US seeks $9 billion for Nvidia superchips to keep intelligence agencies ahead in AI arms race

At a glance:

• US government requests $9 billion to buy Nvidia GB10 superchips for intelligence agencies
• GB10 chip combines a 20‑core MediaTek Arm CPU, Blackwell GPU, 128 GB LPDDR5x memory and 4 TB NVMe, delivering 1 PFLOP FP4 at 140 W
• Rack‑scale GB300 NVL72 systems can house up to 72 GPUs and 36 CPUs, costing $1.8‑4 M per rack

What the GB10 chip delivers

Nvidia’s latest AI accelerator, the GB10, is built around the Grace‑Blackwell architecture and is aimed at the most demanding generative‑AI workloads. The chip integrates a 20‑core Arm CPU sourced from MediaTek—codenamed Grace—with an Nvidia GPU that follows the Blackwell design. Memory is generous: 128 GB of LPDDR5x sits alongside a 4 TB NVMe M.2 SSD, giving the silicon enough bandwidth to feed models that can contain up to 70 billion parameters. In performance terms the GB10 can push roughly one petaflop of FP4 AI calculations while drawing only about 140 watts, a figure that is modest compared with the 1 kW power envelopes of high‑end gaming rigs.

Scaling to rack and data‑center level

When the GB10 is assembled into a rack‑scale system—named the GB300 NVL72—the density skyrockets. Each GB300 unit packs as many as 72 GPUs and 36 CPUs into a single liquid‑cooled chassis. Pricing for a single rack ranges from $1.8 million to $4 million, depending on configuration and support contracts. A modern AI‑focused data centre can host up to 100 000 such racks, meaning total hardware spend can quickly eclipse tens of billions of dollars. The power draw also balloons: while a single GB10 draws 140 W, a fully populated GB300 rack can consume several megawatts, necessitating advanced cooling and power‑distribution infrastructure.

Why the government is funding the purchase

U.S. intelligence agencies—including the CIA and NSA—have flagged AI as both a strategic tool and a national‑security threat. To stay ahead of private‑sector rivals like Anthropic, OpenAI and DeepSeek, the agencies need compute power that matches the scale of commercial models such as Claude, GPT‑5.5 and V4. The $9 billion request, still awaiting congressional approval, would fund the acquisition of GB10‑based hardware and the construction of dedicated data‑center capacity. An interim $800 million from the defense budget has already been redirected to purchase cloud compute, while agencies continue to test Anthropic’s Mythos model despite supply‑chain concerns.

Broader context and next‑gen Vera Rubin platform

The $9 billion request is modest compared with other federal AI investments. Amazon Web Services alone is earmarking $50 billion to upgrade its government cloud, a platform heavily used by U.S. intelligence. Nvidia is already planning the successor to Grace‑Blackwell: the Vera Rubin platform, which will pair a custom Arm‑based CPU called Vera with a high‑performance Rubin GPU. Vera Rubin promises up to ten times the performance‑per‑watt of its predecessor and will leverage HBM4 memory, positioning it for the next wave of AI workloads.

Implications for the AI hardware market

If Congress green‑lights the funding, demand for high‑end AI chips could surge, tightening an already strained supply chain. OEMs such as Dell already sell GB10‑based desktop AI systems starting around $5 000, while rack solutions are priced in the multi‑million‑dollar range. The influx of government money may accelerate price reductions for lower‑tier customers but could also push manufacturers to prioritize custom, large‑scale orders. Competitors like Qualcomm and AMD will be watching closely, as the U.S. government’s procurement choices can set de‑facto standards for future AI infrastructure.

Looking ahead

Beyond the immediate procurement, the request signals a broader strategic shift: AI is now framed as an arms race where hardware supremacy is as critical as algorithmic breakthroughs. As the Vera Rubin platform matures, the government’s hardware roadmap will likely evolve to include even more power‑efficient designs, tighter integration with secure cloud services, and stricter supply‑chain vetting. Stakeholders across the AI ecosystem—from chip designers to cloud providers—should prepare for a landscape where billions of dollars flow toward ever‑more specialized silicon.