A law firm just decided it would rather own the metal than trust someone else’s cloud, and that decision says more about the state of enterprise AI than any vendor keynote this year.
Latham & Watkins purchased Nvidia servers to stand up in-house AI systems. The stated reason, per the firm, is flexibility and independence in how it develops AI. Law.com quoted the framing directly: “Our infrastructure strategy gives us greater flexibility in how we develop, test, a—” and the sentence trails off in the excerpt, but the headline carried the point plainly enough. “Flexibility is Critical.”
That’s the whole story, factually. It’s a short story. But I review AI tools for a living, and I’ve watched enough enterprise pilots die quiet deaths to know when a small procurement decision is carrying a large amount of subtext.
What buying hardware actually signals
Nobody buys GPUs because it’s easy. Buying GPUs means someone has to rack them, cool them, patch them, secure them, and answer for them when they sit idle. The entire pitch of managed AI services is that you skip all of that. You get an API key, you get billed per token, you get to blame your vendor when something breaks.
So when an organization walks away from that convenience, it’s usually because the convenience came with strings it couldn’t accept. In the case of a law firm, the obvious candidates are not hard to guess:
- Client data that cannot sit on infrastructure the firm doesn’t control
- Confidentiality obligations that don’t have a “but the vendor promised” exception
- A desire to test models and workflows without asking permission or renegotiating terms
- Cost behavior that doesn’t scale unpredictably with usage
I want to be careful here. The verified reporting says flexibility and independence. It does not say confidentiality drove the purchase, and I’m not going to put words in anyone’s mouth. But “independence in AI development” is a phrase that only makes sense as a reaction to dependence, and dependence is exactly what the current generation of AI tooling sells.
The dependency problem nobody wants to price
Here’s what bothers me about most AI tools I test. The demo is great. The pricing looks fine at pilot scale. And then you discover that your entire workflow now lives inside someone else’s product roadmap. Models get deprecated. Rate limits change. Terms of service get revised. Prices move. Features you built on get folded into a higher tier.
For a consumer app, that’s annoying. For an organization whose product is judgment and confidentiality, it’s a structural risk. You cannot tell a client that their material was exposed because a vendor changed a data retention policy in a footnote.
Owning the hardware doesn’t eliminate that risk. It relocates it. You trade vendor risk for operational risk, and operational risk is expensive and boring and requires people who know what they’re doing. That’s a real trade, not a free win. But it’s a trade some organizations are now willing to make, and that’s new.
The part that made me laugh
Latham & Watkins also represented Nvidia in an AI compute hosting and guarantee transaction for the PORTS-Pike Technology Campus in Ohio, described as a development of up to 10 gigawatts and one of the largest single data center complexes planned. The transactional team was led by partners Haim Zaltzman, Aida Vajzovic, Joshua Dubofsky, and Dan Van Fleet, with associates including Ross Wasserman.
So the firm advising on one of the largest planned data center complexes in the world also decided it wanted its own servers in the building. Draw whatever conclusion you like. Mine is that when you spend enough time inside the paperwork of the AI buildout, you develop opinions about where you want your own workloads to sit.
What this means if you’re evaluating AI tools
Most organizations should not buy GPUs. That needs saying, because procurement decisions at large firms have a way of becoming aspirational strategy at smaller ones. If you have a dozen people and a modest document workflow, buying servers is a way to spend money and acquire problems.
What you should take from this is narrower and more useful. Ask the questions this purchase implies. Where does your data actually go. What happens if the model you built on disappears in twelve months. What does your cost curve look like at ten times current usage. Can you switch providers without rebuilding, and have you ever tested that.
The tools I recommend are the ones that survive those questions. Most don’t. A firm with resources and legal obligations looked at the same questions and concluded it wanted more control over the answers. You probably can’t afford its solution. You can absolutely afford its skepticism.
🕒 Published: