The AI cloud is heading into the vault. CoreWeave rents high-end GPUs to the biggest AI labs. Now it wants to run that same cloud inside the sealed facilities where America keeps its secrets, with the contractor Leidos as its guide.
The two companies announced the collaboration on Thursday. CoreWeave will place its AI platform inside Sensitive Compartmented Information Facilities, the sealed rooms known as SCIFs where classified intelligence is stored and analysed. Leidos will handle the parts a commercial cloud provider cannot: accreditation, secure architecture, cyber operations and mission integration.
The customers are America’s spy agencies and the Department of War, as the Pentagon is now styled. Both sides call it “sovereign” AI capacity. It is the same stack CoreWeave rents to commercial labs, rebuilt to the rules that classified work demands.
What they plan to sell
The two companies sketched out five things they intend to offer. There is a classified cloud for training and running models, plus tools to help analysts fuse intelligence, exploit imagery and reach decisions.
The rest leans into security and simulation. There are cyber ranges to war-game AI attacks, plus synthetic data and digital twins for training. A final piece links the central cloud to disconnected, forward-deployed units in the field.
Sachin Jain, CoreWeave’s chief operating officer, said federal teams need “secure, scalable platforms” to put AI to work. He said the company wants to extend the tools it gives the world’s top AI labs “to highly secure government environments.”
Why accreditation is the hard part
The constraint here is not chips. It is clearance. A commercial AI cloud is built for the opposite of secrecy: huge shared regions, elastic capacity, customers reaching their clusters over the open internet.
Classified work needs the reverse. The hardware has to sit in a facility cleared for the data, run by cleared staff and cut off from the public internet. Auditing and model governance come built in, not bolted on.
That gap is why a government cannot simply rent a slice of an ordinary cloud region. It is also why an integrator like Leidos is in the deal at all.
Leidos brings the scale for that work. It reported about $17.2bn in revenue last year and employs more than 50,000 people. “This is the next evolution of mission technology,” said Jason O’Connor, president of Leidos Intelligence, describing “sovereign AI compute at scale, secure by design.”
Why CoreWeave needs the work
For CoreWeave, federal money solves a specific problem: concentration. The company grew explosively by renting compute to a handful of giants. It reported $2.08bn in first-quarter revenue, up 112% on the year, against a backlog worth $99.4bn.
Much of that rests on a few names, including a $21bn commitment from Meta and a multi-year deal with Anthropic. Selling to spies and soldiers spreads that risk.
It also fits a wider push. CoreWeave launched a dedicated federal arm last year, aligning with FedRAMP and other government sign-offs. Rivals are chasing the same budgets. Those include military cyber startups and chipmakers handing supercomputers to the armed forces.
Washington is also consolidating its software contracts, and the hunt for sovereign AI now stretches well beyond it, from Europe’s own military-AI push onward.
What is actually signed
Not much yet. CoreWeave’s own filing notes the arrangement depends on definitive agreements that are not yet signed. It says deployment will follow “mission requirements and federal appropriation priorities.” In plain terms, this moves at the speed of Washington, not the speed of a commercial contract.
The capacity is there to draw on. CoreWeave passed 1GW of active power in the first quarter and has contracted more than 3.5GW. Its chief executive, Michael Intrator, says it is heading toward more than 8GW by 2030.
The classified slice will be tiny by comparison. It is limited less by megawatts than by where an accredited rack can legally sit.
Get the TNW newsletter
Get the most important tech news in your inbox each week.