Up to 10 gigawatts. That’s the scale of the Ohio data center complex, the PORTS-Pike Technology Campus, that Latham & Watkins advised Nvidia on — one of the largest single data center developments planned anywhere. It’s the kind of number that normally lives in energy policy papers, not legal press releases. And now the same firm that helped paper that deal has bought Nvidia servers of its own, to run AI systems inside its own walls.
If you’ve been following AI mostly through chatbots in a browser tab, this move might seem like a strange detour. So let’s unpack what it actually means, and why it matters to anyone trying to understand where AI agents are heading.
What “in-house AI” actually means
Most people’s experience of AI is a website. You type, something types back. Behind that simplicity sits an enormous amount of hardware in someone else’s building, owned and operated by a technology company. You rent access. You don’t own the machine, you don’t set the rules, and you don’t decide what happens to your data beyond whatever the terms of service promise.
Buying your own servers flips that arrangement. The hardware sits in infrastructure you control. The models you run on it are your choice. The data never has to leave a boundary you define. In Latham’s own framing, the point is flexibility in how it develops, tests, and deploys these systems — as a partner at the firm put it, “Flexibility is critical.”
Think of it as the difference between renting a commercial kitchen by the hour and building one in your own restaurant. The rented kitchen is faster to start with and someone else fixes the oven. Your own kitchen means you decide the layout, the hours, the menu, and who gets to walk in the back door.
Why a law firm, of all places
Law is an unusually good fit for this kind of decision, and understanding why helps explain a pattern you’ll see repeat across other industries.
- The data is the business. A law firm’s files are other people’s secrets: deal terms, litigation strategy, regulatory exposure. Confidentiality isn’t a preference, it’s the product.
- Clients ask hard questions. Large corporate clients want to know exactly where their information goes and who can touch it. “It’s in the cloud somewhere” is a difficult answer to defend.
- The work is text-shaped. Contracts, briefs, memos, discovery documents. This is precisely the material language models handle well, which makes the potential payoff large enough to justify the cost of hardware.
Add those together and buying the machines starts to look less like a tech vanity project and more like a supply chain decision.
Where AI agents come into it
This is the part I find most interesting, and it’s why I think readers of this site should pay attention.
A chatbot answers a question and stops. An AI agent takes a goal and works through steps: reading files, comparing documents, pulling in relevant precedent, flagging inconsistencies, drafting something for a human to review. Agents are chatty in a very literal sense — they make many calls to a model, not one, and they touch far more of your internal data along the way.
That changes the math. If you’re running an agent that reads thousands of internal documents to do useful work, every one of those documents is travelling somewhere. Owning the infrastructure means you control that route. Renting it means negotiating it.
So when an organisation buys hardware rather than subscribing, it’s often a signal about ambition. You don’t need your own servers to run a search box. You might want them if you’re planning systems that work continuously across sensitive material.
The trade-offs nobody puts on a slide
Running your own AI setup is not simply the better option. It’s a different set of problems.
You need people who can operate the hardware. You need to keep the models current, which is a moving target. You pay upfront instead of monthly, and if your usage turns out to be lower than expected, you’ve bought capacity that sits idle. Cloud providers exist because most organisations would rather rent this complexity than employ it.
The firms that go in-house tend to be the ones where confidentiality requirements, sheer volume of work, and the money to hire specialists all line up. That’s a short list today. It may not stay short.
What to watch for
If you want a simple takeaway, it’s this: the question of where AI runs is becoming as consequential as which AI you use. For years the interesting news in this space was about model capability. Increasingly it’s about ownership, control, and location.
Latham is one firm making one decision. But the reasoning behind it — flexibility, independence, keeping sensitive material close — applies to hospitals, banks, government departments, and any business whose data is genuinely its own. Watch for who follows. That’s where the real story is.
🕒 Published: