Pentagon awards Scale AI a $500m contract to sift battlefield data

WASHINGTON. The Pentagon has awarded Scale AI a $500 million contract to help process battlefield data and assist in operational decision-making, a five-fold expansion of a deal struck only eight months ago and the latest sign that the US military is wiring its planning workflow into Silicon Valley.

The new agreement, reported by Bloomberg on Wednesday, lifts the $100 million ceiling that the Department of Defense imposed in September 2025 to $500 million. Dan Tadross, who runs Scale AI's public sector business, told Bloomberg that the Pentagon was "pushing the limits" on the original contract.

Scale AI, based in San Francisco, supplies the Pentagon with three platforms it markets as an end-to-end AI stack for classified networks. One is a generative AI testbed for defense and intelligence agencies. Another is a decision-support tool called Donovan that pulls usable answers from messy unstructured data. The last is a data engine that converts raw sensitive material into training fodder for military AI models.

Why the deal jumped so fast

The Pentagon's original September 2025 award started with a $41 million tranche for data labeling at Scale's St Louis AI Center, where workers tag geospatial imagery and other sensitive material so models can read it. The Register reported at the time that the contract carried a $100 million ceiling over five years.

Tadross's "pushing the limits" comment points to demand outpacing that ceiling. The new $500 million figure suggests the Department of War, as the Trump administration renamed the Pentagon last year, wants room to plug Scale's tools into more workflows, faster.

Scale's role inside the Pentagon goes beyond data labeling. In March 2025 the Defense Innovation Unit picked the company to lead Thunderforge, a program to put generative AI agents inside large-scale military planning. Anduril provides its Lattice software backbone. Microsoft contributes large language models. Scale ties the pieces together with its own agents and evaluation tools.

Thunderforge is meant to compress the staff work that has long made campaign planning slow. Bryce Goodman, who leads the program at the Defense Innovation Unit, said in March 2025 that "today's military planning processes rely on decades-old technology and methodologies, creating a fundamental mismatch between the speed of modern warfare and our ability to respond." The first Thunderforge users are at US Indo-Pacific Command in Hawaii and US European Command in Germany.

A second Silicon Valley pipeline into the Pentagon

The Scale expansion fits a wider rotation of Silicon Valley capital and talent into defense work. Anduril, founded by Palmer Luckey, has become a counterweight to traditional primes through its Lattice platform and autonomous systems. Palantir last year extended its Maven AI warfare program with a five-year contract worth up to $100 million. OpenAI lifted its ban on military uses of its models in early 2024 and now partners with Anduril on national security work. Anthropic, Palantir and Amazon Web Services have a separate compact to provide Claude models to US intelligence and defense agencies.

The Department of War announced this month an arrangement with seven frontier AI vendors, including SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft and Amazon Web Services, to deploy capabilities on its Impact Level 6 and Impact Level 7 classified networks. Over 1.3 million Department personnel now use the GenAI.mil platform, with hundreds of thousands of agents already live, according to the announcement.

Scale's pitch is that it covers a different gap. The GSA's broader OneGov agreements give federal agencies cheap access to specific models, often at deep discounts and short terms. Scale instead sells the underlying data preparation work, the testbed for trying models against that data, and the agent that runs on top, all certified to operate at Top Secret and Sensitive Compartmented Information clearance levels. Executives have argued that is the difference between renting a model and standing up an actual battlefield AI workflow.

A legal cloud over a key adjacent contract

The new $500 million award lands while Scale is in court with the same department it is now serving. In late January, Scale sued the Department of Defense in the US Court of Federal Claims after losing a separate contract worth up to $708 million from the National Geospatial-Intelligence Agency to a smaller competitor, Enabled Intelligence. That work supports Maven, the Pentagon's flagship imagery and geospatial AI initiative. The Government Accountability Office had dismissed Scale's earlier bid protest two days before the suit was filed.

Most filings remain under seal. Scale has said only that the case "relates to a recent procurement decision" and stressed it "stands firmly with Secretary Hegseth and the Department of War" on getting frontier AI capabilities to warfighters. The case could shape how future high-value classified AI contracts are evaluated, especially when an incumbent loses a follow-on bid.

Meta's stake and a high-profile alumnus

Scale's deepening defense work also draws attention because of who owns nearly half of it. In June 2025, Meta Platforms invested $14.3 billion in Scale for a 49 percent stake, one of the largest single investments in an AI infrastructure provider. Co-founder Alexandr Wang left Scale shortly afterward to become chief AI officer at Meta's Superintelligence Labs and has attended White House AI dinners with President Donald Trump.

Wang earlier wrote an open letter to Trump after his second inauguration urging more federal money for data and computing infrastructure, citing Scale's defense work. Since the Meta investment Scale has cut roughly 200 jobs, about 14 percent of its workforce, and lost commercial AI customers including Google and xAI. Federal contracts are picking up the slack.

What to watch next

The rapid expansion of the Scale contract suggests the Pentagon is willing to bet bigger on a single vendor for the data and decision-support layer of its AI stack, even as it spreads model-level contracts across many providers. The next test is which combatant commands beyond Indo-Pacific and European Command pick up Thunderforge as it rolls out. A second is whether the sealed Court of Federal Claims case forces the National Geospatial-Intelligence Agency to disclose how it rated competing bids on the $708 million Enabled Intelligence award. The third is how quickly agents built on Scale's stack reach actual operational use rather than staff exercises.

For now, $500 million buys five times the runway. Whether the underlying data is clean enough to make the speed pay off is the question that will decide whether the Pentagon's bet on Silicon Valley pays out.