FBI Issues RFI for AI Supercomputers Featuring Nvidia B300 GPUs and Google TPUs

wpnews.pro

The FBI's RFI spans four hardware categories: single-node HGX B300, rack-scale GB300 NVL72, Google TPU v8/v7L AI pods, and Nvidia L40-S inference accelerators [1] - Category 3 specifies performance equivalent to five GB300 NVL72 systems, pointing to a substantial training cluster requirement [1] - The systems would be deployed at the FBI's CJIS Division in Clarksburg, West Virginia, for LLM training, inference, computer vision, and analytics in classified environments [1] - The inclusion of Google TPUs is notable — Google has historically limited on-premises TPU availability outside its own cloud infrastructure [1] - Federal AI spending obligations reached $7.2 billion in 2026, a 966% increase from 2024's $675 million

[2] The FBI has published a Request for Information on SAM.gov seeking market input on AI supercomputer hardware for its Criminal Justice Information Services (CJIS) Division in Clarksburg, West Virginia. The RFI outlines four distinct categories of accelerator systems spanning training, inference, and analytics workloads — including both Nvidia Blackwell Ultra GPUs and Google TPUs [1].

The filing signals the bureau's intent to build dedicated AI compute infrastructure for "large language model training, inference, advanced analytics, computer vision, and other AI-enabled workloads within secure government computing environments," according to the RFI. Multiple systems are anticipated across varying quantities, though no budget figure has been disclosed [1].

The move comes as federal AI spending has surged dramatically, with obligated funds reaching $7.2 billion in 2026 — a 966% increase from $675 million in 2024 [2]. The FBI's RFI is among the more technically specific federal AI procurement signals to emerge from a civilian agency, particularly given its inclusion of Google TPU hardware that has historically been available only through Google Cloud.

Four Categories of Accelerator Hardware #

The FBI's RFI organizes its requirements into four distinct hardware categories, each targeting different workload profiles [1].

Category 1 covers single-node training servers built around Intel Xeon 6767P processors with 2-4 TB of RAM, one Nvidia HGX B300 eight-GPU assembly, SSD storage, 25Gb Ethernet, and Nvidia BlueField-3 data processing units for network offload and security [1].

Category 2 steps up to integrated rack-scale systems, specifying one GB300 NVL72 equivalent — Nvidia's flagship 72-GPU liquid-cooled rack — along with high-speed interconnect fabric, Spectrum-X networking architecture, and in-rack cooling [1].

Category 3 is the most ambitious: an AI pod architecture equivalent to Google TPU v8 or v7L, or better, with performance matching five GB300 NVL72 systems. The RFI requires support for multiple AI frameworks, suggesting the FBI wants vendor-neutral software flexibility even on Google's custom silicon [1].

Category 4 targets inference workloads specifically, calling for Nvidia L40-S accelerators or equivalent with 48GB of GPU memory and support for FP32, FP16, BF16, and INT8 precision formats [1].

The Google TPU Question #

The inclusion of Google TPUs in Category 3 is the most unusual element of the RFI. Google has historically kept its TPU hardware within its own data centers and offered access exclusively through Google Cloud Platform. On-premises TPU deployments outside Google's infrastructure have been extremely rare [1].

For the FBI to receive TPUs in a classified, air-gapped environment at CJIS Clarksburg, Google would need to support a deployment model that diverges from its standard cloud-first approach. The RFI's framing — requiring performance parity with five NVL72 racks — suggests the bureau is evaluating whether a TPU pod could deliver comparable or superior performance to Nvidia's top-end Blackwell Ultra systems for large-scale training. The performance benchmark is significant. A single GB300 NVL72 rack contains 72 Blackwell Ultra GPUs; five racks would represent 360 GPUs with aggregate FP4 performance in the multi-exaflop range. Matching that with TPU v8 hardware would require a substantial pod deployment.

CJIS Division and Secure Compute #

The FBI's CJIS Division operates from a 986-acre campus in Clarksburg, West Virginia, anchored by a 500,000-square-foot main facility that includes a 100,000-square-foot computer center. The division employs more than 3,000 staff and contractors and manages some of the government's most sensitive criminal justice databases, including the National Crime Information Center (NCIC) and the Next Generation Identification (NGI) biometric system [3].

Deploying AI supercomputers at CJIS would bring LLM training and inference capabilities into the same secure perimeter as these existing databases — a logical move for workloads like computer vision applied to biometric data, natural language processing for investigative records, and advanced analytics across criminal justice datasets.

The RFI's emphasis on 'secure government computing environments' underscores that these systems would operate under strict information-handling requirements, likely at classified or law-enforcement-sensitive levels.

Federal AI Compute Buildout Accelerates #

The FBI's RFI arrives amid a broader federal push to build dedicated AI compute capacity. Federal AI spending obligations reached $7.2 billion in 2026, up from $675 million in 2024 — a 966% increase. Potential contract awards ballooned to $91.8 billion, up 1,912% from 2024's $4.6 billion [2].

The Department of Energy has led on large-scale deployments, partnering with Nvidia and Oracle on the Solstice system featuring 100,000 Blackwell GPUs and the 10,000-GPU Equinox system at Argonne National Laboratory [4]. The FBI's requirements, while smaller in scale, represent a significant step for a civilian law enforcement agency building its own on-premises AI training infrastructure rather than relying on cloud services.

The RFI is a market research exercise and does not commit the FBI to any procurement. However, the specificity of the hardware categories — naming exact GPU models, networking architectures, and precision formats — suggests advanced planning rather than exploratory inquiry. Vendors responding to the RFI will help the bureau shape a future solicitation.

What Comes Next #

As a Request for Information, the posting does not guarantee a contract award or set a procurement timeline. RFIs are standard federal acquisition practice to gauge market capabilities and inform requirements before issuing a formal solicitation [1].

The key variables to watch are whether Google can or will offer an on-premises TPU deployment model for a classified government customer, and whether the FBI ultimately selects a single vendor platform or splits across multiple categories — procuring Nvidia hardware for some workloads and TPU pods for others. The Category 3 specification, requiring support for multiple AI frameworks on TPU hardware, suggests the bureau is wary of vendor lock-in even as it evaluates Google's silicon.