In a major step for US science and AI, the Department of Energy and AMD are teaming up to build two powerful supercomputers at Oak Ridge National Laboratory—Lux and Discovery. Lux arrives first in 2026 as America’s AI factory, powered by AMD GPUs and CPUs to speed up breakthroughs in materials, medicine, and energy. Discovery follows in 2028, blending AI and traditional high-precision computing into one seamless system. Together, these machines represent a $1 billion public-private partnership designed to strengthen America’s technological independence, advance research, and unify AI and high-performance computing into a single, sovereign scientific platform.
AMD powers US sovereign AI factory supercomputers, accelerating an open American AI stack. (Source: AMD)
In late October 2025, the US Department of Energy (DOE) and AMD announced a joint effort to expand the nation’s high-performance computing and artificial intelligence capabilities through the development of two new supercomputers at Oak Ridge National Laboratory (ORNL): Lux and Discovery. Together, they represent a $1 billion public-private investment and a strategic milestone in America’s pursuit of AI-driven scientific discovery, energy research, and national security innovation.
The partnership brings together the DOE, ORNL, AMD, HPE, and Oracle, forming an integrated alliance of government laboratories and private industry. ORNL will provide the infrastructure and operational environment, while the corporate partners will design, build, and co-finance the systems. The collaboration extends beyond traditional procurement: The government gains early access to frontier computing for science, while the companies secure a proving ground for next-generation architectures that support both commercial and sovereign AI use.
Lux, scheduled for deployment in early 2026, will become the United States’ first dedicated AI factory for science. It will be co-developed by ORNL, AMD, HPE, and Oracle Cloud Infrastructure, powered entirely by AMD silicon, including Instinct MI355X GPUs, Epyc CPUs, and Pensando networking technologies. The architecture is optimized for AI model training, inference, and data-intensive workflows that support scientific computing, material simulation, and bioinformatics. According to ORNL Director Stephen Streiffer, Lux will deliver roughly three times the AI capability of current systems at the lab and will be central to accelerating research in fusion energy, cancer modeling, and advanced manufacturing.
Lux’s design reflects a shift in supercomputing philosophy. Rather than focusing solely on floating-point throughput, it emphasizes the balance between AI and data handling—scaling performance across thousands of nodes while maintaining efficiency. AMD’s Subthreshold Power Optimized Technology (SPOT) and the energy-aware design of its GPUs allow high computational density without unsustainable power draw. Each MI355X accelerator operates at around 1400W board power, and collectively, the system will reach levels of AI throughput not previously achieved on US soil.
Energy Secretary Chris Wright described Lux as the first example of a commonsense partnership between industry and the government in the AI era. He noted that the system would strengthen the DOE’s AI leadership and provide the infrastructure needed for scientific teams to model complex physical systems—from the plasma dynamics of nuclear fusion reactors to the genomic signatures of disease. Wright also said Lux’s deployment, combined with Discovery’s future capabilities, would give the US a path toward harnessing fusion energy within a few years and making cancer a manageable disease within the decade.
Lux’s success will rely not only on its hardware but also on its integration with Oracle’s sovereign cloud infrastructure, which will enable secure collaboration across agencies and research partners. This hybrid model links ORNL’s on-premises resources with distributed AI capacity, forming a national network for scientific computing. The DOE calls this approach federated AI infrastructure, designed to maintain data security while scaling across institutions.
While Lux targets the near-term demand for AI-driven science, Discovery, arriving later in the decade, will extend the model toward exascale performance and full-spectrum high-precision computation. Set for delivery in 2028 and operational in 2029, Discovery will succeed Frontier and serve as the DOE’s flagship supercomputer for the next generation. Built on HPE Cray GX5000 technology, it will use AMD’s upcoming Epyc Venice CPUs and Instinct MI430X GPUs, components optimized for both scientific accuracy and large-scale AI training. The MI430X architecture integrates one Epyc processor with four GPU dies in a single package, emphasizing bandwidth, low latency, and a unified memory model.
Discovery’s system design follows a bandwidth-everywhere principle. It expands node and global interconnect throughput far beyond first-generation exascale machines while keeping overall power budgets constant. This allows it to deliver higher computational output without a proportional increase in energy cost. Its software environment remains compatible with Frontier, ensuring that existing applications migrate without major rewrites.
The Discovery program embodies the DOE’s intent to merge AI and traditional simulation science into one continuum. With increased precision at both FP32 and FP64 levels, the system will simulate the behavior of advanced materials, battery chemistries, catalysts, and next-generation semiconductors. It will also model biological systems at atomic and molecular scales, supporting drug design and genomic analysis. ORNL expects Discovery to accelerate innovation cycles across energy and materials science, while supporting national security missions that depend on predictive modeling.
AMD Chair and CEO Lisa Su described Discovery and Lux as the company’s most significant collaboration with the DOE to date, a demonstration of how high-performance and AI computing technologies can jointly serve science and national competitiveness. Su emphasized that AMD’s full technology stack—CPUs, GPUs, interconnects, and software—was engineered for open standards, enabling transparency and long-term sovereignty for US research institutions.
From an architectural perspective, Discovery builds directly on the lessons of Frontier but transitions from a pure exascale machine to a hybrid system where AI and simulation operate side by side. Its open-source software environment allows researchers to train models on experimental data and immediately apply those models in high-fidelity simulations. This iterative loop—data to model to simulation—embodies the DOE’s vision of AI-driven science.
HPE President and CEO Antonio Neri described the collaboration as an evolution of supercomputing in the AI era, combining the determinism of HPC with the adaptability of machine learning. By merging these disciplines under one consistent programming framework, ORNL hopes to reduce researchers’ time from concept to discovery.
The two systems represent complementary phases of a single national strategy. Lux addresses the current gap in large-scale AI resources for science, enabling researchers to train and deploy foundation models domestically. Discovery extends that capability into the next decade, unifying HPC and AI within one scalable platform. Together, they advance the US AI Action Plan, strengthen the country’s technological autonomy, and create the foundation for secure, sovereign AI infrastructure.
For AMD, the projects confirm its growing influence in scientific computing and signal that AI architecture will define future supercomputing design. For the DOE, they establish a template for future collaborations where federal science, academia, and private industry jointly build shared infrastructure. When both systems are online, Oak Ridge will host an unprecedented combination of AI throughput and scientific precision—tools intended not for commercial chatbots, but for breakthroughs in fusion energy, materials discovery, and biomedical research that shape the nation’s technological landscape for decades to come.
What do we think?
This is another big win for AMD. Lux and Discovery open clear opportunities: a sovereign AI/HPC stack anchored in open standards, a unified simulation-plus-AI workflow (Frontier continuity), and durable public-private capacity that can accelerate fusion, materials, and biomedical pipelines. AMD benefits from platform credibility for Epyc Venice and Instinct MI430X, while ORNL gains bandwidth-centric architectures and federated, secure integration with OCI.
Threats remain material. Schedule risk to 2026–2029 could erode momentum. Power and cooling envelopes may constrain utilization even with efficiency gains. HBM supply, advanced packaging, and NIC/interconnect availability create single-point bottlenecks. Software portability versus Nvidia’s CUDA gravity could slow adoption. Multi-tenant sovereignty demands rigorous isolation and auditability. Funding cycles and election-driven priorities may shift scope. Success hinges on predictable delivery, robust toolchains, and sustained ecosystem incentives.
WHAT DO YOU THINK? GOOD ENOUGH TO TELL YOUR FRIENDS ABOUT IT? YOU HAVE OUR PERMISSION TO SHARE IT.