Donovan (Scale AI)

A US defense generative-AI decision-support and agent-orchestration platform that lets military users query classified data, build mission-tailored AI agents and compress planning workflows.

Overview

Donovan is Scale AI's federal defense-AI platform for military planning, intelligence fusion and decision support. It is not itself a large language model; Scale and defense-industry reporting describe it as an application and orchestration layer that gives operators access to LLMs, mission data, evaluation tools and no-code agent workflows on government networks. The platform sits in the same command-software class as Palantir AIP and Anduril Lattice, but its starting point is Scale AI's data-engineering and model-evaluation business rather than a legacy battlefield-management system.

Scale describes Donovan as a platform to help operators, analysts and commanders "understand, plan, and act in minutes instead of weeks" by connecting operational data, models and mission-specific workflows inside controlled environments, according to Scale AI. Its public defense profile grew from the May 2023 deployment with the US Army's XVIII Airborne Corps, which Scale and Business Wire described as the first LLM deployment on a US government classified network through Business Wire.

Development

Scale AI unveiled Donovan on 10 May 2023 with the XVIII Airborne Corps partnership, framing the system as a classified-network platform for ingesting more than 100,000 pages of live operational data and making that material queryable in natural language, according to Business Wire. Potomac Officers Club's account of the same rollout described Donovan as supporting commanders, intelligence analysts and planning staffs through faster document search, operational context and planning-cycle support on classified infrastructure via Potomac Officers Club.

By 2024, Scale AI had added Defense Llama, a national-security fine-tune of Meta's Llama 3, to the Donovan ecosystem. DefenseScoop reported that the model was accessible only through controlled government hubs housed within Donovan and was intended for integration with command-and-control systems, intelligence tools and decision-support workflows, according to DefenseScoop. This marked Donovan's evolution from an LLM-access and document-analysis layer toward a broader platform for defense-specific model orchestration.

The next major public inflection was Thunderforge. On 5 March 2025, the Defense Innovation Unit awarded Scale AI the prime prototype contract for Thunderforge, described by DIU as a flagship effort to apply AI to military planning and operations; the program uses Scale's agentic applications and test-and-evaluation expertise alongside Anduril's Lattice and Microsoft-enabled LLMs, according to DIU. Thunderforge is related to Donovan's technology family, but the public record treats it as a distinct multi-vendor prototype program rather than a simple Donovan rebrand.

In September 2025, the Pentagon's Chief Digital and Artificial Intelligence Office awarded Scale an enterprise Production Other Transaction Authority agreement with a $100 million ceiling, later expanded in May 2026 to $500 million because demand exceeded the original ceiling, according to Scale AI. Bloomberg and Forbes framed the enlarged vehicle as one of the larger Pentagon AI-procurement moves of the period and noted Scale AI's Meta ownership context via Bloomberg and Forbes.

Design & capabilities

Donovan is best understood as a secure AI application layer over defense data and LLMs. Scale's public product page says Donovan lets users deploy mission-tailored AI agents "from concept to combat" through integration with the Scale GenAI Platform, including data-source connection, knowledge-base construction, model selection, agent instructions and links into existing systems, according to Scale AI. The platform is therefore closer to an AI workflow and orchestration environment than a single model, sensor or weapons-control application.

The original Donovan deployment emphasized natural-language access to large volumes of operational material. Scale said the XVIII Airborne Corps use case let users ingest orders, situation reports and intelligence reporting, then ask questions without coding or specialist training, according to Business Wire. Publicly described use cases included friendly and enemy-force context, intelligence preparation of the battlespace, crisis planning and Joint All-Domain Command and Control support.

The architecture is model-flexible. DefenseScoop reported that Defense Llama is a Scale fine-tune of Meta's Llama 3 for national-security users and is made available through Donovan-controlled government hubs, according to DefenseScoop. Scale's CDAO announcement also describes a broader stack consisting of Scale Data Engine, the Scale GenAI Platform and Donovan, deployable across NIPR, SIPR and JWICS network tiers, according to Scale AI.

Human oversight is central to the public design claim, but public evidence remains vendor-framed. Scale says the platform uses Reinforcement Learning from Human Feedback and is intended to keep outputs reliable, auditable and trusted as the department moves from pilots to enterprise deployment, according to Scale AI. Independent literature on LLMs in national-security settings identifies hallucination, reliability and trust as unresolved issues for high-stakes decision-support systems, including military planning and intelligence applications, in an arXiv paper.

