The US Federal Bureau of Investigation (FBI) is moving to expand its on-premises artificial intelligence capabilities, issuing a contract opportunity notice for GPU servers and related infrastructure. The solicitation, posted on SAM.Gov and first reported by FedScoop, signals the bureau's intent to build out its own AI compute capacity rather than rely solely on cloud services.
Under a multiple-award indefinite delivery, indefinite quantity (IDIQ) vehicle, the FBI is seeking proposals for a range of enterprise AI compute servers, rack-scale AI systems, pod-scale AI systems, AI inference servers, high-speed network interface cards, and data processing units. Specifically, the agency is requesting four enterprise AI compute servers, two rack-scale systems, one pod-scale system, ten inference servers with expansion hardware, and additional networking components. The collective ceiling for the IDIQ contract is set at $88 million.
The hardware is intended to support a broad set of workloads, including AI, machine learning, large language model training and inference, computer vision, advanced analytics, and general enterprise AI applications. The FBI has not named a preferred vendor or specified the desired compute capability of the GPUs, leaving room for bidders to propose their own configurations.
Industry pricing for comparable systems suggests the scale of investment involved. A rack-scale Nvidia GB200 NVL72 system currently sells in the range of $3 million to $4 million, while the upcoming Vera Rubin rack-scale system is expected to cost between $5 million and $7 million. These figures underscore the significant capital outlay required for agencies to maintain cutting-edge AI infrastructure in-house.
The cost challenge is not lost on the bureau's leadership. Speaking at a recent event, FBI chief AI officer Katie Noyes acknowledged the financial pressure, saying, "If I'm being candid, the price point for us right now in the FBI, it's a little steep." Her comments reflect a broader tension across government agencies as they balance the demand for advanced AI tools against constrained budgets.
The FBI previously issued a request for information in June that offered more detail on its technical expectations. That document indicated a preference for Intel Xeon 6767P processors or equivalent, a minimum of 2TB of RAM with options to expand to 4TB, and at least one HGX B300 eight-GPU air-cooled assembly or equivalent. Storage requirements included two 500GB or larger SSD boot drives and two 7.68TB or larger SSD data drives. Networking specifications called for two 25Gb Ethernet interfaces and two Nvidia BlueField-3 single-port 400 GbE DPUs or equivalent. The rack-scale systems were expected to meet or exceed the performance of one integrated GB300 NVL72 rack-scale system.
It remains unclear whether these specifications have been revised following the consultation period. The earlier RFI also noted that deliveries would be directed to the FBI Criminal Justice Information Services (CJIS) Division facility in Clarksburg, West Virginia.
The FBI operates several data centers across the US, including its primary facility in Pocatello, Idaho, known as "Grey Wolf." That facility opened in 2019, and by 2021 the bureau counted five main data centers nationwide. The current state of the FBI's data center footprint is not publicly documented, but this procurement suggests a continued commitment to expanding on-prem capacity for sensitive AI workloads.
The move comes as federal agencies increasingly seek to deploy AI in mission-critical environments, where data sovereignty and security considerations often push them toward on-premises infrastructure. The FBI's procurement could serve as a benchmark for other government entities navigating the same complex balance between performance, cost, and control.