FBI eyes Nvidia and Google TPU AI systems

The FBI is exploring a purchase of specialized AI computing systems for secure government infrastructure, including hardware for large language model training and inference, data processing, and computer vision.

According to a newly published request for information (RFI), the agency has not yet decided to buy anything. The document is an early market-sounding step, but it says the bureau could acquire multiple systems in different configurations and volumes.

The FBI split its requirements into four categories.

The first covers server-based systems built around Intel Xeon 6767P processors or equivalent hardware, with at least 2TB of RAM expandable to 4TB, plus an 8-GPU Nvidia HGX B300 setup or a comparable platform. These systems must also include SSD storage for the operating system and data, along with high-bandwidth networking.

Recommended reading

Europe’s chip push won’t end US cloud reliance

The second category focuses on integrated rack-scale systems for data-center deployments. The FBI is considering platforms on the level of Nvidia GB300 NVL72 — prebuilt compute racks designed for training and running large AI models. The agency says such systems should include high-speed interconnects between accelerators, Nvidia Spectrum-X-class Ethernet infrastructure, built-in cooling, and support for connecting to water-cooling systems at government facilities.

The third category is aimed at enterprise AI cluster architectures. The documents mention systems comparable to Google TPU v8, TPU v7L, or newer. Required performance must be at least on par with 5 racks of Nvidia GB300 NVL72. The hardware also needs to support large language model workloads and work with a range of AI frameworks.

That TPU reference stands out because Google TPUs have traditionally been offered through Google’s cloud infrastructure, rather than as on-premises systems installed in customers' own data centers. The report notes that Google has been gradually expanding TPU availability beyond its cloud, including through partner compute sites.

The fourth category covers accelerators for inference. Here, the FBI is looking at hardware on the level of the Nvidia L40S with at least 48GB of VRAM for generative AI, embeddings, and computer vision. The accelerators must support FP32, FP16, BF16, and INT8, as well as tools including CUDA, TensorRT, PyTorch, and TensorFlow.

According to the documents, most potential deliveries are intended for the FBI’s Criminal Justice Information Services (CJIS) division in Clarksburg, West Virginia, where the bureau plans to host equipment for secure computing workloads.