Combat record / operational use

Donovan has no public combat record in the sense of a weapon fired or an independently documented battlefield effect. Its operational record is instead a record of classified-network deployment, command adoption and integration into planning workflows. The first publicly identified fielding was with the US Army's XVIII Airborne Corps in 2023, where Scale said the system supported commanders and staff working with live operational data on a classified network, according to Business Wire.

Public reporting later associated Scale's defense-AI stack with multiple US defense components. Scale's CDAO enterprise agreement covers Donovan, Scale Data Engine, Scale GenAI Platform and engineering sprints, and Scale says the agreement gives components across the Department of Defense or Department of War access to these capabilities without each office running a separate procurement, according to Scale AI. The same announcement states the products are available across NIPR, SIPR and JWICS environments.

Thunderforge expands the operational context from staff search and analysis toward theater-level planning and wargaming. DIU says the prototype will initially support US Indo-Pacific Command and US European Command for campaign development, theater-wide resource allocation and strategic assessment, according to DIU. DefenseScoop reported the program was intended to integrate AI-powered planning tools at those combatant commands and scale more broadly if successful, through DefenseScoop.

There is no public independent audit of Donovan's targeting-relevant accuracy, hallucination rate, operational suitability or wartime decision impact. The best-supported assessment is that Donovan is in active US defense use and scaling through enterprise procurement, while its measurable operational performance remains largely non-public.

Advantages

• Secure-network orientation: Donovan was publicly introduced as a classified-network LLM deployment and later described as available across NIPR, SIPR and JWICS environments by Scale AI.
• Planning-cycle compression: Scale and DIU both frame the platform family as aimed at shortening military planning and assessment cycles from legacy staff timelines toward AI-assisted iteration through DIU.
• Model and agent flexibility: Donovan can connect defense data, knowledge bases, models and no-code agent workflows rather than requiring users to rely on one fixed LLM, according to Scale AI.
• Data-engineering base: Scale's broader business in data labeling, model evaluation and AI infrastructure gives Donovan an ecosystem beyond a pure chat interface, reinforced by the CDAO vehicle covering Data Engine, SGP and Donovan through Scale AI.
• Thunderforge adjacency: Scale is prime on DIU's Thunderforge prototype, giving the company a path from Donovan-style decision support into combatant-command planning workflows, according to CNBC.

Drawbacks / limitations

• Public performance evidence is thin: no independent hallucination-rate, red-team or operational-effectiveness audit for Donovan is public, while independent LLM literature flags reliability and trust as unresolved high-stakes issues in an arXiv paper.
• Vendor claims dominate: most specific capability claims come from Scale announcements or contractor-facing reporting, not from released test reports or formal operational assessments.
• Procurement boundaries are easy to blur: Donovan, Defense Llama, Scale GenAI Platform and Thunderforge are related but distinct; conflating them can overstate what any single program has fielded.
• Competitive pressure is material: Business Insider reported that Scale lost a large NGA/Maven data-labeling competition to Enabled Intelligence and later sued the Department of Defense, showing contested access to the data layer that supports defense-AI platforms via Business Insider.
• Ownership scrutiny may affect trust: Reuters reported Meta's 49% non-voting investment in Scale AI and the resulting valuation context, which may sharpen government and commercial questions about data governance, neutrality and vendor dependence through Reuters.

Counterparts

• Palantir AIP (competing defense LLM/agent layer)
• Maven Smart System (DoD AI-targeting program)
• Anduril Lattice (autonomy-mesh C2)

Outlook

Donovan's near-term trajectory is expansion rather than replacement. The $500 million CDAO ceiling gives Scale a broad access path for Donovan, Data Engine and the Scale GenAI Platform across defense components, while Thunderforge gives the company a high-visibility route into combatant-command planning. The central uncertainty is not whether defense organizations will experiment with Donovan-like systems; public procurement already shows they will. The harder question is whether the platform can demonstrate reliability, auditability and decision advantage under classified operational conditions.

The competitive field is also tightening. Palantir's AIP has the advantage of long-running data-integration infrastructure and a central role in Maven, while Anduril's Lattice connects autonomy and C2 workflows more directly to sensors and effectors. Donovan's advantage is Scale's combination of data infrastructure, model evaluation and agent workflow tooling; its vulnerability is that those strengths are less visible to operators if reliability, provenance and integration controls are not trusted.

