DOE Advisers' Genesis Mission Report Sets Fusion, Biology, and Magnet Targets Without Cost Estimates

The Department of Energy's Office of Science has published Genesis Mission Frameworks for AI-Accelerated National Breakthroughs (DOE/SC-2031), a report dated September 2026 from a subcommittee of its Office of Science Advisory Committee (SCAC). Under Secretary for Science Darío Gil introduced it on September 25. The report answers a charge Gil sent the committee on March 26 and works through three examples of AI-accelerated science in depth: a biology campaign, a push to put fusion power on the grid, and a five-year program, titled Magnet Sovereignty, to rebuild U.S. production of rare-earth permanent magnets.

SCAC, which advises the Under Secretary and the Office of Science, approved the subcommittee's response unanimously, with 22 votes in favor, at its July 17 meeting, according to the meeting minutes. The eight-member subcommittee, chaired by Suresh Garimella of the University of Arizona, drew its members from universities, Lawrence Berkeley National Laboratory, Japan's RIKEN, Google DeepMind, Two Sigma, and the Simons Foundation. The report is advice. It is not DOE policy, a funding decision, or an appropriation, and its authors frame what they recommend as suggestions and targets.

For people who build and run supercomputers, two passages carry most of the weight. The fusion chapter argues that fusion's binding constraint is the coordination of resources DOE already owns, and it expects AI to pay off there. The magnet chapter invokes computing up to the zettascale without defining the term. The report attaches no cost estimates to its proposals, and its chapter on the Genesis platform names no DOE system. DOE has published cost estimates for platform hardware in a separate document. The proposed items on that list are being ranked by a different SCAC subcommittee, which reports in November. Gil's charge asked the Genesis subcommittee to coordinate with it on infrastructure.

Fusion's bottleneck, as the report describes it

DOE has invested billions in physics codes, exascale supercomputers, and experimental facilities, the fusion chapter says, yet those resources mostly operate in isolation, with people translating results by hand from one domain to the next. The chapter's thesis follows: "The rate-limiting step is not computational power but the time to orchestrate these resources into coherent, intent-driven campaigns."

The proposed remedy is intent-driven orchestration, in which a researcher states an objective and AI assembles, and keeps reconfiguring, a campaign that spans simulations and experiments at several institutions. The integrating layer would be DOE's AI-fusion Digital Convergence Platform (DCP), which combines simulation codes, foundation models, surrogate models, and digital twins. Under it sits the Fusion Energy Data Ecosystem and Repository (FEDER), built on the FAIR data principles: findable, accessible, interoperable, and reusable. The authors pitch the architecture as a pathfinder for AI across DOE. They are the chapter's lead, David Spergel of the Simons Foundation and the Flatiron Institute, with Tammy Ma of Lawrence Livermore National Laboratory, Mark Papermaster of AMD, and Derek Sutherland of Realta Fusion. The last three are SCAC members from outside the subcommittee, and Ma and Sutherland also sit on the Facilities Subcommittee.

On that basis, the chapter sees an opportunity "to accelerate fusion development by 10- to 100-fold through ecosystem-wide orchestration." The figure is a target, and the report shows no derivation for it or breakdown of where the time goes in a fusion design cycle today. The same range appeared on the subcommittee's July 17 slides. The minutes summarize Spergel as putting the potential at tenfold or more and record him doubting that experimental measurements themselves can be sped up much, though AI can extract more data from each one. Asked about hype generated by the private sector, he cautioned against exaggerated claims about fusion progress and encouraged DOE to be the voice of reason.

From an engineering perspective, this is a claim about the software stack: workflow systems, schedulers, data movement, and the interfaces between simulations and instruments. DOE's own summary of proposed and in-construction facilities lists orchestration across different resources, with different types of scheduling, among the design goals for NERSC-10, the Doudna system under construction at Berkeley Lab. SCN has covered a smaller version of the pattern, an Argonne agent stack that turned one plain-English prompt into 11,182 simulations on Aurora. The fusion chapter proposes orchestration across the entire field, with a universal data-driven simulator open to the entire public-private fusion ecosystem by 2030.

For evidence from live plasmas, the chapter says AI models have forecast tearing-mode instabilities up to 300 milliseconds before onset and then steered plasmas away from disruption, including on tokamaks the models were not trained on. One of the two papers it cites, a 2024 study that demonstrated AI tearing avoidance on DIII-D, reports the 300-millisecond forecast from a discharge in which the controller could not cut beam power any further and the instability was not avoided.

The targets, read as targets

The fusion chapter's headline goal is a U.S. pilot plant delivering net electricity to the grid between 2030 and 2035, and its implementation table places the plant's launch at 2035. The report says this advances the target in the National Academies' 2021 study by five years, which called for an operational pilot plant in the 2035 to 2040 time frame. DOE's Fusion Science and Technology Roadmap, finalized on June 9, 2026, already aims for pilot plants and commercial fusion power in the mid-2030s, and it states that DOE's ability to support those milestones and timelines "is contingent on future public-private partnerships and future Congressional appropriations."

The chapter's milestone tables run through Year 1, Year 2, 2030, and 2035, and the report does not tie Year 1 to a calendar year; the July slides placed the near-term fusion goals on a 2026-2027 axis. Year 1 covers interoperable data across major U.S. fusion facilities and AI disruption-avoidance controllers operating on DIII-D in San Diego and KSTAR in South Korea. Year 2 brings a first fusion foundation model, a finalized Nuclear Regulatory Commission licensing framework for fusion, and the first plasma campaigns on the SPARC tokamak. By 2030, a universal digital twin trained on U.S. and allied data would help down-select among confinement approaches.

The magnet chapter, led by Nadya Mason of the University of Chicago, sets a five-year target to identify, synthesize, and pilot high-performance permanent-magnet compositions designed for domestic manufacturing and end-of-life recycling. Success, in the chapter's terms, means a 50-fold increase in discovery efficiency, a tenfold reduction in the number of experiments needed, model confidence above 80% before physical validation is prioritized, and more than 75% autonomy for routine experimental cycles. The subcommittee's July 17 slide listed the last three; the 50-fold figure appears only in the final report, without a cited source. The steps run from a national magnetics data foundation in Year 1, through an impurity-tolerant neodymium magnet that works with domestic or recycled feedstock within two years, to at least one rare-earth-light or rare-earth-free magnet validated at pilot scale by years three to five, with recovered material feeding a domestic magnet line.

The biology chapter sets milestones for 2028, 2030, and 2035 across six capabilities. Its 2035 column asks for a multiomic blueprint of any organism within 24 hours of discovery, autonomous digital twins that design synthetic cells, a national network of self-driving biofoundries, and first-attempt scale-up from computer design to more than a million liters. One 2035 modeling target is aimed squarely at computing: mature scaling laws linking compute to capability gains. The chapter's lead author is subcommittee vice chair Pushmeet Kohli of Google DeepMind, where the report's coordinating editor, Ali Douraghy, also works. On economics, the report puts the global bioeconomy at about $4 trillion today and says it could reach $30 trillion by 2050, citing a 2025 report from the UN Environment Programme's Copenhagen Climate Centre. It calls for the U.S. share to rise from a quarter to a third, which would increase the U.S. bioeconomy from roughly $1 trillion to $10 trillion.

Zettascale, undefined

The word appears once in the report, in the magnet chapter's opening summary: "AI and computational resources up to the zettascale are what make that speed achievable." It is absent from the July 17 slides, and the report attaches no arithmetic precision, system, date, or cost to it, although the same chapter endorses low-precision, hardware-aware computing to cut the energy cost of model development and to run models at processing sites.

A zettaflop is 1,000 exaflops, and how far off that is depends on the arithmetic. On the June 2026 TOP500 list, which ranks systems by measured 64-bit (FP64) performance on the High Performance Linpack benchmark, the leader is China's CPU-only LineShine at 2,198.4 petaflops, with El Capitan second at 1,809. At FP64, a zettaflop is more than 450 times the fastest measured system. The low-precision AI figures vendors quote put it much closer. NVIDIA said in October 2025 that Argonne's planned Solstice and Equinox systems, with 100,000 and 10,000 Blackwell GPUs, would deliver a combined 2,200 exaflops of AI performance, roughly 2.2 zettaflops, without stating the precision. SCN's analysis of Oak Ridge's Lux, whose FP4 and FP64 ratings sit 126 to 1 apart, shows how wide that spread runs on a single machine.

What the charge asked, and what came back

What follows is SCN's comparison of the charge with the report. Gil's letter asked for an assessment of the current landscape and a decadal roadmap by July 2026, built around three questions. The first asked what should be included in the Genesis platform and what the refresh plan is. The other two covered what should guide a world-leading portfolio of national science and technology challenges and how DOE should work with partners to grow an AI-ready workforce. The letter also called for a roadmap "suggesting milestone-based investments spanning facilities, research, and workforce development."

The subcommittee met the July deadline, presenting on July 17 under the title A Genesis Mission Roadmap for AI-Accelerated National Breakthroughs; the published version is titled Frameworks. On the platform, the report's first chapter answers in concepts. It calls the American Science Cloud the digital fabric linking DOE's computing, data, and experimental facilities, says purpose-built AI hardware is needed alongside the leadership-class and exascale systems already at the labs, and says DOE is building, and will operate, AI machines through federal-industry partnerships. No systems, capacities, architectures, refresh cadences, or costs appear. The domain chapters add data layers, from the DCP and FEDER to a biology data lakehouse and a national magnetics data foundation. The partnership language fits models SCN has examined, from Genesis-Science-1, where Arcee AI secures the compute, to DOE's federation approach to the American Science Cloud.

Workforce gets a single section, describing pathways from K-12 to postdoctoral positions without mechanisms or numbers. For the portfolio, the report explains its three picks as aligned with issues of national importance, DOE's existing strengths, and the Genesis vision, and it offers no general method for adding, weighting, or retiring challenges. It calls fusion one of 26 Genesis challenges, the list DOE published on February 12, 21 of which were targeted by the first $293 million Genesis funding call. DOE's current challenges document lists 33, the count FedScoop reported from the July 22 Genesis Mission summit, and the one Gil's post uses.

Gil's letter also says further charges will follow, and because the work might involve recommendations on infrastructure and facilities, it asks for close coordination with the Facilities Subcommittee. The minutes say the three examples were presented to illustrate the vision and to open a conversation about what else the written report would need.

DOE's cost estimates and the facilities panel

Gil's second charge of March 26 asks SCAC to rank proposed Office of Science facilities and upgrades for 2026 to 2036, including how each would support the Genesis Mission, under scenarios in which $2 billion, $4 billion, or $6 billion is available over the decade. The proposed projects total about $22 billion, according to the July minutes, which record SCAC chair Persis Drell, who also chairs the Facilities Subcommittee, saying the scenarios reflect serious budget constraints and that much larger funding levels would not be realistic. Projects already under construction will not be delayed to make room, and the Facilities Subcommittee will rank only proposed facilities. It is keeping its recommendations confidential until SCAC's November 12 meeting, and the charge asks for a final report by December.

The summary DOE appended to that charge is where the Genesis platform's facilities' cost estimates are provided. It covers facilities under construction as well as proposed ones, and its computing entries fall under Advanced Scientific Computing Research (ASCR), the Office of Science program that runs DOE's open-science supercomputing centers and its research network.

Discovery and Doudna are already being built, so they are not among the projects the panel is ranking. The ecosystem line, whose cost DOE calls highly dependent on available technologies and market forces, runs $4 billion to $6 billion over five years, the same range as the two larger scenarios the panel was given for all new facilities over ten years; the minutes do not say whether the panel is ranking it. The quantum facility sits alongside DOE's separate Quantum Genesis Q competition for lab-verified logical qubits.

The Genesis report's own dollar figures are market sizes and investment comparisons, and the title of a DOE press release appears in its references; it contains no program cost estimate. Cost did come up inside the Genesis process, according to the July minutes. David Siegel of Two Sigma, a subcommittee member and magnet-chapter co-author, suggested adding information on tangible budget figures, and SCAC member Cynthia Friend of the Kavli Foundation, who sits on the Facilities Subcommittee, encouraged the subcommittee to think about the infrastructure needed for experimentation and data collection, especially its costs. The same minutes record approval with the written report expected by the end of August, changed only cosmetically.

Sovereignty and partners

The magnet chapter reads like supply-chain policy: it cites DOE's 2022 supply-chain deep-dive assessment of China's dominance of the rare-earth chain from mining through magnet manufacturing, and adds that the four rare earths in the strongest commercial magnets have no ready substitutes. It keeps the supply chain itself domestic while endorsing cross-border open science in basic AI and materials modeling. The fusion chapter, citing a 2024 Wall Street Journal report, says China outspends the United States by roughly 2-to-1 on public fusion research. And the report describes Executive Order 14363, which launched the Genesis Mission on November 24, 2025, as having "established a sovereign national platform of supercomputers, scientific data, and foundation models."

The fusion plan's blanket and tritium work coordinates with LIBRTI and CHIMERA in the U.K., UNITY-1 in Japan, and UNITY-2 in Canada. RIKEN's Makoto Gonokami co-wrote the magnet chapter, and the July minutes record a five-year, $1 billion Genesis partnership with Japan, announced in June 2026, with each side contributing half. Its eleven joint teams span twelve national laboratories, one Office of Science user facility, and twelve Japanese research institutions.

What to watch

The next checkpoint is SCAC's November 12 meeting in Rockville, Maryland, where, according to the July minutes, the Facilities Subcommittee will present its report. Because the panel has kept its work confidential, that will be the first public look at which proposed facilities it would put first under each scenario. Gil's letter promises further Genesis charges, and the companion report from SCAC's quantum subcommittee, The Quantum Inflection Point, was published on September 17.