Chapter 05 — Frontier, Lux, Discovery

Colleagues —

A new year is a cheap excuse to restate architecture. I am using it anyway, because the Genesis software story (AmSC, ModCon, agents) is easy to over-index and the machine room is easy to treat as background. It is not background. It is the budget.

The Factory has a machine room too. A VA does not live in a prompt. It lives on phone, SMS, calendar, CRM, a human checkpoint, and a runbook. Software only matters if those boxes stay full.

Frontier is still the exascale floor

The Oak Ridge Leadership Computing Facility's Frontier remains the system of record for "we already did exascale." OLCF's own page: first to 1 exaFLOP in FP64 (2022), HPE Cray EX architecture, Slingshot, 3rd Gen AMD EPYC plus AMD Instinct 250X, peak 1.6 EF. The science case on that page is the Genesis case in older language — whole-device fusion models, materials at new timescales, cosmology with AI-oriented analysis, SMR neutronics, LCLS image reconstruction in minutes rather than weeks.

Lux allocations can share data with Frontier via Orion, OLCF's Lustre file system, over 600 PB. That sentence is the 2027 workflow: simulate or generate on Frontier, train or infer on Lux, without a truck.

DOE Moderate security controls on Lux match Frontier. OLCF says it plans Moderate-Enhanced, which would add ITAR and PHI. If your Genesis workflow touches either, that enhancement is the gating item to track — I will not claim it has shipped until OLCF says so.

Lux: the 2026 AI injection, with a commercial half

DOE announced Lux and Discovery on October 27, 2025. The public-private model is the news, not the press-conference adjectives. AMD co-invests. OLCF operates. Half of Lux's annual 3.5 million node-hours are dedicated to the Genesis Mission. The other half is paid proprietary access at the DOE user-facility rate. Proprietary projects can start any month and run six months to five years. Minimum buy: 175,000 node-hours per six-month commitment — OLCF's example is 400-node (3,200 GPU) jobs for eighteen days, or ~40 nodes continuous for six months. Contact is olcf-partnerships@ornl.gov.

Published specs, from OLCF:


GPUs	4,000+ AMD Instinct MI355X in 500+ nodes
HBM	1.1 PB total; 288 GB per GPU, 8 TB/s bandwidth
AI inference	40 exaFLOPS (FP4)
AI training	20 exaFLOPS (FP8)
HPC	317 petaFLOPS (FP64)
Node local	>24 TB NVMe per node
Scheduling	Slurm and Kubernetes
Network	Ultra Ethernet, 400G NIC per GPU

ORNL's July 22 Genesis release said installation was underway and Lux "will be completed this year." If you are reading this in January 2027, the operational question is not the announcement. It is: who got the Genesis half, what ran, and is the proprietary half actually selling. I will not invent those answers. Ask OLCF.

Discovery: 2028, and the HPC/AI/quantum convergence bet

Discovery is the traditional procurement. HPE Cray Supercomputing GX5000, next-gen AMD EPYC "Venice" CPUs, AMD Instinct MI430X GPUs, DAOS-based HPE Cray Storage Systems K3000. Delivery: 2028. DOE's October 2025 release said performance "far exceeding Frontier," and that Discovery will "pioneer the convergence of high-performance computing, artificial intelligence, and quantum systems." ORNL's March 2026 update repeats: significantly greater performance than Frontier across every part of the system; modeling and simulation integrated with AI; path to HPC + AI + quantum.

Together with Lux, DOE cited more than $1 billion in public-private investment with AMD and HPE.

If you are planning 2027–2028 science, Discovery is a capacity and architecture event, not a branding event. The software question is whether AmSC's identity and data APIs will still be the way a PI in 2028 submits a job that spans Lux leftover, Discovery, and a beamline — or whether we are back to ssh and tribal knowledge.

AmSC + ESnet: the system of systems

Leadership machines without a network are museums. ESnet is DOE's high-performance networking user facility — the "data circulatory system" for the 17 labs and 28 user facilities. Inder Monga's AmSC description is still the right one going into 2027: common identity, interoperable service interfaces, AI-enabled orchestration, intelligent storage, plus an extension to commercial cloud and frontier AI models.

PNNL, co-lead of the Models pillar, describes AmSC as unified APIs, secure enclaves, and hybrid commercial-cloud integration — the backbone ModCon rides. That is the runtime for agent cards I wrote about in December.

Office of Science still operates 28 user facilities as shared, peer-reviewed resources used by tens of thousands of researchers a year. Supercomputers, accelerators, light sources, neutron sources, nanoscience, genomics. Genesis only "integrates experimental facilities" if those 28 actually show up as AmSC services. SYNAPS-I and MOAT (accelerators/synchrotrons) and OPAL (bio design labs) are the prototypes. Watch whether 2027 user-facility proposal calls start requiring or offering AmSC-native data paths. I will not claim they already do.

The grid side of the machine room

SPARK's anticipated award window was October 2026–January 2027. If selections posted, they belong in your 2027 infrastructure map next to Lux, not in a separate "energy policy" folder. The round is $1.9 billion for reconductoring and advanced transmission technologies, including a topic explicitly aimed at transmission that unlocks new large loads.

DOE's July 2025 reliability report, cited on the Speed to Power page, warned that without new firm capacity, blackout frequency could increase by up to 100 times by 2030. Genesis compute that cannot stay up is not a science platform.

The AI Automation Factory

the stack the VA actually runs on

A new year is a cheap excuse to restate architecture. Genesis software (AmSC, ModCon, agents) is easy to over-index; the machine room is the budget. Same split in the Factory.

The VA does not live in a prompt. It lives on a stack: phone / SMS, calendar, CRM or sheet, email, a human checkpoint, a runbook. Software only matters if those boxes stay full — the same sentence I am using for Frontier, Lux, and Discovery.

January 2027 is after the cash date. Sudipa's desk is still the first bench: if that loop ran, the question is not the announcement. It is which loop ran, did the proprietary half (the $2,000/month retainer) actually sell, and is the stack boring enough that a second broker can attach without a custom build. If the first client did not pay, I will say so rather than invent a 2027 customer I cannot verify.

Offer stays: $2,500 setup + $2,000/month. 2027 lead machine: this newsletter + one weekly ask. No spray. The stack is what makes a second desk possible without starting from tribal knowledge.

Worth clicking

Lux: ORNL's AI + HPC Supercomputer — specs, Genesis split, proprietary path, Orion.
Frontier — still the FP64 floor.
ORNL, AMD and HPE to deliver DOE's newest AI supercomputers: Discovery and Lux — Discovery architecture, 2028 delivery, GX5000 / MI430X / Venice.
The American Science and Security Platform — the legal/program name for the integrated instrument.
Office of Science User Facilities — the experimental half of the same instrument.

February: how outsiders attach. The consortium, the 24 named MOU shops, SBIR, user-facility access, and the partnership inbox. Most of this stack is useless if only the 278 can touch it.

The AI Automation Factory. Virtual assistants for sales and broker desks. First workflow: a real-estate broker/operator (Sudipa). Offer: $2,500 setup + $2,000/month. First paid client / cash by 31 Dec 2026. 2027: this newsletter + one weekly ask. No spray.

Reply if one loop on your desk is eating the week. deep-heap.com

Deep Datta
Deep Heap

Intelligence has a substrate.

I write this for people who already know the labs, the machines, and the constraints — and for operators who need a virtual assistant that actually runs a desk. If I cannot verify a source, I leave it out.