Key specifications

Sources

1. Scale AI — "Thunderforge: AI for American Defense." https://scale.com/blog/thunderforge-ai-for-american-defense
2. Defense Innovation Unit — "DIU's Thunderforge project to integrate commercial AI-powered decision-making." https://www.diu.mil/latest/dius-thunderforge-project-to-integrate-commercial-ai-powered-decision-making
3. CNBC — "Scale AI announces multimillion-dollar defense military deal." https://www.cnbc.com/2025/03/05/scale-ai-announces-multimillion-dollar-defense-military-deal.html
4. DefenseScoop — "DIU Thunderforge: Scale AI, combatant commands, INDOPACOM, EUCOM." https://defensescoop.com/2025/03/05/diu-thunderforge-scale-ai-combatant-commands-indopacom-eucom/
5. Breaking Defense — "AI for war plans: Pentagon innovation shop taps Scale AI to build Thunderforge prototype." https://breakingdefense.com/2025/03/ai-for-war-plans-pentagon-innovation-shop-taps-scale-ai-to-build-thunderforge-prototype/
6. Business Wire — "Scale AI Partners with XVIII Airborne Corps for First LLM Deployment to a U.S. Government Classified Network." https://www.businesswire.com/news/home/20230510005630/en/Scale-AI-Partners-with-XVIII-Airborne-Corps-for-First-LLM-Deployment-to-a-U.S.-Government-Classified-Network
7. Potomac Officers Club — "Scale AI deploys Donovan platform on classified government network." https://www.potomacofficersclub.com/scale-ai-deploys-donovan-platform-on-classified-government-network/
8. DefenseScoop — "Scale AI unveils Defense Llama large language model for national security users." https://defensescoop.com/2024/11/04/scale-ai-unveils-defense-llama-large-language-model-llm-national-security-users/
9. Scale AI — Donovan product page. https://scale.com/donovan
10. Scale AI — "Scale AI and Pentagon CDAO expand agreement to $500 million." https://scale.com/blog/scale-ai-pentagon-cdao-500-million-agreement
11. Bloomberg — "Meta-backed Scale AI wins $500 million Defense Department deal." https://www.bloomberg.com/news/articles/2026-05-06/meta-backed-scale-ai-wins-500-million-defense-department-deal
12. Forbes — "Pentagon hands Meta-backed Scale AI $500 million contract." https://www.forbes.com/sites/aliciapark/2026/05/06/pentagon-hands-meta-backed-scale-ai-500-million-contract-5-times-last-years-deal-report-says/
13. Reuters — "Meta finalizes investment in Scale AI valuing startup at $29 billion." https://www.reuters.com/business/finance/meta-finalizes-investment-scale-ai-valuing-startup-29-billion-2025-06-13/
14. The Verge — "Meta, Scale AI and Alexandr Wang." https://www.theverge.com/meta/685711/meta-scale-ai-ceo-alexandr-wang
15. Wikipedia — Alexandr Wang. https://en.wikipedia.org/wiki/Alexandr_Wang
16. Business Insider — "Pay cuts, poaching and pivoting inside Scale AI after Meta." https://www.businessinsider.com/pay-cuts-poaching-pivoting-inside-scale-ai-meta-2025-12
17. Business Insider — "Scale AI sues Department of Defense after NGA Maven contract loss." https://www.businessinsider.com/scale-ai-department-of-defense-lawsuit-court-meta-2026-2
18. CSIS — "What is the Maven Smart System and what does it do?" https://www.csis.org/analysis/what-maven-smart-system-and-what-does-it-do
19. Palantir — AIP for Defense. https://www.palantir.com/platforms/aip/defense/
20. Scale AI — "Scale AI and DoD expand Army R&D partnership." https://scale.com/blog/scale-ai-dod-expand-army-rd-partnership
21. arXiv — "Large Language Models in National Security Applications." https://arxiv.org/pdf/2407.03453
22. i10x.ai — "LLMs in military decision-making risks." https://i10x.ai/news/llms-in-military-decision-making-risks

🔒 The Analyst Section on Donovan (Scale AI) is for BattlePolicy Pro members. Below the line: procurement and contract history, the suppliers and companies behind the system, our head-to-head analyst view of the systems it actually competes with, and the BattlePolicy assessment note. Unlock the full file with BattlePolicy Pro →