NVIDIA and Oracle have announced the construction of the largest AI supercomputer of the Department of Energy (DOE) in the United States, a system named Solstice that will incorporate 100,000 NVIDIA Blackwell GPUs and will be installed at the Argonne National Laboratory. The project, introduced during GTC Washington, is complemented by a second system, Equinox, with 10,000 Blackwell GPUs which is expected to be available in the first half of 2026. Both systems will be interconnected with NVIDIA networks and will jointly deliver 2,200 exaflops of AI performance, designed for training and reasoning with state-of-the-art models oriented toward large-scale scientific discovery.
The declared goal is to accelerate the productivity of U.S. R&D across sectors ranging from health to energy, including advanced materials and climate. Additionally, it aims to democratize the use of AI agents—the so-called “agentic scientists”—that assist researchers in tasks such as hypothesis exploration, experimental design, analysis, and validation. To achieve this, Argonne’s supercomputers will leverage NVIDIA Megatron-Core (large-scale training) and the TensorRT inference stack, capable of scaling frontier models and reasoning workflows reproducibly.
Two systems, one shared goal: next-generation open science
- Solstice will serve as the pillar of the initiative, with 100,000 Blackwell GPUs and state-of-the-art interconnection, designed for training large-scale AI models and multimodal reasoning agents.
- Equinox, with 10,000 Blackwell GPUs, will enhance Argonne’s computational capacity starting in the first half of 2026, distributing workloads and reducing research timelines for thousands of scientists.
According to the lab, both systems will integrate with leading DOE facilities like the Advanced Photon Source, enabling seamless flows of simulation, experimental data acquisition, and AI-driven analysis, creating a streamlined end-to-end workflow.
Beyond hardware: AI agents and “agentic AI” for public R&D
One key focus of the announcement is the promotion of AI agents in open science. Unlike traditional AI—which focuses on isolated tasks—these agents connect multiple steps: formulating hypotheses, consulting databases, running simulations, evaluating results, and iterating. With Solstice and Equinox, Argonne, the DOE, and NVIDIA aim to industrialize this approach, providing the community with infrastructure and libraries to develop “virtual scientists” that complement human teams, ensuring traceability, verifiability, and risk control typical of public sector projects.
A public-private collaboration model
The DOE frames the project within a new collaboration paradigm with industry, blending investment and use cases to accelerate deployment. In the statement, U.S. Energy Secretary Chris Wray emphasizes that “winning the AI race requires creative alliances”, and that Argonne’s two systems represent a new approach to turning shared innovation into a national strength. Jensen Huang, founder and CEO of NVIDIA, states that AI is the “engine” powering a new wave of scientific discoveries and that accelerated computing will enable progress across diverse disciplines.
What it means for researchers
- Unparalleled capacity for training and reasoning: the combined 2,200 exaflops of AI open doors to frontier models and reasoning agents that were previously impractical in the public sector.
- Integrated workflows: with Megatron-Core and TensorRT, the same stack supports training and deployment at scale, reducing iteration cycles.
- Connection to experimental facilities: integration with sources like X-ray, neutron instruments, and other large DOE tools will close the loop between experimental data and modeling with unprecedented latencies and volumes.
- Productivity boost: AI agents can handle repetitive tasks like literature searches, data cleaning, and simulation orchestration, freeing human teams to focus on science.
Risks and precautions
As with all cutting-edge infrastructure, success will depend on:
- Governance of models and metrics (accuracy, bias, reproducibility).
- Security and data sovereignty within a federal and multi-institutional framework.
- Training and adoption: thousands of researchers will need to learn how to design, measure, and audit agents operating with conditional autonomy.
- Availability: Equinox is planned for the first half of 2026; Solstice will require logistical ramping, energy, and cooling resources commensurate with its scale.
A milestone in the national laboratories’ road map
The announcement of Solstice and Equinox fits into a broader vision: seven new systems at Argonne and Los Alamos utilizing NVIDIA technology to drive scientific leadership and strengthen strategic capabilities. Concurrently, NVIDIA is advancing the creation of an AI Factory Research Center in Virginia, serving as a testing ground for super-factories of AI across multiple generations. Overall, this signifies a paradigm shift in how science is produced with AI in the U.S.
Frequently Asked Questions
What exactly are Solstice and Equinox?
Solstice will be the primary DOE supercomputer at Argonne with 100,000 Blackwell GPUs. Equinox, with 10,000 GPUs, will bolster capacity starting in the first half of 2026. Both systems will be interconnected via NVIDIA networks and will provide 2,200 exaflops of AI for training and deploying frontier models and reasoning agents.
What does “2,200 exaflops of AI” mean?
It’s an aggregated measure of AI compute (using reduced precision formats typical in AI), reflecting the power for massive training and inference. It should not be directly compared 1:1 with exaflops in FP64 used in traditional supercomputers.
How will “agentic” AI serve scientific research?
To automate workflow chains: literature searches, proposing experimental designs, simulating, analyzing, and iterating. These agents operate in a “human-in-the-loop” paradigm with decision logs, ensuring productivity gains while maintaining control.
What software will the systems utilize?
The Megatron-Core library for large-scale training and the TensorRT stack for inference, along with NVIDIA tools for agents and reasoning. The goal is to unify the entire stack from training to deployment.
Sources (selected): Official NVIDIA Newsroom announcement on Solstice and Equinox (Argonne), including figures of 100,000 and 10,000 Blackwell GPUs, 2,200 exaflops of AI, timelines, and agentic AI objectives; and NVIDIA’s overarching statement on expanding AI infrastructure at Argonne and Los Alamos.