NVIDIA and Oracle plan DOE's largest AI supercomputer for scientific discovery

NVIDIA and Oracle said on March 16, 2026 that they will build the U.S. Department of Energy's largest AI supercomputer for scientific discovery at Argonne National Laboratory. The announcement centers on two systems, Solstice and Equinox, and positions the project as a public-private effort to expand compute available to researchers working across science, energy, and national security.

The headline numbers are large even by current frontier-AI standards. NVIDIA said Solstice will use 100,000 Blackwell GPUs, while Equinox will use 10,000 Blackwell GPUs and come online in the first half of 2026. The two systems are expected to deliver a combined 2,200 exaflops of AI performance and be connected with NVIDIA networking. Oracle is contributing cloud and infrastructure support through OCI, while the DOE and Argonne will operate the systems as part of broader research programs.

• Solstice: 100,000 NVIDIA Blackwell GPUs
• Equinox: 10,000 NVIDIA Blackwell GPUs
• Location: Argonne National Laboratory
• Combined target: 2,200 exaflops of AI performance

According to the release, the systems will support training of frontier models and AI reasoning models for open science, using NVIDIA Megatron-Core for model development and TensorRT for inference. Argonne also said the new machines will connect to experimental facilities such as the Advanced Photon Source, which matters because it ties large-scale AI compute directly to real scientific instruments rather than treating the cluster as a standalone corporate asset.

The strategic angle is as important as the hardware. NVIDIA framed the deployment as a foundation for agentic scientists and faster R&D, while the DOE described it as part of a new partnership model that mixes public research goals with industry investment. For research institutions, that could mean quicker iteration on materials, healthcare, energy, and other data-intensive workloads if the infrastructure is delivered on schedule.

What makes the story significant is not only the raw GPU count, but the fact that one of the largest announced AI systems is being aimed at public scientific work instead of a purely commercial chatbot or cloud product. The announcement is still a forward-looking plan, so timelines and realized throughput will depend on execution, but it sets a high bar for how national labs and AI vendors may collaborate in the next wave of scientific computing.