Mission and its AI framing predate Genesis; its public requirements remain centered on stockpile simulation and application performance.
On July 16, the National Nuclear Security Administration presented two forthcoming Los Alamos National Laboratory supercomputers, Mission and Vision, under the Trump Administration’s Genesis Mission. Both will be built by Hewlett Packard Enterprise on the NVIDIA Vera Rubin platform. NNSA said the systems’ codesign could “reduce the time needed for discovery from months to minutes.”
The concrete near-term disclosure was narrower: the first Vera CPU server is scheduled to arrive at Los Alamos this summer for testing. Full deployment of Mission and Vision is expected across 2027 and 2028.
The systems themselves were not new announcements.
NNSA, LANL, and HPE had already announced Mission and Vision on October 28, 2025. At that point, HPE had been selected as prime contractor; the HPE Cray Supercomputing GX5000, NVIDIA Vera Rubin processors, and Quantum-X800 InfiniBand network had been identified; and both systems were being presented as products of the AI era. The economics of the Vera Rubin platform SCN has traced elsewhere, in its capital-cost curve and the Korean memory stack underneath it. NNSA specifically assigned foundation-model and agentic-AI work to Vision, while LANL described Mission as combining HPC and AI capabilities for classified modeling and simulation.
Executive Order 14363 launched the Genesis Mission on November 24—27 days after those announcements. The order directed DOE to construct an integrated AI platform using national-laboratory supercomputers, scientific data, models, agents, and experimental facilities. It did not appropriate money, and its implementation remained subject to available appropriations.
The public record therefore supports a narrower—and more defensible—conclusion than the idea that a simulation machine received an AI paint job to match a new presidential initiative.
Genesis did not originate Mission, its procurement, or the AI positioning around Mission and Vision. It later became the institutional and political umbrella under which NNSA presented them.
For Mission specifically, however, the public engineering record remains simulation-led.
The systems discussed in the recent announcements are related, but they are not interchangeable.
The ATS-5 evidence characterizes Mission. It should not be applied wholesale to Vision or Veritas.
LANL released the 61-page ATS-5 technical-requirements document on September 3, 2024. It calls for a system to be deployed in 2027 to replace Crossroads as a tri-lab computing resource for the weapons program. Its stated mission is to run some of NNSA’s largest three-dimensional stockpile simulations and reduce their completion time from months to days.
The document’s four architectural goals are:
It singles out memory bandwidth, sparse and semi-sparse memory access, strong scaling, and branch-heavy code as areas of particular value. It also defines “time to insight” as the wall-clock time required to complete NNSA’s largest and most complex stockpile simulations.
That emphasis is consistent with irregular, memory-sensitive simulation workloads. It is not a specification built around a single accelerator-throughput number or a dense matrix-multiplication benchmark.
LANL explicitly states that Linpack performance is not the measure of success. Instead, proposals and final system acceptance are tied to a seven-application suite:
The same benchmark suite is used in procurement evaluation and acceptance. Offerors were asked to describe systems capable of at least a tenfold scaled single-node improvement, or SSNI, over Crossroads. A tenfold scalable system improvement, or SSI, target was to be negotiated before award.
AI is present in the requirement and should not be edited out of the history.
MLMD trains and runs inference on machine-learned interatomic-potential models. The software requirements also ask vendors to describe optimized libraries for machine-learning and AI workloads and support for distributed deep-learning frameworks such as PyTorch DDP and Horovod.
The requirement hierarchy puts those capabilities in context. TR-1 requirements form the proposed baseline system. The optimized AI libraries are listed at TR-2, which adds capabilities or productivity, while distributed deep learning is listed at TR-3, the document’s stretch-goal tier.
AI is therefore neither absent nor newly attached. It is a supported workload, part of the software environment, and one component of the benchmark suite.
It is not the procurement’s stated mission need. That remains stockpile simulation, with application performance and memory-system behavior determining success.
The distinction matters because the 2024 and 2026 performance claims do not have the same referent.
ATS-5’s months-to-days objective applies to identified hero-class stockpile simulations. Its definition of time-to-insight is tied to the wall time of those simulations, and its acceptance framework uses SSNI and SSI results from published application benchmarks.
NNSA’s July months-to-minutes statement applies broadly to “discovery.” It does not identify a model, simulation, agent workflow, baseline system, dataset, level of numerical precision, or acceptance method.
LANL’s companion account is more qualified. It says agentic AI can reduce time-to-insight “in some cases” from months to minutes by allowing AI systems to form hypotheses, choose tools, launch simulations, analyze results, and refine subsequent steps. That is a claim about workflow acceleration, not necessarily a claim that a hero-class physics simulation will execute in minutes.
Those could both be useful advances. They are not technically equivalent.
SCN found no public ATS-5 acceptance criterion corresponding to the months-to-minutes statement. Until NNSA or LANL identifies a workload, baseline, quality threshold, and measurement procedure, the figure should be treated as a strategic performance claim rather than an engineering acceptance target.
The July announcement was primarily a reframing at the NNSA level, but the broader set of June and July disclosures was not purely rhetorical.
NVIDIA said on June 22 that Mission’s planned configuration will combine Vera Rubin GPU nodes with 2,300 standalone Vera CPUs in HPE GX240 blades. The same account described Veritas and its planned approximately 1,150 standalone Vera CPUs.
NVIDIA also published early LANL testing results. It said a Vera CPU exceeded the Crossroads CPUs by more than threefold on Branson, one of the seven ATS-5 application benchmarks. The post gives no compiler settings, CPU operating conditions, problem configuration, node-level normalization, power envelope, or other methodology sufficient for an independent comparison. The number is relevant because Branson is part of the procurement suite, but it remains a vendor-published result sourced to the customer.
LANL subsequently put broader design targets on the record. The laboratory says Mission and Vision are designed to provide more than three times the per-CPU performance of Crossroads, more than four times the memory per core, and lower power consumption. It also says each Rubin GPU is designed to deliver more than 12 times the AI performance of the Hopper GPUs in Venado.
The CPU target maps clearly onto Mission’s published concern with per-core memory bandwidth and irregular applications. The 12-times AI figure is less reproducible from the release: LANL does not identify the numerical precision, model, kernel, batch size, sparsity assumptions, or power basis behind it.
LANL’s account also identifies Starlight, due later in 2026, as the first of three purpose-built architectures for Mission. The full Mission and Vision systems are then expected to use Vera CPUs and Vera Rubin NVL4 rack-scale systems. The release does not disclose Mission’s total node count, overall CPU-to-GPU balance, or the other two purpose-built architectures.
This is why “the machine did not change” goes beyond the evidence. The public record shows continuity in the selected prime contractor and named core platform. It also shows that the disclosed design became materially more detailed. The documents do not establish whether every later detail was present in October, matured during codesign, or changed afterward.
DOE published its FY2027 Weapons Activities budget justification on April 3, 2026. It is a 375-page budget request to Congress, not an enacted FY2027 appropriation.
The capital-equipment table lists “ATS-5 System, LANL” with a $250 million total estimated cost. It shows:
The table therefore shows $130 million through FY2026. “Enacted” should not be conflated with money already spent, paid to HPE, or obligated under a single contract.
The budget narrative also uses the combined name “ATS-5/Mission,” confirming that the budget line and the publicly named supercomputer refer to the same system. An exact-text search of the document found no occurrence of “Genesis.” It separately discusses AI tools and infrastructure across the nuclear-security enterprise.
The sources reviewed by SCN do not disclose the value of HPE’s award for Mission and Vision.
HPE says the systems are part of a $370 million DOE investment. That statement does not say Mission and Vision themselves cost $370 million, and HPE does not break out the scope included in that figure.
NNSA’s October announcement cites $115 million provided under Public Law 119-21 to accelerate national-security work through AI. It does not assign the entire $115 million to Mission or Vision and does not provide a system-specific price.
NNSA’s classified-AI plans also predate Genesis. The FY2026 Weapons Activities justification called for deploying open-source and proprietary large language models on unclassified and classified networks and integrating AI software into stockpile-design tools and workflows. It also proposed an Artificial Intelligence for Nuclear Security program focused on secure models, AI-assisted design and inspection, materials discovery, and classified AI hardware.
Genesis did not initiate NNSA’s movement toward AI. It consolidated and elevated work that was already underway.
Set the October and July NNSA announcements side by side, and the most important change is the institutional frame.
October established:
July foregrounded:
LANL’s companion account restores the procurement chronology that NNSA’s July release omits. It opens by noting that the laboratory unveiled Mission and Vision the previous fall, then adds Veritas, Starlight, and the new design targets. In the version reproduced by HPCwire as a LANL release, Genesis is not invoked.
These are two institutional accounts of the same systems. The laboratory account is anchored to procurement history, codesign, architecture, and delivery. The agency account is anchored to Genesis and its promise of AI-accelerated discovery.
Neither account necessarily invalidates the other. Genesis was expressly designed to incorporate existing DOE computing resources into a unified platform. Attaching Mission and Vision to that initiative after their selection is therefore programmatically coherent. It is not the same as saying Genesis caused their procurement.
One schedule discrepancy remains unresolved. LANL’s October release and NVIDIA’s June account expected both Mission and Vision to be operational in 2027. The July releases describe full deployment across 2027 and 2028. The public documents do not establish whether that reflects a schedule slip or a distinction between initial operation and full deployment.
The public record establishes continuity in procurement, prime contractor, and core platform. It does not establish a frozen design.
Mission predates Genesis. The AI positioning around Mission and Vision does too. July did not create either one; it connected them to a later national initiative and added limited delivery and processor detail.
The distinction matters because program branding and system acceptance are different things. SCN reported earlier this week that simulation supercomputers are increasingly pitched in AI terms; that piece could report the pressure only as practitioner complaint. Here it can be read off government paper.
Mission’s published engineering standard remains performance against seven application benchmarks, using Crossroads as the reference and SSNI and SSI as the acceptance framework. “Discovery in minutes” is not among the public criteria.
Until NNSA publishes a workload and a reproducible measurement behind that phrase, months to minutes is a vision statement (no, not the supercomputer.)
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.
