Photo By: Immo Wegmann
Latham & Watkins has made an unusual investment for a law firm: its own AI infrastructure.
The firm has purchased Nvidia-powered servers and is using them to fine-tune open-weight AI models, while continuing to work with commercial AI providers. The move gives Latham an internal option for workloads where it wants greater control over its infrastructure, models and sensitive information.
The interesting question is not whether every enterprise will follow Latham and start buying GPUs. It is what the decision says about a changing definition of ownership in enterprise AI.
For years, companies could largely treat AI as software: choose a provider, connect the relevant data and consume the capability. As AI becomes embedded deeper into business operations, that model is becoming more complicated. The technology stack now includes infrastructure, models, proprietary data, workflow systems and the applications that turn model outputs into business decisions.
Enterprises may not need to own all of those layers. But they increasingly have to decide which ones matter enough to control.
AI Is Becoming a Stack, Not a Tool
The easiest way to think about enterprise AI is no longer simply as a model.
Underneath an AI application sits an infrastructure layer: computing resources, deployment environments and data systems. Above that sits the model itself, which may be proprietary, open-weight or some combination of both. Then come the systems that connect the model to company data, business rules, workflows and human decisions.
That creates multiple places where an enterprise can choose to buy, build or customize.
A company might rent computing infrastructure but own its application. It might use a commercial foundation model while keeping proprietary data inside its own environment. It might fine-tune an open-weight model for a specialized workflow without attempting to develop a foundation model from scratch.
The result is that “owning AI” no longer has a single meaning.
The Model May Not Be the Most Valuable Layer
The AI industry often frames competition around which company has the most capable model.
For an enterprise, the more consequential question may be different: What part of the system contains knowledge or capabilities that competitors cannot simply purchase from the same provider?
That could be proprietary data. It could be a workflow built around years of operational experience. It could be a company’s internal evaluation and orchestration systems. Or it could be a specialized decision engine that combines AI with domain-specific quantitative methods.
This distinction matters because access to the same frontier model does not give every company the same competitive advantage. If thousands of companies can access a model, the differentiation may increasingly come from what each company builds around it.
That is also why open-weight models are becoming part of the enterprise conversation. Gartner’s September 2026 research identifies open-weight models as an option for enterprises seeking greater cost efficiency, control and innovation across different use cases.
The attraction isn’t necessarily independence for its own sake. It is the ability to make more deliberate choices about where intelligence lives and how it is deployed.
Ownership Does Not Have to Be All or Nothing
Latham’s strategy is notable precisely because it doesn’t appear to be an attempt to replace commercial AI altogether.
Bloomberg Law reported that the firm began developing its own Nvidia server infrastructure several years ago and is using it to fine-tune open-weight models alongside off-the-shelf AI tools. That gives Latham additional flexibility in determining which technology is appropriate for different tasks.
That creates a layered approach to ownership.
Commercial models can handle workloads where general-purpose capabilities are sufficient. Internal infrastructure can support applications where control or customization matters more. Open-weight models can provide another option when a company wants to modify or deploy a model within an environment it controls.
The same principle could apply across other industries.
A financial institution may care most about controlling sensitive data and risk systems. A manufacturer may prioritize integration with proprietary operational data. A consumer brand may care less about owning the underlying model than about owning the decision system that understands its pricing, customers and demand.
The strategic question is therefore not simply: Should we build AI or buy AI? It is: Which layers of AI are strategically important enough that we should control them ourselves?
The Rise of Specialized Enterprise Intelligence
That question also changes how companies should think about AI applications.
General-purpose models are designed to serve many users across many contexts. Enterprise decision systems can be designed around a much narrower set of problems.
Kapnova, for example, describes its platform as a decision-intelligence system for consumer brands that uses specialized agents to identify opportunities across pricing, promotions, marketing, inventory and demand, while quantitative models determine the value of potential decisions.
The distinction is important. Kapnova does not need to own the entire AI infrastructure stack to build specialized intelligence for consumer brands. Its differentiation comes from the system surrounding the models: proprietary business data, outside market signals, specialized quantitative methods and a workflow designed around revenue, gross profit and contribution-margin decisions.
That illustrates another possible form of AI ownership: owning the intelligence layer that makes general technology useful to a particular business.
So, Who Should Own the AI Stack?
There may not be one answer.
Some companies will have little reason to operate their own infrastructure. The technical and financial burden may outweigh the benefits.
Others may find that certain workloads are too sensitive, strategically important or specialized to leave entirely in the hands of an external provider.
And many may settle somewhere between those extremes: commercial models for some tasks, open-weight models for others, proprietary data and workflows throughout, and internal systems controlling the pieces that matter most.
Latham’s decision does not establish that enterprises are about to become AI infrastructure companies. It does, however, demonstrate that some organizations are beginning to treat control over parts of the AI stack as a strategic consideration rather than simply an IT procurement decision.
The next enterprise AI question may therefore not be who has the best model. It may be who owns the parts of the system that make that model valuable to the business.
