The largest firms are building AI they control

On 10 September 2026 the Financial Times reported that Latham had bought Nvidia servers and was adapting open models to run on them. Bloomberg Law reported the next day that the servers are in a data centre only Latham staff can enter, that the firm has several Nvidia H200 chips, and that it has been buying them for about three years.

Rene Mendoza, the firm’s chief information officer, told the Financial Times why. Some client information is sensitive enough that Latham does not want to give it to any cloud vendor. Michael Rubin, who chairs the firm’s AI committee, told Bloomberg Law that the point is to run those models with the firm’s own software, on the firm’s own machines.

Latham has more than 900 technology staff. It did not say what the servers cost. It has not stopped using other AI products. This is a second path, for the work that should not leave.

On 28 May 2026 Reuters reported that Kirkland would spend $500 million of its revenue over three to four years, including $100 million in 2026, on a platform it builds itself. The firm’s revenue last year was $10.6 billion. It said it would still license some outside tools.

Jon Ballis, the chair, told the Financial Times the aim in plain words. Take the firm’s own knowledge and put it to work across the practice. Own the technology, so the companies helping to build it cannot sell the same thing to everyone else. About 180 technology people are on it. The design comes from how 250 of Kirkland’s own lawyers do the work.

In June the firm announced a Palantir tool for private-fund work: fund documents, side letters, and the promises a fund has already made to investors. Erica Berthou, a partner in that group, said the point was to put senior judgment into a system more than a thousand lawyers can use.

Bloomberg Law also reported that Kirkland was hiring someone to run on-premise GPU clusters. The firm would not say whether it had bought the chips. We will not treat a job advertisement as a purchase. What Kirkland has confirmed is the money, the ownership, and the reason: a tool every competitor can license does not know how Kirkland practises.

A Hong Kong practice is not going to hire 900 technologists, or spend half a billion dollars, to reach the same conclusion. The conclusion is still the right one.

There are two things those firms are paying for. The file stays on machines they control. And the system is built from their own work, so it is not the same product the firm next door rented this morning.

That is what we build. Makebell runs on the firm’s own machines. We never receive the file, so we cannot see it, train on it, or keep a copy of it. The company that trained the model does not receive it either. The firm’s precedents, playbooks and past matters sit in the system, and they stay the firm’s. A lawyer still decides what goes out.

Latham can afford to assemble that from chips, engineers and a leased hall. Most firms should not try. The principle does not require their budget. It requires a system that was designed to run inside the firm, on the firm’s own knowledge, from the beginning.