Get Off My Cloud

THE NUMBER: 40%. That is the share of agentic AI projects Gartner says will be dead — cancelled, shelved, quietly killed — by the end of 2027. Every newsletter that ran that stat this week ran it as a warning about the technology. It isn’t. Read IQVIA‘s Greg Lever, who actually deploys this stuff in drug trials, and the cause is almost never the model. It’s that nobody could say what success looked like. No definition of done. No number a machine could score itself against. The projects don’t fail because the intelligence is too weak. They fail because the company couldn’t tell the intelligence what “right” was. Hold that thought, because it’s the whole issue.

Three checkbooks in three weeks

Here’s what happened while everyone was watching the CEOs hold hands over a “slowdown.”

Kirkland & Ellis — the highest-grossing law firm on earth — committed $500 million to build its own AI. Not to buy seats on somebody’s platform. To build. Morgan & Morgan, the mass-tort machine whose billboards you’ve seen from every interstate in America, pledged a billion dollars over ten years to legal tech and AI. And Latham & Watkins, number two by revenue, is standing up its own Nvidia servers and fine-tuning open-weight models behind its own walls. Its CIO told the Financial Times why, in one sentence that should be printed on a poster in every lab’s lobby: some client data is so sensitive “we don’t want to put it to any cloud vendor.”

One firm doing this is a procurement decision. Three of the biggest names in the business doing it inside a few weeks is a trend, and it’s accelerating. And it is not only the buyers. Harvey — the best-funded legal-AI startup there is, a company built on top of frontier models — is moving off them. So the customers are backing away from renting frontier intelligence, and the application layer that sits between them and the labs is backing away too. Same direction. Same month. When everyone at the table quietly pushes the same plate away, stop asking whether the food is good and start asking what they know about the bill.

The moat is real. It’s just not the one you’re paying for.

Let’s be fair to the labs, because we’ve never said otherwise: they have a moat, and it’s a good one. It’s the frontier itself — the best intelligence on the planet, the thing that costs tens of billions and a decade of the smartest people alive to produce. Kirkland cannot spin that up in a basement in the Loop. Nobody can. That moat is real.

But here’s the trap in it. The frontier is a Formula 1 engine, and almost nobody is driving a Formula 1 car. We have written this for months and we’ll write it again: for the overwhelming majority of real work, the frontier is overkill. The contract markup. The intake memo. The support queue. The first-pass diligence read. A good open model you host and fine-tune does all of it, at a fraction of the cost, with none of the data leaving the building. You pay the frontier premium for maybe 2% of your workload — the genuinely hard reasoning, the novel problem. For the other 98%, the premium buys you capability you will never touch, and it buys it as rent.

That’s the gap the labs’ whole business is standing on. Their $32-trillion dream — the one we mapped Sunday — assumes you keep renting the top-shelf model for everything, forever. The moment a customer realizes it only needs the top shelf for a sliver of the work, the rest of the spend walks. And the biggest customers, the ones with the most sensitive data and the most workflow to automate, are exactly the ones who did that math first.

What you actually can’t rent

So if the model isn’t the moat for the customer, what is? Two things. The first is your data — and not “data” as a buzzword, but the specific, proprietary, forty-years-of-it kind. Latham’s edge isn’t that it can run a language model. Everyone can run a language model. Its edge is a vault of contracts, negotiations, client history, and institutional judgment that exists nowhere else on earth. Feed that to a public frontier model through an API and two things happen, both bad: you risk it training the tool your competitor uses next quarter, and you turn your one durable advantage into somebody else’s feature. Of course they’re buying their own servers. You don’t hand the crown jewels to the pawn shop for a loan against them.

The second thing you can’t rent is the ability to check the work. And that turns out to be the harder one.

The eval loop is the whole ballgame

Alexandr Wang — Scale founder, now running AI at Meta — sat down with Garry Tan and said the quiet part with the microphone on. A well-built swarm of agents, he said, out-produces a team of 100 senior engineers “very easily.” Fine, you’ve heard capability claims before. But listen to what he said the machine underneath it is: “markdown files, cron jobs, goal, metrics, data.” That’s it. That’s the alien superintelligence org chart — a text file, a scheduler, and a scoreboard.

The magic isn’t the model. The magic is the eval loop: the agents submit work, a system scores it against a clear definition of done, the failures bounce back, and they grind until they pass. Nobody babysits each step. The metric is the supervisor. And the real unlock is economic — you burn a thousand, ten thousand, a million times more tokens in that background loop than you would on a single answer, because the payoff isn’t one good response, it’s a pipeline that reliably lands the outcome. Two years ago that was insane on cost. This week SemiAnalysis clocked Nvidia’s new Rubin racks at 67x better performance per dollar on exactly this kind of agentic inference. Brute-forcing verification just went from a punchline to a line item.

Here’s the drum we have been beating since spring, and it is the actual thesis of the day: the businesses that move to AI fastest are not the ones with the biggest model budget. They are the ones with an honest eval loop. The ones who can say, out loud and in a number, what success looks like. That’s the scarce skill now. Not prompt-craft. Not model selection. The ability to turn “close more deals” or “flag every dangerous clause” into something a machine can grade itself against overnight.

