Our Take
Kirkland is building in-house for control and data privacy, not because third-party tools can't do the work—the bet is that owning the stack beats customizing someone else's.
Why it matters
This signals how tier-one law firms now see legal AI as a core competitive asset worth a half-billion dollars and real engineering talent, not just a procurement decision. It also shows which firms can afford to go vertical versus those stuck with off-the-shelf layers.
Do this week
General counsel: audit your firm's legal AI spend and vendor lock-in before Kirkland-sized competitors finish their builds; know whether you're renting or owning.
Kirkland's GPU and Workflow Ambition
Kirkland & Ellis has posted roughly 85 AI-related roles across the United States as part of its $500 million innovation project, with two newly listed AI Infrastructure Director positions (Houston and Chicago, $302,000–$335,000 salary range per company posting) that specify expertise in "on-premise GPU environments" and Microsoft Azure ML services. These roles sit alongside dozens of AI Innovation Adviser positions requiring hands-on experience with Harvey, Legora, CoCounsel, and Lexis+ AI platforms, as well as workflow mapping and prompt engineering at scale.
The job specifications indicate Kirkland intends to design and manage its own on-premises GPU infrastructure to fine-tune open-source language models using the firm's proprietary legal data. The Infrastructure Director role explicitly covers "AI Infrastructure Ownership," "Innovation Enablement" through "secure, governed environments for experimentation," and "Platform Design & Delivery" for custom AI solutions. The AI Innovation Adviser position, by contrast, focuses on embedding within practice groups to translate legal workflows into scoped AI solutions, with direct accountability for output quality.
Kirkland has stated it will deploy a team of around 180 people to execute this project, though hiring remains ongoing. Job postings date back to March, with new roles added every few days.
The Privacy and Control Calculation
Building an internal LLM stack demands rare expertise and capital. Most law firms license existing platforms from Thomson Reuters, LexisNexis, or specialized legal AI vendors. Kirkland's approach trades that simplicity for two potential advantages: data isolation and operational control.
On-premises GPU infrastructure running a fine-tuned open-source model means client work and proprietary legal reasoning remain within the firm's own network, not routed through a vendor's cloud or shared tenant infrastructure. For a firm handling sensitive M&A, litigation discovery, and high-stakes regulatory work, that architectural difference can matter to clients and risk committees, even if the actual output quality converges with customized third-party solutions over time.
The tradeoff is real. Building and maintaining custom AI infrastructure requires sustained engineering investment. Kirkland's scale (roughly 2,600 lawyers globally) and capital position make this viable; few other firms can absorb that cost.
Thomson Reuters, for comparison, is fine-tuning open-source models for its own products but retains a cloud-based distribution model. Kirkland's choice to keep inference and training on-premises is architecturally different and signals confidence in owning the full software stack end-to-end.
What Law Firms Should Consider Now
For general counsel and legal operations leaders at firms evaluating AI investment, the message is clear: there is no single path. Kirkland is betting on vertical integration and data sovereignty. Other tier-one firms (Latham & Watkins, Skadden) are layering third-party platforms with custom point solutions and have embedded those workflows deeply. The latter approach is faster and cheaper; Kirkland's is more defensible if clients or regulators scrutinize where legal data flows.
Mid-market and smaller firms should inventory current vendor commitments and contractual lock-in terms before Kirkland-tier competitors finish their builds and potentially offer data-sovereign services as a client differentiator. The question is not whether to adopt legal AI (you should), but whether to own the infrastructure or rent the capability. Your choice now shapes your options in 12 to 24 months.