Mark to Market

Three Clocks, One Bill

Ensemble Health Partners runs the revenue cycle for hospitals you've probably been treated in, and this year it's spending up to $100 million on AI. This week its people told The Information something the frontier labs would rather you not hear: they moved a stack of that work off OpenAI's best model onto one 23 times cheaper, same vendor, and it worked fine. On the routing platform that carries those tokens, 65% of everything processed in June came from open-source models. A nine-figure buyer just said the expensive model is optional.

We've been saying the model was never the moat for a month — Buy Wins, Coffee's for Closers, Fulcrum, Central Casting. I'm not taking a lap. I'm marking the book. When a buyer this size puts a name and a number on the trade, the thesis stops being ours and becomes the market's. So thank you, Ensemble, for making the point more eloquently than we managed.

Here's the part the cheerleaders skip. Three clocks are running. The labs invent fast. The open models commoditize what the labs invent almost as fast — we watched a banned American model (Fable 5) get replaced by a Chinese one over one weekend this month. The enterprise deploys slow, in quarters: you test it, wire it, train people, get legal comfortable. Line those up and you get the cruelest arithmetic in tech, where by the time you deploy the frontier capability you paid a premium for, it's already free. You bought express shipping on something that landed on the clearance rack.

That's the missing ROI everyone keeps writing about. MIT found 95% of corporate AI pilots show no profit-and-loss impact, and the consultants blame change management. It's simpler and worse: the premium evaporates before you can catch it. The math only flips where you deploy as fast as the thing commoditizes — coding, trading, ad auctions, the tight-loop jobs. Everywhere else the clock eats your lunch.

Which is why the labs are scrambling on two fronts. OpenAI rolled out its own chip this week, Jalapeño, built with Broadcom, because the only way to defend a margin buyers won't pay is to crush your own cost of making a token. The market clocked the squeeze from the other side: the Philadelphia chip index fell 7.9% and Korea halted its stock exchange as the memory trade that priced infinite demand cracked. Revenue leaking out the front, capex piling up the back. Tough sledding.

So if the model is free and the premium can't be caught, what's left to own? The one thing that doesn't commoditize: knowing whether the work is right. Validation. The shop that can grade the output is the shop that can deploy before the capability goes free — and beating the clock is the only reason speed ever pays. The model was never the moat. Reading the tape was.

At COAI today: the full Signal/Noise — the three clocks, why coding is the one place the premium survives, and the validation playbook — is live at getcoai.com.

Outsider Labs Advisory →

23×. That's how much cheaper the model Ensemble Health swapped to is than the OpenAI frontier tier it dropped — same vendor, same hospital revenue-cycle work, a $100-million AI budget on the line, and it held. Across that routing platform, 65% of June's tokens now run on open-source models. The frontier capability isn't worthless. It's that by the time most companies finish deploying it, the cheap bench already does the job. The moat was never the model. It was knowing that.

Audit one workload's model bill today. Take your highest-volume AI task, run it on a cheap or open model against the same test set, and check the output holds. That is exactly the move that found Ensemble a 23x cut. If nobody in your shop owns the question "is this the cheapest model that does the job," you just found your assignment.

Build the scoreboard before you scale. The reason 95% of pilots show no return is that nobody could grade the work, so nobody trusted it enough to ship. Pick one workflow and write down, in plain words, what "right" looks like. The team that can grade the output is the team that deploys it before the capability commoditizes.

Standardize on a routing rule, not a model. Frontier tokens for the 20% of work where the extra smarts change the outcome; cheap or open for the rest. Write the rule down and give someone the keys. Today's frontier is next quarter's commodity, and a static default quietly overpays every day you leave it alone.

Thursday, June 25

A $100M buyer calls the frontier 23× too expensive. (Full analysis above.) Ensemble Health moved hospital revenue-cycle work to a model 23 times cheaper than OpenAI's top tier and said it worked; 65% of its platform's June tokens were open-source. The expensive model just became optional.

OpenAI becomes a chip company. (Full analysis above.) The Jalapeño "Intelligence Processor," built with Broadcom, is OpenAI's first in-house silicon and its bid to go full-stack. Read it plainly: the only way to defend a token margin buyers won't pay is to own the cost of making the token.

The memory trade cracked. The Philadelphia chip index dropped 7.9%, Micron and SanDisk roughly 13% each, and Korea's Kospi tripped its circuit breaker — one session before Micron's earnings call. The market's first real vote that the AI-capex bill could be near its peak, not its floor.

Agility Robotics heads to Wall Street at $2.5B. The warehouse-humanoid maker is going public via SPAC — labor-as-a-service underwritten by the public markets. The question isn't whether Digit walks. It's whether the math pencils against a forklift and a person.

Gemini 3.5 Flash gets native computer use. Google baked agentic computer-use into its cheap, fast tier. A capability that was a premium standalone six months ago now ships down-market — the commoditization clock, running exactly on time.

Where the cycle caught up to us this week.

The model isn't the moat. "The lever is free; the value is the fulcrum — proprietary data, judgment, distribution" (us, Show Me Where to Put the Fulcrum, Jun 16) → Microsoft's Brad Smith, amplifying Emergence Capital's Gordon Ritter, "the lasting value of AI won't reside in any single model" (his phrase: "above the model"), reposted by Satya Nadella (X, Jun 24).

Own the router, not the model. "When the model is a rental, own the layer that decides which one gets the part" (us, Central Casting, Jun 23) → Implicator on Sakana Fugu: buyers get "one black-box router… a new layer they still cannot see" (Implicator.ai, Jun 24).

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: