DOE is funding a grid model designed to evaluate 1 billion scenarios a day. But queue delays arise as much from process, staffing, speculative requests, and construction as from solver time.
On September 1, U.S. Department of Energy Office of Electricity Assistant Secretary Catherine Jereza visited Brookhaven National Laboratory to announce a three-year Genesis Mission Phase II project for a grid foundation model. The goal, Brookhaven said, is to build a model capable of simulating 1 billion scenarios in 24 hours as utilities assess new loads.
That morning, Reuters published an analysis of utility and grid data showing more than 700 gigawatts of electricity requests from very large users, mostly data centers, across parts of the Midwest, Mid-Atlantic and South. That is more than 10 times industry estimates of current U.S. data-center power use.
The shared date was a coincidence of the news cycle. The shared subject was the interconnection process: the technical and administrative work required before a utility or transmission provider agrees to serve a large new load. Together, the announcements raise a narrower question: If AI can run some grid calculations much faster, how much faster can it make the queue?
The project is titled "Foundation Models for the Electric Grid: From Proof of Concept to Real-world Impacts," or GridFM 2.0. DOE's Office of Electricity announced it as an $11.5 million research project. Brookhaven described a $14.2 million total project, with $3.9 million allocated to the lab and the balance shared among collaborators. Neither announcement explains whether the $2.7 million difference reflects cost share, in-kind support, or another accounting distinction.
The $293 million Genesis Mission funding opportunity, previously covered by SCN, set a Phase II range of $6 million to $15 million over three years, a figure DOE's RFA FAQ calls an estimate derived from Phase I budgets and says can move during award negotiation. Idaho National Laboratory's $60 million Prometheus project was identified by INL as the first Phase II award announced under the program. Brookhaven's announcement places GridFM 2.0 among the first publicly identified Phase II projects, but the available sources do not establish a definitive ordinal.
Within the wider Genesis program, the American Science Cloud and its planned federation of supercomputers provide computing infrastructure, while the Genesis-Science-1 partnership with Arcee AI focuses on a model of scientific literature. GridFM 2.0 targets a different layer: a foundation model of a physical system.
Brookhaven names 26 collaborating organizations. They include the National Laboratory of the Rockies and six universities; AWS, Microsoft, NVIDIA, IBM Research and other companies; the New York Power Authority, Alabama Power, the Long Island Power Authority, National Grid, Con Edison and Great River Energy; and the New York Independent System Operator, NYSERDA, the National Rural Electric Cooperative Association, GE Vernova Research and LF Energy. DOE says the project will include two deployments with utility partners, but neither announcement identifies the hosts.
The award is intended to scale work already underway. GridFM emerged from a 2024 effort led by researchers at IBM and partner organizations. A perspective paper posted to arXiv in July 2024 argued that a model trained across grid topologies and operating conditions could support multiple planning and operations tasks. Its authors came from IBM, Hydro-Québec, ETH Zürich, Argonne and other institutions.
LF Energy's Technical Advisory Council approved the open-source project in October 2024, with IBM and Hydro-Québec listed as leads. OpenGridFM is now classified as an Incubation-stage project.
Hendrik Hamann, now Brookhaven's chief AI scientist for Innovation, Science, and Security, worked at IBM Research for 26 years, most recently as chief science officer responsible for AI geospatial foundation models. Two related deliverables arrived this summer. On June 30, Brookhaven and AWS announced GridSearch, an application intended to prescreen thousands of potential interconnection locations. On August 10, researchers posted GENCO, a neural solver for steady-state grid analysis, along with an open-source development framework.
The project also addresses an input that IBM says is scarce: access to real-world grid data. Utilities restrict detailed network and operating data for security and commercial reasons. That makes utility and system operator participation central to the project.
Large-load interconnection studies do more than run one kind of calculation. Depending on the project and region, a transmission provider may run power-flow, stability, and short-circuit studies, as well as electromagnetic-transient simulations. RMI's 2026 overview of large-load interconnection describes that broader study stack.
GENCO addresses three steady-state tasks: power flow, optimal power flow, and state estimation. Power flow is the most direct link to interconnection screening because it calculates bus voltages and line flows for a specified network, generation pattern and load. Optimal power flow adds an optimization objective and operating limits; it matters to wider grid planning and operations, but it is not a universal substitute for the other analyses in an interconnection study.
GENCO uses a graph-transformer architecture to map grid topology and injections to an AC operating state, followed by physics-based corrective steps that reduce power-balance errors. It can batch many cases on a GPU. The paper's runtime analysis compares an NVIDIA H100 with an 84-core AMD EPYC 9634 system running parallel classical solvers, avoiding the misleading practice of comparing a parallel GPU model with a single-threaded baseline.
The published results are benchmark results, not timings for an end-to-end interconnection study:
Grid (buses) | Power flow: speedup vs. AC-PF | Optimal power flow: speedup vs. AC-OPF |
|---|---|---|
14 | 0.8x | 16.1x |
118 | 1.8x | 48.4x |
500 | 2.1x | 54.3x |
2,000 | 28.2x | 84.5x |
10,000 | 29.1x | Not reported |
Source: GENCO paper, version 2, Tables 6 and 7. Power-flow figures use GENCO Tiny; optimal-power-flow figures use GENCO Small.
For power flow on grids through 500 buses, the paper says the classical AC solver remains preferable. At 2,000 buses and above, GENCO returns a full AC state, including voltage magnitudes and reactive power, at roughly twice the runtime of the simpler DC approximation. For optimal power flow, GENCO is faster at every reported grid size, with a worst-case optimality gap of 0.3% in the reported tests.
The limitations matter. The authors say zero-shot transfer to an unseen grid remains an open challenge. They report real-grid validation on a roughly 1,200-bus Hydro-Québec network using a year of SCADA data, but that is not the same as production deployment in an interconnection workflow. IBM's account of the work says GENCO is not positioned to replace established tools in real-time grid operations. It describes a hybrid approach in which the neural model screens many cases, residual checks identify suspect results, and classical solvers provide authoritative answers for the cases that require them.
Three different kinds of numbers appear in the public record, and they should not be treated as interchangeable.
First are the measured GENCO benchmark results: up to 29.1x faster than the paper's AC power-flow baseline and up to 84.5x faster than its AC optimal-power-flow baseline on the reported grid sizes.
Second are broader statements about early applications. A Stony Brook account of Hamann's July 22 Genesis Mission Summit keynote says some analyses have shown acceleration of up to 1,000x. It also quotes him saying data-center interconnection simulation studies can move from "months, sometimes years" to "minutes." The article does not define the timed workflow, baseline, hardware, or share of the interconnection process included in that comparison.
Third are the GridFM 2.0 program goals. DOE says the project aims to evaluate 1 billion potential scenarios in 24 hours, increase planning throughput by more than 10,000x, and make key grid calculations more than 1,000x faster than traditional approaches.
One billion scenarios a day is about 11,600 scenarios per second, sustained. DOE has not published the scenario definition, hardware configuration, model size, or validation workflow behind that target. It therefore cannot be derived from the paper's 30x result, nor used to estimate an accelerator count from the public information. The benchmark and the program target describe different layers of performance.
The near-term role supported by the available evidence is a screening layer: use the neural surrogate to evaluate many candidates, then send selected or questionable cases to established engineering tools. Whether utilities will accept that workflow for formal studies is one of the questions the two planned deployments are meant to test.
Generator interconnection and large-load interconnection should not be conflated. Lawrence Berkeley National Laboratory's Queued Up series measures generation and storage projects seeking transmission interconnection. Large loads such as data centers follow different processes, with responsibilities divided among transmission providers, FERC, and state regulators.
The generator data still illustrate the scale of the process problem. LBNL reports that a typical project entering service in 2024 spent 55 months between its interconnection request and commercial operation. That period includes more than computation: queue administration, studies and restudies, project development, permitting, financing and construction.
DOE's i2X Transmission Interconnection Roadmap attributes backlogs and delays to rapid growth in requests, inefficient processes and staffing constraints. It also describes how withdrawals can force restudies and recommends stronger readiness requirements, firm study timelines, better data access and greater automation.
FERC's response to generator backlogs was primarily procedural. Order No. 2023, issued July 28, 2023, in docket RM22-14, moved public-utility transmission providers toward first-ready, first-served cluster studies, established a 150-day cluster-study timeline and imposed delay penalties, subject to safeguards and compliance proceedings.
Large-load rules are now moving on a separate track. On June 18, 2026, FERC issued show-cause orders to the six regional grid operators under its jurisdiction. The orders direct each operator to justify its existing tariff or propose reforms. The first of the five reform categories is the study process, and Day Pitney's summary of the orders says the Commission's targeting study timelines of 60 to 90 days. SCN previously examined those proceedings when the proposed large-load fast lane arrived with a curtailment clause.
Texas is different because most of the ERCOT system sits outside FERC's traditional rate jurisdiction. After Governor Greg Abbott ordered an audit of data-center requests on August 3, ERCOT paused advancement of its Batch Zero process. ERCOT has said it aims to file its Batch Zero eligibility-verification report by December 10, according to Utility Dive's report on the commission meeting.
The case against treating solver time as the main queue bottleneck is straightforward:
The case for more study throughput is also real:
The most plausible near-term way GridFM could shorten a queue may be by improving what enters it. A developer that can rank thousands of connection points before filing has less reason to submit multiple requests simply to discover which site is viable. That could reduce speculative volume, but it doesn't prove an end-to-end reduction in interconnection time.
The next evidence will come from deployment. The GridFM community's sixth conference is scheduled for October 29 and 30 in Porto. FERC must decide what action to take after the six regional operators' August responses to its show-cause orders. ERCOT is targeting December 10 for its Batch Zero verification report. DOE and Brookhaven have not identified the two utilities that will host the GridFM 2.0 deployments.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
