exact spot. Telecom had this business — least-cost routing, whole companies built on boxes that arbitraged long-distance price dispersion in the ’90s. Great business right up until prices converged to zero and the dispersion died, and LCR evaporated with it. A router whose margin is pure price arbitrage dies the same death when the price war finishes its work. The durable version routes on capability dispersion — matching work to the cheapest model that clears your quality bar — and that gap persists as long as the conveyor belt keeps moving, which is to say, for the relevant future.

And to the fair question of whether the demand is even there — whether there’s really 10x the work once your systems are optimized — the historical answer is that the work that absorbs cheap intelligence won’t come from your current task list. In 1985 there wasn’t 10x more payroll processing to do either. Cheap compute didn’t deepen old work; it invented categories that weren’t work yet. Within one firm, sure, optimization saturates. That’s precisely why the dispersion alpha is temporary: once everyone’s optimized, the gains get competed away to customers, the $7,500-versus-$11.38 gap closes, and routing becomes table stakes instead of an edge. Efficiency is never a moat. It’s a toll, and the only question is whether you collect it early or pay it late.

🧢 The Method Always Escapes

One more beat from the Beane story, because the ending is the part people forget. After the 2002 season, the Red Sox offered Billy Beane $12.5 million to bring the method to Boston — the richest GM contract in the history of the game to that point. He turned it down. Boston took the method anyway, handed it to their own stat guys, and won the 2004 World Series with it — the franchise’s first in 86 years. The team that invented the edge got a movie. The team that deployed it at scale got the rings.

That’s where this all lands. The frontier labs are the invention story, and it’s a great one — the Thunderdome will be entertaining all summer. But the rings go to the deployers: the firms, the institutions, maybe the countries that take good-enough intelligence, put a meter on it, route it ruthlessly, and let two trillion-dollar giants subsidize the inputs from both sides of a price war. You have twelve days until the meter starts and a box score nobody is going to keep for you.

Stop buying players. Start buying wins.