Get Off My Cloud

The tenants are buying the building.

💲 Kirkland & Ellis put up $500 million to build its own AI. Morgan & Morgan pledged a billion over ten years to legal tech and AI. Latham & Watkins is buying its own Nvidia servers and fine-tuning open weights behind its own walls, because — its CIO told the FT — some client data is so sensitive "we don't want to put it to any cloud vendor." Three of the biggest names in law, three different checkbooks, same instinct inside a few weeks. Hey, you. Get off of my cloud.

One firm doing this is a headline. Three is a trend, and it's accelerating. And it isn't only the customers — Harvey, the best-funded legal-AI vendor there is, is climbing off the frontier models it was built on. When your buyers and your app layer both back away from renting frontier intelligence in the same breath, that tells you what the frontier is actually worth to them.

Now, the labs do have a moat. It's the frontier — genuinely the best intelligence on earth, and Kirkland can't spin that up in a basement. We've never argued otherwise. But the frontier is a Formula 1 engine, and almost nobody is driving a Formula 1 car. The contract review, the intake memo, the support queue — a good open model you control handles all of it. Pay the frontier premium for the 2% of work that needs it. For the other 98%, you're renting something you could own.

So what actually separates the winners? Alexandr Wang, now running AI at Meta, said it to Garry Tan: a well-built agent swarm out-produces 100 senior engineers "very easily," and the machine underneath is just "markdown files, cron jobs, goal, metrics, data." The magic isn't the model. It's the eval loop — the agents grading each other against a clear definition of done until the work is right. Nvidia's new Rubin racks cut the cost of that by 67x, which turns "burn a million tokens checking the work" into a plan instead of a punchline.

This is the drum we keep beating: the businesses that move to AI fastest aren't the ones with the fattest model budget. They're the ones with an honest eval loop — the ones who can say what success actually looks like and write it down as a number a machine can score. Gartner says 40% of agentic projects die by 2027, and the cause is almost never the model. It's that nobody could define the win.

So the market splits clean. Firms with proprietary data and an honest eval walk off the labs' land and take the crown jewels. Everybody else rents good-enough thinking by the sip. The labs keep the frontier — and get to watch the highest-value work they assumed they'd bill for get built in-house, on cheaper models, by the customers who know exactly what they need.

Decide which side of the fence you're on — sovereign or tenant — before a lab decides for you.

At COAI today: the full Signal/Noise — the map of which verticals close to the frontier next, and why law went first — is live at getcoai.com.

Outsider Labs Advisory →

40%. That's the share of agentic AI projects Gartner says will be dead by 2027 — and the killer is almost never the model. It's that nobody could say what success actually looks like. The companies pulling ahead aren't the ones renting the biggest model; they're the ones with an honest eval loop — a clear, scoreable definition of done. The frontier isn't the moat. Knowing what "right" looks like, and owning the data to prove it, is.

Write down what success looks like — as a number — before you buy another model. The fastest AI adopters aren't the ones with the biggest budget; they're the ones with an honest eval loop. Take one workflow and define "done" as something a machine can score. If you can't write that number, no model will save you — you're already in Gartner's 40%.

Move one high-stakes workflow onto a stack you control. Take the process where being wrong is expensive — the contract review, the compliance memo — and pilot it on an open-weight model you can fine-tune and audit, not a rented black box. Keep the cheap public models for the low-stakes stuff. The point isn't the token bill. It's owning the data and the verification.

Ask the disintermediation question out loud. Name the outcome your company actually sells — the diagnosis, the filing, the design. Then ask whether the lab supplying your intelligence could sell that same outcome directly. If yes, your data is your only moat, and piping it through their API is mailing them the blueprint. Decide that on purpose.

Monday, September 14

— Harry and Anthony

Sources: