Nvidia to build seven supercomputers for US Energy Department

Nvidia CEO Jensen Huang announced the company will build seven new supercomputers for the US Department of Energy, including one in partnership with Oracle that will use 100,000 Blackwell chips.

The supercomputers will support nuclear weapons research and alternative energy projects like nuclear fusion.

Huang also said Nvidia currently has US$500 billion in bookings for its Blackwell and Rubin AI chips over the next five quarters.

He told reporters that Nvidia has not applied for US export licenses for its newest chips to China, citing the Chinese government’s decision to restrict Nvidia’s market access.

Nvidia revealed a US$1 billion investment for a 2.9% stake in Nokia, aiming to target the AI communications sector, and introduced Arc, a product line for telecom equipment.

The company also announced a partnership with Palantir focused on commercial logistics solutions, and unveiled Hyperion, a new self-driving car technology platform, with plans to work with Uber on a robotaxi network.

Nvidia’s GTC event was held in Washington, D.C., for the first time.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Nvidia’s $500 billion in bookings doesn’t guarantee near-term revenue conversion

Huang said $500 billion in Blackwell and Rubin bookings span the next five quarters. The piece does not say if these are binding contracts, conditional orders, or prepayments. That detail sets revenue timing.
Q2 fiscal 2026 revenue came in at $46.7 billion, with Blackwell Data Center up 17% sequentially 1. That pace sits far below what $100 billion per quarter would deliver if all bookings turned into sales.
The gap between bookings and revenue depends on manufacturing capacity, High Bandwidth Memory (HBM) supply, and advanced packaging limits (the complex chip assembly step). The announcement left these factors out, yet they shape delivery on the stated timeline.

Secondary GPU market opportunity emerges from massive fleet refreshes

A US Department of Energy (DOE) supercomputer with Oracle will use 100,000 Blackwell chips. That push, along with wider Blackwell rollouts, will displace many H100 and A100 GPUs (prior-generation Nvidia data center GPUs) as operators refresh gear.
Cloud resellers, managed service providers, and AI infrastructure startups can source these GPUs at discounts. They can refurbish units and sell lower-cost compute for inference or smaller training work to budget-conscious firms.
Timing drives success. Buyers need to acquire GPUs before supply floods the market, then build ties with hyperscalers (the largest cloud providers operating at massive scale) or data center operators running refreshes now.