Pentagon classified network expansions
The U.S. Department of War signed agreements on May 1 to allow eight frontier AI firms to run models on classified IL6 and IL7 networks. These environments manage secret and highly sensitive national-security data. Approved vendors include SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, Oracle, and Amazon Web Services. Scale AI secured a contract for the Thunderforge planning system in March 2025. Defense leaders seek $961.6 billion for the 2026 budget, and $33.7 billion is earmarked for science, technology, and autonomous systems to support the military’s transition toward an AI-first fighting force. More than 1.3 million personnel used the unclassified GenAI.mil platform, which launched in December 2025, to generate hundreds of thousands of agents in five months. This platform cuts task times from months to days. While the Pentagon uses various models, the NSA began deploying Anthropic’s Claude Mythos AI model on classified networks despite ongoing disputes. Anthropic remains absent from the May 1 agreements because negotiations about surveillance and weapons guardrails stalled. Palantir also pulled its Claude models from Department of Defense platforms. The May 1 list follows a pattern from 2025, where the Pentagon contracted Scale AI in March 2025 and signed deals for OpenAI’s ChatGPT and xAI’s Grok in July 2025.
Scale AI revenue and market rivalry
Meta purchased a 49% stake in Scale AI for $14.3 billion, which implies a $29 billion valuation. This investment creates friction because rival labs might reconsider sending sensitive research to a supplier tied to a competitor. Microsoft and xAI are reportedly considering moving away from Scale AI after this deal. Snorkel AI reports an annualized revenue run rate of $375 million, a massive increase from $20 million the previous year. Snorkel also holds a $3.5 billion valuation after it shifted from labeling software to delivering finished datasets and reinforcement-learning environments. Other competitors grab significant market share in the expert work sector. Mercor reports $2 billion in gross annualized revenue and employs 30,000 experts. Surge AI generated $1.2 billion in 2024. Handshake manages a network of 3 million graduate-level candidates and 500,000 PhDs. Handshake revenue jumped from $10 million at the start of the year to nearly $1 billion a few months later. Micro1 grew its revenue from $100 million to $500 million in less than a year and holds a $4 billion valuation. AfterQuery reached a $3.2 billion valuation after its revenue run rate passed $100 million within 14 months. Turing produces technical datasets with revenues reaching $300 million. Turing’s Frontier Research Lab spends an average of 58 hours of expert work on a single task. Turing also releases the CyberStrike benchmark, which contains 200 security tasks that even strong frontier-agent configurations solve less than one-third of the time. Will the Pentagon favor independent vendors over those tied to Meta? You should note that these marketplaces move large sums through contractors.
Scale AI service specifications
Scale AI provides data services for frontier labs and government agencies. The company uses its Outlier subsidiary for LLM training and Remotasks for autonomous vehicle annotation.
| Service Type | Target User |
|---|---|
| RLHF | Frontier AI labs |
| Autonomous vehicle data | Robotics companies |
| 3D/LiDAR annotation | Computer vision teams |
| Red team testing | AI safety researchers |
| LLM evaluation | Research divisions |
Scale Labs produces benchmarks to test model performance. The company works with major players like OpenAI, Google, Meta, and Microsoft. Scale AI maintains an enterprise-only model with custom pricing and no public pricing. This structure results in long sales cycles that make the company unsuitable for small projects or startups with limited budgets. Small teams should instead look to Roboflow or Labelbox for multimodal annotation.
