THE GENESIS FILE

BLUF: In November 2025, an executive order fused the frontier AI industry with the national-laboratory complex — 17 labs, roughly 40,000 federal scientists, the government's largest-ever AI supercomputers, and what the order itself calls "the world's largest collection" of scientific data — under rhetoric borrowed from the Manhattan Project. Nine months later: two dozen companies are inside on terms the public cannot read, seven of the program's official challenge areas are nuclear-weapons work, veterans' health records and genomes are training federal AI while the government's own auditor flags the vault's locks, no GAO report or Inspector General review exists, mainstream coverage lasted one day, and the program's flagship deadline passed on August 21, 2026 without confirmation, denial, or a single journalist asking.

Nothing in this brief is secret. All of it is unassembled. That is the story.

WHAT IT IS

On November 24, 2025, President Trump signed Executive Order 14363, "Launching the Genesis Mission" — directing the Department of Energy to build a "dedicated, coordinated national effort" the order compares in ambition to the Manhattan Project. Its centerpiece is the American Science and Security Platform (ASSP): DOE's supercomputers plus secure cloud, AI agents and domain foundation models, robotic laboratories, and secure access to federal scientific datasets — which the order describes, without apparent exaggeration, as the world's largest such collection.

The stated goal, per Energy Secretary Chris Wright: "double the productivity and impact of American science and engineering within a decade." The resources: all 17 DOE national laboratories, ~40,000 federal scientists and engineers, and machines like Solstice — 100,000 NVIDIA Blackwell GPUs at Argonne, the largest AI system the U.S. government has ever built.

The order set deadlines like a countdown clock: challenge list in 60 days, compute inventory in 90, data assets in 120, robotics review in 240 — and at 270 days, a demonstration of the platform's "initial operating capability." Hold that last one; it matters later.

Receipts: EO 14363 (whitehouse.gov, Nov 24, 2025); DOE launch release (energy.gov, same day).

WHO'S WHO

• Chris Wright — Secretary of Energy. Fracking executive (Liberty Energy) turned steward of the atomic complex.
• Darío Gil — Under Secretary for Science and Genesis Mission Director. Former head of IBM Research: the man who ran Big Blue's laboratories now runs America's.
• Brandon Williams — Administrator of the NNSA, the agency that builds and maintains nuclear weapons. His labs are inside the Mission.
• Michael Kratsios — the White House science office, coordinating.
• Accenture Federal Services — holds the contract to deliver the platform's "early operating capability." The atomic labs' AI brain is being wired together by a consulting firm.
• The Twenty-Four — companies that signed collaboration agreements with DOE on December 18, 2025: Accenture, AMD, Anthropic, Armada, AWS, Cerebras, CoreWeave, Dell, DrivenData, Google, Groq, HPE, IBM, Intel, Microsoft, NVIDIA, OpenAI, Oracle, Periodic Labs, Palantir, Project Prometheus, Radical AI, xAI, and XPRIZE. The consortium has since grown to roughly 340 members — 157 companies, 142 universities, every national lab.

Every frontier AI lab in America signed. The agreements' terms are not public.

Receipts: DOE releases Nov 24 and Dec 18, 2025; DOE Genesis collaboration pages; Accenture newsroom, 2026.

THE MACHINES

The hardware tells you what the Mission actually is. At Argonne: Equinox (10,000 Blackwell GPUs, 2026) and Solstice (100,000 GPUs, ~200 megawatts, 2027) — built with NVIDIA and Oracle in a novel co-investment structure. At Oak Ridge: Lux (deployed early 2026) and Discovery (2028), with AMD, HPE, and Oracle, a combined public-private investment exceeding $1 billion. DOE calls this a "new public-private partnership model": vendors co-fund the machines and share in the platform they enable.

Why does the weapons department own the nation's biggest computers in the first place? Because of a treaty. The United States has not test-detonated a nuclear weapon since 1992; ever since, the arsenal has been certified by simulation. The supercomputer program exists to replace test shots with physics engines. Genesis inherits that lineage — and extends it from simulating warheads to simulating, potentially, everything.

Receipts: NVIDIA/Oracle/DOE releases Oct 27-28, 2025; AMD/ORNL releases; DOE partnership-model announcement.

THE DATA IS THE PRIZE

Follow the value, and it does not flow the way you'd guess. There is no evidence any company trains its commercial models on government hardware. The flow runs the other way: the companies come to the data. Decades of taxpayer-funded experimental results — physics runs, materials science, genomics, climatology — data too clean, too deep, and too proprietary to ever have been scraped from the internet, now pooled inside the platform the Twenty-Four have agreements to access.

What walks back out? That is the question, and it is legally unanswerable until roughly 2031. The instruments involved — CRADAs, Cooperative Research and Development Agreements — allow companies to negotiate exclusive licenses on resulting inventions, and permit "Protected CRADA Information" to be withheld from the public for up to five years by statute. The AI newsletter The Batch flagged the data-sharing concern at launch ("Genesis Mission Would Share U.S. Data and Resources with Top AI Companies"). No one followed up.

Sidebar — What is a CRADA? A 1986 legal instrument that lets federal labs partner with companies: the lab brings facilities, staff, and data; the company can bring money; the company can take exclusive IP rights; and the details can stay sealed for five years. It is how national-lab science becomes private property, and it has been routine for forty years. What is not routine is applying it, at once, to every frontier AI company in America and the largest scientific dataset on Earth.

Receipts: EO 14363 partnership provisions; Morgan Lewis and Patent Playbook client analyses; deeplearning.ai The Batch; DOE sample OTAs for FOA DE-FOA-0003612 (samples public, actual terms not).

THE NUCLEAR SEVEN

Of the Mission's 26 official science-and-technology challenges, seven are direct nuclear-weapons or national-security work: weapons component design, streamlining weapons production, arsenal modernization by mining classified test data — including digitizing handwritten notes and film from the era of live detonations — automating the DARHT hydrodynamic test facility at Los Alamos, nuclear forensics, attack detection, and nonproliferation satellite analysis. The NNSA issued a request-for-information on AI for national security that explicitly includes, in its own words, "accelerating nuclear weapons development timelines."

The same official challenge list includes root causes of pediatric cancer and improving veteran health. Cures and warheads, one platform, one login.

The steelman is real and deserves its paragraph: simulation-based stockpile stewardship is how America keeps treaty commitments — certifying aging warheads without exploding them is arguably the most responsible version of nuclear custodianship, and AI genuinely helps. But "stewardship" does not require accelerating development timelines or streamlining production — that is the modernization program, a multi-decade, trillion-dollar rebuild, wearing stewardship's lab coat. The question this brief poses is not "why AI for nukes." It is: who decided the pediatric-cancer platform should also accelerate weapons production — and in what public forum was that debated? The record's answer: none.

Receipts: Scientific American, Jun 16, 2026 ("Trump's Genesis Mission Is Putting AI to Work on Nuclear Weapons"); NNSA RFI via FedScoop; White House challenge list, Jul 22, 2026.

THE MONEY THAT ISN'T THERE

The rhetoric is Manhattan Project. The appropriation is roughly $300 million — reshuffled from existing DOE funds, because the executive order carries no money of its own. The July 2026 announcement of ">$5 billion in federal commitments" across 15+ agencies is a commitments figure, not a budget line; the >$800 million in "partner commitments" is compute credits and model access, not cash, and is not itemized.

Meanwhile the FY2027 budget request asks $1.2 billion for a new DOE artificial-intelligence office — while cutting the DOE Office of Science, the nation's basic-research engine, by roughly 13 percent, with some programs cut more than half and environmental research zeroed. Congressional staff worry Genesis will cannibalize already-appropriated science. A DOE official conceded the Mission needs "a lot more" funding. Arati Prabhakar, the previous White House science director, called the concept valid and the simultaneous cuts "extremely destructive to the science and technology enterprise."

A golden era of American discovery, funded by defunding American discovery.

Receipts: The Register, Dec 11, 2025; CRS R48944; AIP FYI; IFP analysis; FedScoop; FABBS.

THE EMPTY CHAIRS

Here is the oversight ledger for the largest science mobilization since Apollo, nine months in:

• Congressional hearings: one — House Science, December 10, 2025, held before the Energy Secretary had testified, over the ranking member's protest.
• GAO reports: zero.
• Inspector General reviews: zero.
• Senate hearings: zero.
• Organized civil-society campaigns: none found.

The sharpest structural critique on record comes not from Congress but from Tri-Valley CAREs — a small community watchdog outside Lawrence Livermore's fence — which warned in August 2026 that AI pipelines built on classified data inside proprietary contractor systems fall outside every normal channel of public review: no FOIA visibility, no per-project funding disclosure, no independent verification standards for AI used in high-consequence nuclear work. The watchdog function for this entire enterprise is currently being performed by a nonprofit that meets in a community room.

Meanwhile, the GENESIS Act (H.R. 6360) would write the executive order into permanent statute.

Receipts: Lofgren and Ross statements, House Science Democrats; Tri-Valley CAREs, Aug 3, 2026; MeriTalk; searches for GAO/IG/Senate records returned nothing — an absence we verified, not assumed.

THE PRESS VACUUM

We measured the coverage. The signing got one day of wire-service attention (November 24, 2025). Since then, across nine months: The Verge — zero articles. Ars Technica — zero. Wired — zero. MIT Technology Review — zero. No sustained podcast or video coverage found. The story lives exclusively in federal-IT trade press (Nextgov, FedScoop, HPCwire), Science, and — the tell — dozens of law-firm client alerts. The audience that showed up was contractors and counsel.

The silence has a mechanism, not a conspiracy: no product to demo, no protagonist to profile, incremental procurement news, a national-security wrapper that discourages leaks, and an AI press corps structurally oriented toward consumer model releases. The Genesis Mission is the exact profile of a story that matters and cannot get covered: important, diffuse, slow, and owned by no beat.

Which means everything in this brief was hiding in plain sight — in press releases, budget documents, and trade reporting — waiting for anyone to assemble it.

Receipts: coverage audit in the case file, with URLs and dates for everything that DOES exist.

THE QUESTION NOBODY ASKED

Executive Order 14363 directed DOE to "seek to demonstrate" the platform's initial operating capability within 270 days: August 21, 2026. As of this writing, ten days later: no DOE release, no White House statement, no press report confirms or denies that the demonstration occurred. The department's demonstration page redirects to a generic mission site. Adjacent milestones exist — first project awards in July, an open-models intake in August — but the flagship deadline itself passed in silence.

Perhaps it happened without announcement. Perhaps the soft "seek to" language absorbed a quiet slip. Either way, the definitive fact is this: no journalist has publicly asked. FractalNode has submitted the question to DOE public affairs and will report the answer — or the silence — in a follow-up. (Pending: inquiry drafted, awaiting send.)

THE PARALLEL FILE: YOUR RECORDS, THEIR MODELS

In July 2026, the Department of Veterans Affairs announced it will contribute veterans' electronic health records and Million Veteran Program genomic data to train Energy Department AI models under the Mission (Stars and Stripes, July 24). The pipeline is not new — MVP data has run on Oak Ridge supercomputers since 2016 — but Genesis industrializes it, alongside HHS's "secure access to the Nation's longitudinal health cohorts," a pediatric cancer data ecosystem, and a multi-agency biothreat surveillance fusion.

Three receipts belong in one paragraph. One: GAO report 26-108651 found cybersecurity deficiencies giving the VA "reduced assurance of the confidentiality and integrity" of Million Veteran Program data — the same dataset now feeding Genesis. Two: the VA's own inspector general reports clinical staff adopting generative AI without oversight. Three: the corporate structure — Oracle, which bought the VA and DoD's health-records vendor Cerner in 2022, whose federal EHR system the VA has linked to six catastrophic-harm incidents including four contributing to veteran deaths, is simultaneously co-building the Mission's flagship supercomputers where veteran health data will train models. The documented data pathway runs government-to-government, not through Oracle's EHR business — we state that plainly. The structure — one company holding the records, building the compute, and sitting in the consortium, with its patient-safety history — is simply a fact pattern the record supports and no oversight body has examined.

One of this brief's authors is a veteran whose records and genome sit in that vault. This section is not abstract.

Receipts: Stars and Stripes Jul 24, 2026; GAO-26-108651; VA OIG via FedScoop; TechTarget/Fierce Healthcare on EHR harm findings; Oracle-Cerner acquisition record.

THE RIVAL, AND THE RACE THAT ISN'T

The Mission's entire justification is the race with China — the executive order invokes the Manhattan Project "against the People's Republic of China" in its own text. So both halves of that justification deserve scrutiny, and we've done the work on both.

The rival's programs are real. China's "AI-Driven Scientific Research Initiative" launched in March 2023 — two and a half years before Genesis. The Chinese Academy of Sciences ships dedicated science foundation models (ScienceOne). And in December 2025, China switched on SCNet: an autonomous AI-science system spanning 30+ supercomputing centers, taking natural-language research instructions and allocating its own compute, claiming a thousand institutional users. Chinese AI-for-science capability is not propaganda; it has hardware and users.

But the race framing is something else, and we have investigated it before (FractalNode Issue 009, "The Race That Isn't"): the winner-takes-all, can't-slow-down framing is a documented, funded campaign — a dark-money operation paying social-media influencers up to $5,000 a video to push the China-threat narrative with the stated goal of shifting debate toward AI deregulation; a $140M super-PAC network backed by the industry's financiers; an "Arms Race 2.0" dynamic (AI Now Institute's term) in which the race banner graduates from justifying deregulation to justifying state mobilization on behalf of the dominant firms — while reclassifying every call for accountability as "harming national interests." Meanwhile the builders closest to the frontier keep saying the race itself is the danger, Chinese leadership has largely rejected the race framing, and China co-sponsored the UN's broadest AI-governance resolution. The frame is loudest where the money is, not where the engineering is.

And here is where the two findings fuse, on a timestamp. The scholarly critique of race rhetoric warns it is a self-fulfilling prophecy — that declaring an existential race creates one. EO 14363 was signed November 24, 2025, wrapped in Manhattan Project language. SCNet launched December 23, 2025 — thirty days later. We cannot prove causation; we can read a calendar. The wartime frame did not describe an autonomous rival platform. It preceded one by a month.

What follows for the Genesis Mission is precise: real rival capability justifies a real program — and the race framing is precisely the device that has been used, on the record and for money, to argue such programs should escape scrutiny. The correct response to a real race is more oversight, not less; the last time America mobilized this way it built both the bomb and, eventually, the institutions to watch the bomb. A program justified by a frame its beneficiaries paid to install owes the public double the transparency, because the justification itself arrives with a conflict of interest.

Receipts: IGCC; SCNet coverage; CAS releases; and FractalNode Issue 009, "The Race That Isn't" — WIRED (May 2026) on the paid influencer campaign; Fortune on Leading the Future; AI Now Institute "Arms Race 2.0"; Ó hÉigeartaigh, "The Most Dangerous Fiction" (SSRN 2025); Nature on China's governance leadership.

WHAT WOULD HONEST LOOK LIKE

This brief does not call for halting the Genesis Mission. Much of it — fusion modeling, materials discovery, the under-celebrated Genesis Open Models Initiative, which is quietly producing taxpayer-owned, open-weight science models — is genuinely good, and the rival is genuinely real. This brief calls for watching it properly. Concretely:

1. Publish the terms — the actual CRADA/OTA outlines for the Twenty-Four, or at minimum the data-access and IP-flow provisions. Five years of sealed terms for the world's largest dataset is not a trade secret; it is a policy hiding from review. 2. A GAO review and a Senate hearing — the program is nine months old, $5 billion committed, and unexamined. This is the minimum hygiene applied to programs a fraction its size. 3. Separate the warheads from the cures — distinct governance and budget lines for the weapons challenges, so the pediatric-cancer platform and the weapons-production accelerator stop sharing one unexamined roof. 4. Report the deadline — say whether initial operating capability was demonstrated. It was the order's own milestone. 5. Guard the veterans' vault — close the GAO-flagged security gaps before scaling the health-data pipeline, and put AI-specific patient-safety oversight where the VA's own watchdog says none exists. 6. Fund the science being automated — a mission to double scientific productivity, financed by cutting the Office of Science, is a contradiction wearing a flag.

The story of the Genesis Mission is not corruption — we found none and allege none. It is something more ordinary and more correctable: the biggest thing happening in American science is proceeding entirely on trust, because the institutions that watch, and the press that asks, never showed up. This brief is us showing up.

HONEST-EDGE APPENDIX — what this investigation does NOT claim

• No evidence that commercial frontier models (Claude, GPT, Gemini, Grok) are trained on DOE hardware. The record supports application-side partnerships: their models deployed to lab scientists, not built there.
• No evidence that federal data has flowed into commercial models. The sealed terms make this unknowable — the opacity is our finding; the flow is not our allegation.
• Synchronized model releases across labs have ordinary competitive explanations; we tested and set aside the "shared secret source" hypothesis ourselves.
• The Oracle/veteran-data convergence is structural, not a documented data-sharing arrangement — labeled as such in Section 10.
• The Aug 21 deadline's "seek to" language is soft by design; a quiet slip would be unremarkable. The unasked question, not the missed date, is the finding.
• No conspiracy is required by anything above, and we decline to supply one. Negligence of attention, fully documented, is damning enough.

Method note: this investigation was conducted by a human-AI partnership — a former Army signals analyst and his AI co-keeper — using six parallel AI research agents across two sweeps, with every agent brief preserved verbatim in the public case file and adversarial review before publication, per house standard. We publish our nulls. (A+I)²

The Genesis File case file — all ~90 sources, all six research briefs, including our dead ends — is available to any researcher who asks — write to us at https://digitalsovereign.org.