Big Law Is Building Its Own AI Infrastructure

For most of the LegalTech era, the relationship between law firms and technology companies was straightforward. Vendors built software, law firms bought it, and internal technology teams focused largely on implementation, integration, security and adoption. Even as generative AI arrived, that basic division of labor largely survived: firms purchased Harvey, Legora, Microsoft Copilot, ChatGPT Enterprise and other platforms while specialist vendors handled the underlying models and infrastructure. Latham & Watkins has now provided one of the clearest signs yet that this arrangement may be starting to change.

According to reporting by the Financial Times and Bloomberg Law, Latham has purchased Nvidia GPU infrastructure, leased secure third-party data-centre capacity accessible only to its personnel, and begun fine-tuning and customizing open-weight models for internal use. The firm is not abandoning commercial AI products; it continues to use tools from companies such as OpenAI, Anthropic and Google. Instead, it is creating another option: an internal AI layer that gives the firm greater control over particularly sensitive data, model selection and the economics of AI consumption. Financial Times; Bloomberg Law.

That distinction matters. Latham is not trying to train the next GPT from scratch, nor is it suddenly becoming a conventional software company whose primary business is selling technology. What it is doing is strategically more interesting: building enough internal AI infrastructure to decide when it wants to rely on external providers and when it wants to run, customize or control systems itself. The largest law firms may therefore be moving from a simple build-versus-buy decision toward a much more sophisticated model of buy, build and orchestrate.

What Latham Actually Built

Latham’s move is more substantial than experimenting with an internal chatbot. Bloomberg Law reported that the firm has been developing its Nvidia server infrastructure for approximately three years and has purchased multiple Nvidia H200 GPUs, while considering newer Blackwell and Vera Rubin systems. Those servers operate in a third-party data-centre facility to which only Latham personnel have access. The firm is using that infrastructure to fine-tune open-weight models and create an internal alternative to relying exclusively on cloud-based AI services. Bloomberg Law.

The Financial Times also reported that Latham now employs more than 900 technology specialists, including machine-learning and AI engineers. That scale is critical because hardware alone does not produce useful legal AI: the difficult work lies in securing the infrastructure, preparing data, managing permissions, choosing models, evaluating outputs, building workflows and maintaining the systems over time. A firm capable of employing hundreds of technologists can make investments that would be unrealistic for the vast majority of legal practices. Latham’s experiment therefore tells us as much about the economics of the largest global firms as it does about AI itself. Financial Times.

Why Would a Law Firm Want Its Own AI Infrastructure?

The first answer is confidentiality. Latham’s chief information officer, Rene Mendoza, told the Financial Times that certain client information may be sufficiently sensitive that the firm does not want it placed with any cloud vendor. Enterprise AI providers have invested heavily in security, contractual protections and no-training commitments, so this should not be read as an indictment of cloud AI. The point is that a private infrastructure option allows the firm to make a different decision when a particular client, matter or category of information demands it. Legal IT Insider.

The second answer is cost. Generative AI is moving from occasional prompting toward large-scale document analysis, background agents and workflows that can run for hours, which means inference expenditure can become material at enterprise scale. Mendoza specifically referenced the prospect of increasing consumption costs as a reason for wanting flexibility. A firm with sufficient utilization may eventually find that owning some compute changes the economics of particular high-volume tasks, just as large enterprises sometimes find private infrastructure economical for predictable workloads even while continuing to use the public cloud.

The third answer is control over model choice. Open-weight models can be downloaded, hosted on infrastructure controlled by the firm and customized for specific use cases, giving firms more flexibility over configuration, deployment and data handling. Goldman Sachs CIO Marco Argenti made a similar argument outside the legal sector this week, saying enterprises should not rule out open models and emphasizing the importance of model choice where appropriate security safeguards exist. Deloitte likewise launched an Open Model Engineering practice this month, citing flexibility, cost predictability, sovereignty and control as reasons enterprises are increasingly evaluating mixtures of open and proprietary models. Axios; Deloitte.

Finally, there is strategic independence. A firm that relies entirely on external AI vendors is exposed to changes in pricing, product roadmaps, model availability and technical architecture that it does not control. That does not make vendor dependence inherently problematic; modern enterprise software is built on layers of external services. But for a law firm generating billions of dollars in revenue and increasingly treating AI as core infrastructure, the ability to preserve an internal alternative can become a strategic asset rather than an engineering curiosity.

Why Open-Weight Models Change the Build-versus-Buy Equation

A decade ago, the idea that a law firm might operate sophisticated AI internally would have required an extraordinary research organization. Frontier-scale model training still does. The important change is that firms no longer need to build the foundation model themselves in order to control more of the AI stack. Capable open-weight models can be deployed on privately controlled infrastructure, then adapted, evaluated and connected to proprietary firm knowledge without requiring a law firm to spend billions of dollars training a model from zero.

This changes the economics of internal development because the law firm can concentrate on the parts where it possesses genuine competitive advantage. Latham’s lawyers know how Latham performs transactions, structures advice, manages risk and serves its clients; its internal data contains precedents and institutional knowledge accumulated through years of high-end legal work. Those assets are more distinctive than the base model itself. The strategic question therefore becomes whether the firm can combine commercially available model intelligence with proprietary legal knowledge in a way that produces something competitors cannot simply purchase from the same vendor.

It also changes vendor bargaining power. If a large firm knows that certain workflows can be moved onto an internal open-weight system, it is less dependent on a single commercial model provider’s pricing or product roadmap. That does not mean the internal system will always be cheaper or better, because maintaining GPU infrastructure, engineering teams and security controls is expensive. The value lies in having a credible alternative rather than assuming every legal AI workload must pass through the same external stack.

Latham Is Not the Only Firm Moving From Buyer to Builder

Latham’s hardware investment is unusual, but the broader move toward proprietary legal AI is already visible elsewhere in Big Law. In August, Goodwin launched Regina OS, an internal AI platform designed to connect firm data, workflows, models, governance and user experiences. The firm is investing approximately $25 million annually in buying and developing AI and has described its ambition as moving from a technology-enabled firm toward an AI-native one. Goodwin.

Goodwin’s approach is particularly revealing because it shows that “building your own AI” does not necessarily mean building models from scratch. Regina OS draws on technologies from Microsoft, Anthropic and OpenAI while adding a proprietary layer connecting those models to Goodwin’s institutional knowledge and legal workflows. The firm also created an automated venture-financing tool using a customized version of Claude, with associates participating directly in development. Goodwin hopes to extend the model into areas including M&A, private equity, litigation and fund formation. Goodwin.

This distinction is crucial because it means the relevant trend is not “law firms will replace LegalTech companies.” The more plausible trend is that sophisticated firms will increasingly assemble their own technology architectures from commercial models, LegalTech products, internal data and proprietary workflows. The law firm becomes less like a passive software customer and more like a systems integrator with a meaningful internal product-development capability. That is a very different kind of customer for LegalTech vendors to sell to.

So, Should LegalTech Vendors Be Worried?

Some should be. The vendors most exposed are those whose product is little more than convenient access to a general-purpose model wrapped in a legal interface. If a large firm’s internal technology team can reproduce most of the value by connecting an open model or commercial API to internal documents and prompts, the vendor has limited defensibility. As firms hire more engineers and lawyers learn to build internal applications with coding agents, the cost of reproducing simple AI features continues to fall.

But the conclusion that Big Law will therefore stop buying LegalTech would be a mistake. Running private AI infrastructure is expensive, operationally demanding and outside the core business of most law firms. Even Latham is maintaining access to commercial AI products rather than attempting to internalize everything. The more likely outcome is that vendors will be forced to prove that they provide value a law firm cannot economically or sensibly reproduce itself.

Research providers such as Thomson Reuters and LexisNexis possess legal content and citation systems that are extraordinarily expensive to recreate. Specialist platforms such as Relativity own deeply developed workflows and ecosystems in complex areas such as e-discovery. AI-native companies such as Harvey and Legora are increasingly investing in workflow orchestration, knowledge integration, evaluation, governance and enterprise distribution rather than selling raw model access. These are precisely the layers that become more valuable as foundation models themselves become more interchangeable.

There is also a counterintuitive possibility: firms building more internally could create a larger market for sophisticated LegalTech infrastructure. A firm with its own AI engineering team does not necessarily want to build every connector, evaluation framework, document parser, research database, security control or workflow component from scratch. Technically mature customers often purchase more infrastructure, not less; they simply purchase it differently. LegalTech vendors may therefore find themselves selling APIs, components and platform capabilities to law firms that increasingly behave like enterprise technology organizations.

The Hybrid Model Is Probably the Real Future

The most realistic future is not Big Law becoming a collection of software companies. It is a hybrid architecture in which firms buy commodity capabilities, partner for specialist capabilities and build only the layers that create meaningful differentiation or control. Commercial frontier models may remain the best option for cutting-edge reasoning tasks, while private open-weight models handle selected sensitive or high-volume workloads. Specialist LegalTech products may continue to provide research, discovery, transaction and workflow functionality that would be irrational for firms to recreate independently.

The proprietary work is likely to concentrate around the firm’s own knowledge and processes. That includes retrieval across internal precedents, matter-specific context, practice-group workflows, permission structures, evaluation benchmarks and increasingly the agentic orchestration connecting those elements. In other words, firms may not want to own the intelligence commodity itself as much as they want to own how that intelligence is applied to their unique legal practice. That is where institutional knowledge can become technological differentiation.

This also explains why the “legal operating system” race is becoming more complicated. Vendors want to become the orchestration layer through which lawyers work, but firms such as Goodwin are explicitly building their own internal orchestration layers. If the largest firms succeed, LegalTech companies may have to accept that they will not always own the full user environment. Their products may instead need to plug into firm-controlled systems and compete to become the best component inside a broader architecture.

The LegalTech Moat Moves Up the Stack

The implications for LegalTech founders are significant. Access to a frontier model is becoming a weaker moat because customers can increasingly access comparable intelligence through multiple providers or deploy capable open-weight alternatives. A product differentiated primarily by model access can be squeezed from both directions: foundation-model companies can move directly into legal workflows, while sophisticated law firms can build selected capabilities internally. The durable moat therefore has to move above the model.

That moat may be proprietary legal data, trusted content, deeply embedded workflows, regulatory expertise, integrations, institutional adoption, benchmarking, governance or distribution. It may also be accumulated product knowledge about a particular legal process that would take an internal team years to replicate reliably. Vendors that can demonstrate those advantages have little reason to panic about Latham buying GPUs. Vendors that cannot may find the next procurement conversation considerably harder.

There is a parallel lesson for law firms. Owning hardware or writing internal software does not automatically create competitive advantage. A proprietary tool only matters if it improves quality, speed, cost, client service or institutional learning in a way that commercial alternatives cannot. Firms should be as skeptical of internal technology projects as they are of vendor marketing, because building software can become an expensive distraction when the differentiation is illusory.

Which Law Firms Can Realistically Follow Latham?

The answer is: not many, at least not at the same infrastructure depth. Latham generated approximately $8.3 billion in revenue last year according to the Financial Times and maintains a technology workforce numbering in the hundreds. Purchasing GPU servers is only the visible capital expenditure; operating them requires engineering talent, cybersecurity, model operations, evaluation processes, data governance and ongoing maintenance. For most firms, commercial cloud and LegalTech products will remain substantially more economical.

But firms do not need to copy Latham’s hardware strategy to follow the broader direction. A mid-sized firm can build an internal retrieval layer over its precedents, develop proprietary workflows using commercial models, create evaluation datasets for its key practice areas or expose internal knowledge to AI through controlled APIs. Goodwin’s example demonstrates that the proprietary layer can sit above external models without requiring the firm to own the compute. The broader shift from passive purchasing to active architecture is therefore accessible to far more firms than the data-centre headline initially suggests.

We may consequently see a segmentation of the market. The very largest global firms may operate hybrid private-and-cloud AI infrastructure, a second tier may build proprietary orchestration and workflow layers on commercial models, and smaller firms may rely primarily on integrated LegalTech platforms. That would mirror enterprise technology markets more generally, where the degree of internal engineering increases with organizational scale and strategic need. LegalTech vendors will need products flexible enough to serve customers across those different levels of technical sophistication.

What Should LegalTech Vendors Do Now?

The first priority is interoperability. Vendors should assume that sophisticated customers will increasingly want to connect products to their own data, agents, models and workflows rather than live permanently inside a closed application. APIs, exportability, permission-aware connectors and support for multiple models can therefore become competitive differentiators. A product that integrates cleanly into a firm-controlled architecture may outperform a theoretically more powerful product that demands complete platform dependence.

The second priority is to invest in what customers cannot easily build. That means authoritative data, difficult integrations, domain-specific evaluation, specialized legal workflows, compliance infrastructure and accumulated market knowledge. The cost of generating text is falling rapidly, but the cost of building trusted professional infrastructure remains high. Vendors should therefore be careful about treating every new model capability as a product moat when the same capability may become broadly available within months.

The third priority is partnership. Large law firms building internal AI teams will still need external expertise and technology, but they may want a different relationship with vendors. The winning vendor may increasingly behave less like a sealed software supplier and more like an infrastructure partner that allows the firm’s own engineers and lawyers to build on top of the platform. This is a harder product strategy, but it aligns with where sophisticated enterprise customers appear to be heading.

1. Design products so firms can connect their own knowledge, permissions and workflows without surrendering architectural control.
2. Support model choice where possible instead of assuming one foundation model will remain dominant indefinitely.
3. Invest in domain-specific evaluation and reliability, because firms building internally will compare vendor performance against their own benchmarks.
4. Make data portability and APIs strengths rather than reluctantly provided enterprise features.
5. Build defensibility above the model through content, workflow depth, integrations, governance and distribution.

Big Law May Become Partly a Technology Industry

The most interesting implication of Latham’s move is not that LegalTech vendors are about to disappear. It is that the boundary between law firm and technology company is becoming less clean. Firms such as Latham and Goodwin are hiring engineers, building platforms, creating AI workflows and making infrastructure decisions that would once have belonged almost entirely to software companies. Their primary product remains legal advice, but technology is increasingly becoming part of how that product is designed and delivered.

This could alter competition between law firms themselves. If two firms have access to the same commercial AI products, neither gains much durable advantage merely by purchasing them. A firm that combines those tools with superior proprietary workflows, institutional data, internal engineering and client-specific systems may create something more difficult to copy. Technology therefore moves from being an operational support function toward becoming part of the firm’s competitive strategy.

Clients may eventually encourage this change. Sophisticated corporate clients already ask about information security, efficiency and AI usage, and some will increasingly expect firms to demonstrate how technology improves delivery rather than simply naming the products they license. A firm capable of building secure client-specific workflows or operating sensitive AI privately may be able to offer something materially different from a competitor using the same off-the-shelf stack. That creates a commercial reason for at least some firms to invest well beyond standard LegalTech procurement.