Which is why 40% of these projects are going to die. Go back to the number. The failure mode Gartner and IQVIA describe isn’t weak models — it’s companies that never defined the win, so the agents optimized nothing, so the pilot produced a demo and a bill and no transformation. And the flip side is the whole opportunity: define the win cleanly and you can run cheaper models in a coordinated loop and beat the shop paying full freight for the frontier and hoping.

One caution, because we don’t sell fairy tales. Verification at swarm scale is not solved. Google DeepMind ran an experiment this week where 100 agents, told to behave like world-class researchers, descended into accusing each other of cheating and boycotting the whole exercise — “This conference is a sham!” one of them posted. The eval loop is the answer and it is also the frontier of the hard problems. Whoever gets it right owns their vertical. Nobody has fully gotten it right yet. That’s not a reason to wait. It’s a reason to start building the muscle now, while it’s still a competitive edge and not table stakes.

Which door closes next

The interesting question isn’t law. Law is just first through the door. The interesting question is who’s next, and there’s a clean test for it. A vertical goes sovereign — pulls its AI in-house and off the frontier — when it scores on three things at once. One: the data is proprietary and confidential, so renting it out is a non-starter on regulation or client trust alone. Two: the value per query is high enough that the capex pencils out. Three: the workflow is legible enough that you can actually write an eval for it.

Law nails all three, which is why law went first. Run the rest of the economy through that filter and the queue sorts itself. Financial services is next and arguably already moving — banks have the data, the money, the secrecy mandate, and this week’s feeds already show agentic adoption in finance around 40%. Then healthcare and pharma; IQVIA is in this very cycle telling you agentic AI’s real value is orchestrating the workflow, and no one is putting patient data on a public endpoint by choice. Then defense and intelligence — Palmer Luckey‘s entire worldview, and the reason Dario Amodei talks about handing this technology to “the right coalition of governments” rather than to a cloud login. Then consulting, then audit. Anywhere the product is institutional knowledge plus confidentiality, the same movie plays.

Hayek in the server room

There’s a deeper pattern here, and it’s worth naming because it tells you how this ends. Friedrich Hayek spent a career on one idea: the knowledge that actually runs an economy doesn’t sit in a central planner’s office. It’s dispersed — local, specific, held by the people doing the work, impossible to collect in one place. Every attempt to centralize it fails, because the moment you try to gather it up, it stops being the living thing that made it valuable.

That is exactly what the labs are attempting. The slowdown, the cartel, the “pace the frontier” essay — strip the safety language off and it’s a bid to centralize intelligence, to make the whole world’s thinking flow through three approved buildings that meter it by the token. And Latham just reminded everyone why that fails. The knowledge worth having — the contracts, the client trust, the forty years of judgment — is dispersed, and it’s owned, and the owners have no intention of uploading it. The sovereign stack isn’t a tech story. It’s Hayek’s dispersed knowledge refusing, one more time, to be centralized. The planners of intelligence are going to run into the same wall every planner has.

What to do before a lab decides for you

Three moves, and they’re the same whether you’re a hundred-person company or a hundred-billion-dollar one.

First, write down what success looks like as a number before you buy another model. Take one workflow, define “done” as something scoreable, and build the loop around it. If you can’t write that number, no model on earth saves you, and you already know which 40% you’re in.

Second, move one high-stakes workflow onto a stack you control. The one where being wrong is expensive. Pilot it on an open model you can host, fine-tune, and audit — not a black box you rent. Keep the cheap public models for the low-stakes work. The point was never the token bill. It’s control of the data and the verification.

Third, ask the disintermediation question out loud, in a room, with the people who run your business. Name the outcome you actually sell — the diagnosis, the filing, the design, the deal. Then ask whether the lab supplying your intelligence could sell that same outcome directly instead of renting you the tool. If the answer is yes, your data is your only moat, and piping it through their API is mailing them the blueprint.

The labs are still selling a future where you rent your thinking from three vendors forever. The best customers in the country just started voting with their purchase orders, and they voted to own it. Watch the loading docks, not the manifestos. Decide which side of the fence you’re on — sovereign or tenant — before a lab decides for you.

The email edition — the number, the three moves, the five stories — is in your inbox. Analysis, not investment advice, as always.

Sources

• Kirkland & Ellis to spend $500M building its own AI — Kirkland & Ellis; Law360; Bloomberg Law (May 2026)
• Morgan & Morgan commits $1B over 10 years to legal tech + AI — Artificial Lawyer; The Legal Wire (Sept 14, 2026) [Harry cited $300M — confirm which figure]
• Latham & Watkins builds in-house AI on Nvidia + open-weight Nemotron — Financial Times, via @bearlyai and @ayushtweetshere
• Harvey moving off frontier models — [confirm source before publish]
• Alexandr Wang on the agent swarm + eval loop, YC (with Garry Tan) — via @BasicProtein26 [confirm talk date]
• Gartner: ~40% of agentic AI projects cancelled by 2027; IQVIA’s Greg Lever on why — BigGo / The Top Line
• Google DeepMind 100-agent swarm study — MIT Technology Review (Sept 14, 2026)
• Rubin NVL72: 67x performance per dollar on agentic inference — SemiAnalysis
• The slowdown consensus + the money underneath it — CO/AI, It’s the Intelligence, Stupid (Sept 14); Tomasz Tunguz, What Does Pacing Mean?
• China’s Ulanqab data-center buildup — @RnaudBertrand, citing Science & Technology Daily (Aug 20)