US seeks $9 billion for Nvidia GB10 superchips to fuel AI arms race

The US government is escalating its AI ambitions with a secret $9 billion request for Nvidia's GB10 superchips, aimed at helping intelligence agencies like the CIA and NSA keep pace with private-sector leaders such as Anthropic and OpenAI. The move underscores a broader arms race where national security hinges on cutting-edge compute power. The GB10, built on Nvidia's Grace Blackwell architecture, delivers 1 petaflop of FP4 AI performance while drawing only 140 watts — a stark contrast to the multi-kilowatt consumption of larger AI systems. Each chip pairs a 20-core Arm CPU (designed by MediaTek) with a Blackwell-based GPU, 128GB LPDDR5x memory, and 4TB of NVMe storage, capable of fine-tuning 70-billion-parameter models.

Scaling from a single $5,000 rack-mounted unit to data-center proportions reveals the true cost of AI dominance. Nvidia's GB300 NVL72 — a liquid-cooled rack with up to 72 GPUs and 36 CPUs — can cost between $1.8 million and $4 million per unit. A full-scale data center may house up to 100,000 such racks, driving astronomical power and cooling demands. While the $9 billion request has received executive thumbs-up, it still requires congressional approval. The funding would prioritize rapid deployment, ensuring the US intelligence community can train and run cutting-edge models without falling further behind the blistering pace of corporate AI development.