CoreWeave and Leidos are creating a secure AI cloud for U.S. intelligence agencies.

      The AI cloud is moving into secure locations. CoreWeave provides high-performance GPUs to major AI labs and now aims to set up its cloud within the Sensitive Compartmented Information Facilities (SCIFs), the secure areas where classified information is stored and analyzed, with the help of Leidos.

      The partnership was announced on Thursday. CoreWeave plans to implement its AI infrastructure in these SCIFs, while Leidos will manage the elements that a commercial cloud provider cannot address: accreditation, secure architecture, cybersecurity operations, and the integration of missions.

      The target customers are the U.S. intelligence agencies and the Department of Defense, which is now referred to as the Pentagon. Both companies describe the offering as “sovereign” AI capabilities, using the same technology stack that CoreWeave provides to commercial labs, but modified to meet the requirements of classified projects.

      They outlined five primary offerings: a classified cloud for training and operating models, alongside tools to assist analysts in integrating intelligence, interpreting imagery, and making decisions. Additional features focus on security and simulations, providing platforms for testing AI-related cyber threats, synthetic data generation, and digital twins for training purposes. The last component connects the central cloud to isolated, forward-deployed units in the field.

      Sachin Jain, CoreWeave's chief operating officer, emphasized the need for federal teams to have “secure, scalable platforms” for AI applications, stating that the company aims to deliver tools typically reserved for leading AI labs to secure government settings.

      The significant challenge in this endeavor is not the hardware but rather the clearance required. A commercial AI cloud operates on principles contrary to secrecy: expansive shared areas, adaptable capacity, and customers accessing their systems via the open internet.

      In contrast, classified work necessitates a different approach. The hardware must be located in a facility cleared for the sensitive data, operated by authorized personnel, and disconnected from the public internet. Comprehensive auditing and model governance must be integrated within the system rather than added later.

      This requirement reveals why the government cannot merely lease space in a standard cloud. It also explains why an integrator like Leidos is essential for this project.

      Leidos offers the necessary scale for this type of work, having reported around $17.2 billion in revenue last year and employing over 50,000 individuals. Jason O’Connor, president of Leidos Intelligence, characterized this as “the next evolution of mission technology,” highlighting “sovereign AI computation at scale, designed with security in mind.”

      CoreWeave sees the federal opportunity as a solution to a critical problem: dependency on a small number of clients. The company has seen rapid growth by providing computing power to a few major players, reporting $2.08 billion in revenue for the first quarter, a 112% increase year-over-year, against a backlog of $99.4 billion.

      Much of this business hinges on a few clients, including a $21 billion commitment from Meta and a long-term agreement with Anthropic. Selling to government agencies like intelligence and military sources would help mitigate this risk.

      This move also aligns with a broader strategy. CoreWeave established a dedicated federal division last year, meeting standards such as FedRAMP and other governmental requirements. Competing companies are pursuing the same funding, including military cyber startups and chip manufacturers supplying supercomputers for armed forces.

      The U.S. government is also consolidating its software contracts, and the quest for sovereign AI now extends beyond its borders, including Europe’s initiatives in military AI.

      Currently, there hasn’t been much formal agreement. CoreWeave’s filing indicates that the collaboration relies on definitive agreements that are still pending. The deployment will proceed based on “mission requirements and federal appropriation priorities,” meaning progress will be aligned with government timelines rather than the pace of commercial deals.

      The capability is substantial. CoreWeave surpassed 1GW of active power in the first quarter and has contracts for over 3.5GW. Its CEO, Michael Intrator, indicated projections of exceeding 8GW by 2030.

      However, the classified portion will be small in comparison, constrained not by power capacity but by the legalities of placing accredited systems.