Latham buys AI servers: what matters is what it wants to control

One of the world's major law firms is buying servers to run artificial intelligence on its own infrastructure.

The news invites a simple headline: lawyers are starting to build their own AI. But the interesting part is not that a law firm is buying powerful computers. It is what it intends to do with them, what it wants to keep under its control, and why it does not consider it sufficient to subscribe to the same tools its competitors can buy.

The Latham & Watkins case raises a question that also affects much smaller organisations:

Which part of our capacity to work do we want to buy externally, and which part do we want to build and retain internally?

What Latham is doing

According to Bloomberg Law's reporting on 11 September 2026, based on interviews with the firm's leaders, Latham has acquired several NVIDIA H200 GPUs and is experimenting with adapting open-weight models. The initiative began approximately three years earlier.

Its servers are operating in a third-party data centre, with access restricted to Latham personnel. They are therefore not necessarily machines installed inside its offices, but firm-owned infrastructure housed at an external facility. The firm has not disclosed the cost of the investment. Bloomberg Law, reproduced on Latham's website.

For a non-technical reader, three elements are worth distinguishing.

GPUs are processors particularly suited to the operations many AI models require.

Open-weight models allow their parameters to be downloaded and run on one's own infrastructure, under their respective licences. They are not necessarily unrestricted software.

And model adaptation, or fine-tuning, means adjusting an existing model. It does not mean training an AI comparable to those of the major laboratories from scratch.

But neither the servers nor the model constitute a legal system on their own. They need surrounding software, sources, procedures, permissions, evaluations and people responsible for checking the results.

It is not abandoning external providers

This is the nuance easily lost on social media.

In a statement reported by The American Lawyer on 10 September, Rene Mendoza, Latham's chief information officer, explains that the strategy seeks flexibility to develop, test and deploy AI without relying on a single provider. The report confirms that the firm maintains relationships with external providers and access to ChatGPT, Claude and Gemini. It also mentions open Nemotron 3 models. The American Lawyer, reproduced on Latham's website.

The choice is not necessarily “everything inside” or “everything outside”.

It can mean combining both worlds: external tools where they add value and internal capacity where there are reasons to retain greater control.

My reading is that the important decision is not to dispense with providers, but to avoid making a subscription to one the only possible way of working.

First reason: deciding where information is processed

Before asking which model answers best, an organisation should ask what information it needs to give the model in order to obtain an answer.

Working with a public judgment is not the same as working with sensitive internal documentation. Nor is sending preselected excerpts the same as transmitting complete original files.

An organisation's own architecture can provide a way to process certain information without handing it to an external model provider. But that possibility should not be confused with an automatic guarantee of confidentiality.

A system hosted on one's own infrastructure can still send information through connectors, tools, logs, backups or auxiliary services. And a poorly protected privately owned server can also be breached.

The question is not simply who owns the machine. It is who can access it, which programs are involved, which communications are permitted and how compliance with the boundaries is verified.

Control does not come bundled with the hardware: it has to be designed, operated and checked.

Second reason: turning the firm's knowledge into its own capability

Bloomberg reports that Latham aims to integrate AI with its internal systems and develop tools incorporating its professionals' expertise. This ambition goes beyond protecting data: it seeks to turn accumulated knowledge into a distinctive working capability. Bloomberg Law.

This is where the strategic question emerges.

If two firms subscribe to the same program, both gain access to a similar capability. That does not mean they will produce identical answers: their professionals, sources, instructions and review will still matter. But access to that tool ceases, by itself, to be an exclusive advantage.

Differentiation may shift elsewhere:

• how each firm organises its knowledge;
• which procedures it has formalised;
• which errors it knows how to detect;
• how it distributes tasks;
• and how it verifies and authorises the result.

A document library does not automatically become usable experience. It must be cleaned up, structured and connected to the work. The same applies to procedures: what a partner knows how to do intuitively needs to be made explicit before it can be automated with some reliability.

Building one's own system is therefore not simply a matter of installing a model. It means deciding which parts of professional practice can become reproducible processes without losing their exceptions or professional oversight.

Third reason: preserving choice

Technological dependence does not arise only when a supplier raises prices.

It can also arise when documents, evaluations, instructions, integrations and procedures become so closely tied to a platform that switching becomes extraordinarily difficult.

Owning infrastructure does not remove all dependencies either. It may shift them towards a hardware manufacturer, a family of models, particular software components or a handful of people who understand the system.

The reasonable objective is not absolute independence. It is to retain real alternatives.

That requires the ability to replace components without rebuilding the entire organisation: change models, move information, reuse evaluations and preserve authorisation rules even when the technology performing a task changes.

Architecture serves a business purpose here: preserving room to manoeuvre.

Fourth reason: controlling the system's economics

Latham acknowledges that token costs are a factor, although not the main driver of its decision. We do not have a public figure that would allow us to calculate the return on its investment. Bloomberg Law.

Broadly speaking, buying API access makes part of consumption a variable expense. Running one's own models shifts some of that expense towards equipment, staff and operations.

But running locally does not mean running for free.

The calculation must include hardware acquisition and renewal, electricity, maintenance, security, availability and specialist work. It must also include time spent solving problems that a supplier would handle in a purchased service.

For a large firm with a sufficient volume of suitable tasks, that investment may pay off. For an organisation that makes little use of the infrastructure, the opposite may be true.

The relevant comparison is not simply price per token. It is the total cost of producing a sufficiently reliable result, including verification and correction.

What requires a large firm today may become more accessible

That economic equation is not static.

Advances in hardware and software efficiency could allow some tasks that require expensive equipment today to run with more modest resources tomorrow. Quantisation, for example, seeks to represent parameters using less memory; other improvements make better use of equipment or employ specialised models. Every alternative has limits and must be evaluated for its effect on the quality of the work.

There is no need to run the world's most powerful model locally. What matters is having enough capability to perform a specific task well.

This opens an interesting possibility for smaller firms and businesses: affordable local capacity for certain processes and sensitive information, combined with external services for work requiring other capabilities.

We do not know when or how far those cost reductions will go. Requirements, usage volumes and supervision needs could also increase.

The conclusion is therefore not to buy oversized infrastructure today. It is to build modularly, so that adding better and cheaper local capability does not require starting from scratch.

What a small firm should learn

Copying the server purchase would confuse the visible decision with the reasoning behind it.

The starting point should be much simpler: identify a real process, understand its risks and establish what architecture it needs.

Comparing versions of public documents, organising internal knowledge and allowing an agent to send communications are different problems. They do not require the same resources or permit the same permissions.

Before investing, it is worth answering:

What do we want to improve? What information is involved? How will we measure quality? What must remain under our control? Who can stop the system?

Only then can we decide what to buy, what to configure and what to build.

Initial tests should use public or synthetic documentation. A working prototype does not yet demonstrate that a system is ready to receive sensitive information.

The news is not the GPUs

We do not know whether this investment will give Latham a lasting advantage. Nor would it be serious to present its decision as proof that every firm should follow the same path.

But it does highlight an important issue: adopting AI can become a decision about the organisation of the firm, not merely its tools.

What knowledge it retains.

What processes it develops.

What information it shares.

Whom it depends on.

And who controls decisions with effects outside the organisation.

My conclusion is not that law firms should become hardware companies. It is that they should avoid outsourcing their judgment about how to organise their own capacity to work.

The tool may belong to the provider. The architecture of the work and responsibility for its results must remain with the organisation.