Why Enterprises May Be Entering an AI ‘Build and Buy’ Era

Photo By: Igor Omilaev Latham & Watkins recently made an unusual investment for a law firm: its own Nvidia-powered AI servers. The firm is using that infrastructure to fine-tune open-weight AI models while continuing to use commercial tools such as ChatGPT, Claude and Gemini. Rather than replacing outside AI providers, Latham is building an internal option alongside them, giving its engineers more control over how certain models are customized and deployed. That distinction may matter beyond the legal industry. As AI becomes embedded in increasingly sensitive and specialized business processes, enterprises may be entering a new phase of AI adoption in which the question is no longer simply whether to build or buy. Instead, companies may increasingly combine the two: buying general-purpose capabilities while building or controlling the layers that are most important to their own businesses.

The Build-or-Buy Question Is Becoming More Complicated For much of the generative AI boom, the enterprise model was straightforward. A company could access increasingly capable models through an API or software platform without having to operate the underlying computing infrastructure. For many applications, that still makes sense. But the calculus changes when AI becomes deeply connected to proprietary data, specialized workflows or consequential business decisions. An enterprise may care not only about what a model can do, but also where it runs, how it can be customized, how much control it has over its data and how dependent it becomes on a particular provider. That is one reason open-weight models have attracted growing enterprise attention. Recent analysis from the Financial Times points to their appeal around customization, cost and data privacy, while Gartner has identified open-weight models as an increasingly relevant part of enterprise AI strategies. The result is less a rejection of commercial AI than a diversification of the stack.

Buy the Commodity. Build the Differentiator. The emerging enterprise AI architecture may therefore look less like a binary choice and more like a portfolio. A company might use a frontier commercial model for a general reasoning task, an open-weight model for a specialized workflow, its own infrastructure for particularly sensitive workloads, and custom software to connect those systems to proprietary data. The objective isn’t necessarily to build a better foundation model than the companies spending billions of dollars to develop them. It is to control the parts of AI that are strategically important to the business. That can include the data layer, evaluation systems, workflow orchestration, domain-specific customization or the infrastructure required for particular workloads. For some companies, the case for owning infrastructure may be compelling. For others, the capital, engineering and operational requirements will make an external provider the more rational choice. The important question is therefore not simply, “Should we build our own AI?” It is “Which parts of our AI stack are important enough to own?”

Why Latham’s Strategy Is Interesting Latham offers a useful case study because its approach doesn’t require choosing one side. The firm has reportedly spent several years developing its own Nvidia server infrastructure and is using it to fine-tune open-weight models, while retaining access to commercial AI systems. The strategy gives the firm additional flexibility in choosing the right model for different tasks, while keeping particularly sensitive work within the infrastructure it controls. That creates something enterprises increasingly value: optionality. If a commercial model is the best tool for a particular task, the company can use it. If a workload requires greater customization or control, the company has another route. This also changes the way enterprises can think about vendor dependence. Instead of locking an entire AI strategy into one provider, companies can build an architecture in which different models and deployment environments serve different purposes. That doesn’t mean every company should start buying GPUs. It means that as AI becomes more strategically important, some enterprises may decide that having an internal capability is valuable even if they continue buying most of their AI externally.

From General AI to Specialized Enterprise Intelligence The same logic applies at the application layer. The most important enterprise AI systems may not be the ones that simply provide access to the most powerful general-purpose model. They may be the systems designed around a specific business problem, proprietary data environment and decision-making process. Kapnova is one example of this approach. The company describes itself as an agentic revenue and profit optimization system for consumer brands, combining AI-driven signal discovery with quantitative methods designed to evaluate commercial decisions. Its platform is built around questions such as pricing, launches and marketing spend rather than general-purpose conversation. That reflects a broader shift: enterprises don’t necessarily need to own the entire AI stack to create specialized intelligence. They can combine general-purpose models, proprietary data, specialized quantitative systems and domain-specific software into something tailored to their own operating environment. For Kapnova CEO and co-founder James Sun, that distinction is central to how businesses should think about AI: the model is only one component of a much larger decision system.

The AI Stack May Become Hybrid by Design Latham’s investment doesn’t establish that enterprises are about to become AI infrastructure companies. Nor does it mean commercial models are losing their relevance. It points to a more nuanced possibility. As AI becomes increasingly embedded in business operations, enterprises may become more selective about what they buy, what they customize and what they control themselves. The next phase of enterprise AI may therefore not be defined by whether a company chooses to build or buy. It may be defined by knowing which parts of the AI stack are worth owning, and which are better left to someone else.