By Scale AILearning Objectives
- Understand what Scale AI's Donovan and Thunderforge products do for national security
- Evaluate the use of large language models inside classified planning environments
- Assess the automation-bias and accountability risks of AI in military decision support
What Are Donovan and Thunderforge?
Donovan and Thunderforge are the defense products of Scale AI. Donovan delivers fine-tuned large language models (LLMs) — built on Meta's Llama — for national-security missions inside controlled US-government and top-secret environments, supporting search, summarization, and operational decision support.
Thunderforge is a Defense Innovation Unit (DIU) prime-contract program under which Scale, working with Anduril's Lattice and Microsoft's LLMs, builds AI agents for campaign development, theater-wide resource allocation, and strategic assessment. It was initially focused on Indo-Pacific Command and European Command and was awarded via DIU in March 2025.
Scale's defense business includes a 2025 US Army research-and-development contract worth $99 million (running through 2030) and a separate five-year Department of Defense deal with a $100 million ceiling for AI on top-secret networks. As of 2025, Meta holds a large minority stake in Scale AI.
💡Key Concept
AI in military planning: Donovan and Thunderforge apply LLMs and AI agents not to weapons but to the planning layer — drafting courses of action, allocating resources across a theater, and summarizing intelligence. The appeal is speed at scale; the risk is that planners may over-trust fluent AI output, and a language model that fabricates a detail inside a classified planning loop can quietly shape a real decision.
Key Capabilities
- Donovan LLMs — fine-tuned models (built on Meta Llama) for classified national-security search, summarization, and decision support
- Controlled environments — deployed inside US-government and top-secret networks
- Thunderforge agents — AI agents for campaign development, resource allocation, and strategic assessment
- Partnered stack — Thunderforge combines Scale, Anduril's Lattice, and Microsoft's LLMs
- Command focus — Thunderforge initially targeted Indo-Pacific Command and European Command
⚠️Warning
Applying AI to military planning raises automation-bias and accountability concerns: the more capable the tool, the easier it is to defer to it. LLM hallucination inside a classified planning loop is a serious risk — a confidently wrong summary or recommendation can propagate into decisions that are hard to audit. Over-trust in AI planning outputs is a genuine failure mode, and Meta's large minority stake in Scale raises questions about data governance and independence.
Company Details
| Detail | Info |
|---|---|
| Company | Scale AI |
| Products | Donovan (classified LLMs) and Thunderforge (AI planning agents) |
| Thunderforge | DIU prime contract awarded March 2025; partners include Anduril and Microsoft |
| Contracts | 2025 US Army R&D contract $99 million (through 2030); five-year DoD deal with a $100 million ceiling for AI on top-secret networks |
| Ownership | Meta holds a large minority stake (2025) |
| Type | Private |
| Website | scale.com |
Key Takeaways
- Donovan delivers fine-tuned Llama-based LLMs for classified national-security work, and Thunderforge builds AI planning agents under a DIU prime contract with Anduril and Microsoft
- Scale's defense business includes a $99 million Army R&D contract and a five-year DoD deal with a $100 million ceiling for top-secret networks
- Putting AI into military planning raises real automation-bias and hallucination risks, and Meta's large minority stake raises data-governance and independence questions
