Picture the procurement meeting. Someone from IT has a slide up showing next year’s projected token spend. Someone from the practice side is asking whether client documents ever leave the building. And someone in the corner asks the question that changes the shape of the whole conversation: what if we just bought the hardware?
At Latham & Watkins, that question apparently got a yes. The firm has purchased Nvidia servers to stand up in-house AI systems, according to the Financial Times, with the stated goal of greater flexibility and independence in how it builds AI. Law.com quoted the firm’s framing directly in its headline: “Flexibility is critical.”
I review AI tooling for a living, which mostly means I spend my days poking at APIs, dashboards, and pricing pages. So my first reaction to this news was not about law firms at all. It was about the buy-versus-rent decision that every team using AI tools eventually runs into, and what it means that a firm this size just picked buy.
What “flexibility” actually means when you own the box
Vendors love the word flexibility. It usually shows up in a marketing paragraph about how many integrations they support. Owning your own compute means something narrower and more concrete.
- You choose the model, including the ones nobody is hosting for you anymore.
- You choose when to upgrade, which means you also choose when not to.
- You control where the data sits, physically, and can point at a rack when someone asks.
- Your unit economics stop moving every time a provider adjusts a price sheet.
That last one deserves attention. The thing I hear most often from teams running AI in production is not “the model isn’t good enough.” It’s “we can’t forecast the bill.” Metered inference is wonderful when you’re experimenting and miserable when you’re budgeting. Hardware flips the problem: a big number up front, then a curve you can actually draw. For an organization built around billable hours and predictable cost structures, that trade probably reads as obvious.
The independence part is the interesting half
Flexibility is the word Latham used. Independence is the word doing more work here. Right now, most AI capability in most organizations is rented from a handful of providers. That’s fine, until a provider deprecates a model you built a workflow around, or changes terms, or has an outage during a filing deadline.
Owning infrastructure doesn’t eliminate that dependency. You’re still buying Nvidia hardware, still probably running models someone else trained, still patching a stack you didn’t write. But it moves the failure modes into your own building, where you can do something about them. For a firm whose product is confidential judgment applied to sensitive material, that relocation has obvious appeal.
There’s a small irony worth pointing out. Latham also represented Nvidia on an AI compute hosting and guarantee transaction for the PORTS-Pike Technology Campus in Ohio, a development of up to 10 GW and one of the largest single data center complexes planned anywhere. The firm has been reading the fine print on AI infrastructure deals from the inside. It’s not shocking that some of that knowledge turned into a purchase order.
What I’d tell a smaller team watching this
Don’t copy it. Not yet, anyway. A firm of Latham’s scale has the volume to amortize hardware, the IT staff to run it, and the security requirements to justify it. Most teams have none of those three, and buying GPUs to solve a problem that a well-negotiated API contract would solve is an expensive way to feel in control.
What I would take from this is the habit of asking the question. Specifically:
- What is our monthly inference spend, and what does it look like at 5x current usage?
- Which vendor decisions could break our workflows tomorrow without our input?
- What data are we sending out that we’d rather not, and is that a compliance risk or just a discomfort?
- If we owned the stack, who on the team would actually operate it?
If the answers point toward rented infrastructure, stay there and negotiate harder. Renting is the right default for the overwhelming majority of teams, and the tooling around hosted models is far more mature than the tooling around running your own.
The signal underneath the story
What makes this newsworthy isn’t the hardware. It’s that a professional services firm treated AI compute as infrastructure rather than software. Software you subscribe to. Infrastructure you own, staff, depreciate, and plan around for years.
That reclassification is the part I’d watch. When AI stops being a line item under SaaS and starts appearing under capital expenditure, it changes who makes the decisions, how long the commitments last, and how much switching costs matter. Latham’s purchase is one firm, one set of servers, one strategy nobody outside the building has audited. But it’s a real data point in a debate most teams are still avoiding, and I’d rather see someone commit and report back than watch another year of pilots that never leave the sandbox.
🕒 Published: