13 With respect to large awards, we observe a handful of high-profile contracts for AI products, including a $250 million blanket purchase agreement issued to Scale AI in 2022, Project Maven, a 5-year $480 million contract awarded to Palantir, and most recently, Thunderforge, a large DIU program involving several major players including Scale AI, Anduril, Microsoft, and more. Thunderforge, specifically, is intended for the development and integration of generative AI and AI agents in operational and theater-planning military functions, and while the overall value is unspecified, the initial prototyping contracts are incorporating the largest AI developers, signaling potentially high value awards long term. CDAO also launched Rapid Capabilities Cell, a $100 million program designed to incubate new applications of AI tools. On the other hand, while contracts are already going to major generative AI developers and their partners, over 300 different contractors delivered smaller-scale AI contracts (both generative AI and other AI forms), most of which were between six and seven figures in value.22 These contracts are going to small and mid-sized businesses, as well as startups working on new and emerging applications of AI technology. The DOD’s contracting approach reflects a common theme among its historical acquisition patterns—a handful of large contracts issued to ‘primes,’ with set-asides for small and mid-sized organizations. This demonstrates a logical initial step to integrate AI into the DOD, balancing the need for leveraging market leaders with the need to cast a broad net in identifying and developing new technologies and applications. Long term, however, we believe the approach presents several risks that could be detrimental to the DOD’s long-term AI transformation, including Generative AI. We cite these risks below. Dependencies in the Tech Stack The DOD’s complex web of technology architectures and data systems may not be optimized to integrate generative AI tools, presenting a long-term challenge with acquiring the tools too soon. Key to understanding this challenge is how tools such as LLMs and AI Agents interact with an enterprise’s resources to function. LLMs ‘learn’ by repeated trial and error in responding to user requests. They pull data from massive data sets like the internet or internal databases, plot the information on a word map, and generate answers by pulling them from the map. Search engines send back raw results of the closest data found, while LLMs combine what was found with the user’s prompt to generate answers that, statistically speaking, are likely to answer the prompt coherently. AI Agents go further, determining if their model was trained on data sufficient to fulfill the request. If not, AI Agents then ‘self-prompt’ to act autonomously (hence being an ‘agent’) to resolve the request. This might entail pulling in data sources not already included on its language map through software applications called Retrieval Augmented Generation (RAG) or interacting with graphical user interfaces (GUIs) to perform historically human-driven work tasks. LLMs and AI Agents therefore depend on easily navigable, comparable, and trustworthy digital assets in an enterprise to perform accurately. This is where the problem arises for their rollout in the DOD, given its labyrinthine array of security requirements, digital systems and data repositories across the six service branches. A clear understanding of these architectures, and their maturity relative to other areas of the organization, is critical to understanding the technological dependencies that can either enhance or constrain a generative AI’s effectiveness. While DIU mentions assessing DOD technology architectures as part of its services, we believe that this should have elevated priority in the near-term, as foundational technology assets are modernized and aligned for Gen-AI. Further, extending these assessments beyond the scope of existing integration programs such as Thunderforge can allow DIU to preemptively identify, categorize, and prioritize areas of the DOD for modernization before programs reach those areas of the organization, which can allow for proactive and targeted investment decisions. Generative AI Adoption in the US Military