Latham & Watkins advised Nvidia on an AI compute hosting deal for a data center campus in Ohio sized at up to 10 gigawatts. The same firm just bought its own Nvidia servers to run AI in-house.
One of those facts is a law firm doing law firm things. The other is a law firm deciding that renting inference from someone else’s API wasn’t good enough. As someone who spends most days wiring bots together, the second fact is the one I keep turning over.
What the firm actually said
The stated reason, per Law.com, is flexibility. Their infrastructure strategy gives them more room in how they develop and test. Reported alongside that: independence in AI development. That’s the whole substance of the news, and it’s enough.
Because “flexibility is critical” is not a press release phrase when you’ve actually built something. It’s the exact sentence you say after your third rate limit, your second surprise deprecation notice, and the day a model provider silently changed behavior on a prompt chain you’d spent a month tuning.
Why owning metal changes how you build
Most bot architecture today assumes a hosted model at the far end of an HTTP call. That assumption shapes everything downstream:
- You design around token budgets instead of compute budgets
- You batch requests to dodge rate limits rather than to fill a GPU
- You treat the model as a fixed dependency and build all your cleverness in the orchestration layer
- You accept that fine-tuning means uploading your data somewhere
Flip to owned hardware and each of those inverts. Compute becomes a fixed cost you’ve already paid, so the marginal call is nearly free. That single change rewrites your retrieval strategy. When every query costs money, you build tight retrieval and send the smallest possible context. When compute is sunk cost, you can afford to run five retrieval passes, rerank aggressively, and let the model see far more before it answers.
Same with evaluation. If you’re paying per token, thorough eval suites get trimmed to whatever the budget tolerates. On your own boxes, you can run the full suite on every commit. That’s the difference between shipping on vibes and shipping on numbers, and it’s mostly an economics problem disguised as a discipline problem.
The confidentiality angle nobody can hand-wave
A law firm has a structural reason to care where data sits that most of us can gesture at but not really claim. Client privilege isn’t a preference. It’s the product. If you’re building bots for healthcare, finance, or anything touching regulated records, the Latham move is a useful reference point: the compliance answer for “where does the prompt go” gets a lot shorter when the answer is “that rack.”
This is the part I’d push builders to think about before they think about GPUs. Local inference is not primarily a performance play. For most workloads, a hosted frontier model will beat what you can serve yourself. The trade you’re making is capability for control. Whether that trade is smart depends entirely on how much your control problem costs you today.
What this does not mean
It does not mean you should buy servers. The failure mode here is obvious and I’ve watched people walk into it: hardware sitting idle because nobody built the serving layer, the model registry, the monitoring, or the deployment pipeline around it. GPUs are the cheap part of self-hosting. The expensive part is the platform work, and a firm the size of Latham has staff who can absorb that. A three-person bot shop generally does not.
The portable lesson is architectural. Build so the model is swappable. That means a thin adapter between your orchestration logic and whatever generates tokens, prompt templates stored as data rather than baked into call sites, and an eval use that runs against any endpoint you point it at. Do that, and moving from a hosted API to a local model later is a config change instead of a rewrite.
Most teams skip this because the adapter feels like premature abstraction when you only have one provider. Then the provider changes something and the abstraction stops feeling premature.
The signal worth reading
What makes this story interesting isn’t the hardware. It’s who bought it. A law firm is not a tech company. When an organization whose core business is judgment and confidentiality decides it wants AI infrastructure under its own roof, that says something about how central these systems have become to the actual work. You don’t buy servers for a pilot project.
Latham advised Nvidia on a data center deal, then bought Nvidia servers. Somewhere in between, someone looked at the arrangement and decided both sides of it were worth being on.
🕒 Published: