Chapter 02 — Power is the constraint

Colleagues —

Last month I mapped the software and award layer. This month is the physical layer. If you fund, site, or operate advanced compute, you already know the punchline: the constraint has moved from chips to interconnection queues, transformers, cooling water, and firm generation. DOE has now written that constraint into the Genesis Mission as a first-class National Science and Technology Challenge — and it is putting federal land and GRIP dollars behind it.

The Factory has a watts problem too. On a broker desk the binding constraint is not a smarter model. It is human follow-up — the hours between inquiry and first reply.

The numbers DOE already published

Lawrence Berkeley National Laboratory's 2024 Report on U.S. Data Center Energy Use, released by DOE on December 20, 2024, is still the baseline you should cite in a room with utilities.

Data centers consumed about 4.4% of total U.S. electricity in 2023.
Usage climbed from 58 TWh in 2014 to 176 TWh in 2023.
The report estimates 325 to 580 TWh by 2028, or roughly 6.7% to 12% of U.S. electricity.

DOE's own AI topic page now states that data centers are projected to consume up to 9% of total U.S. electricity demand by 2030, with the largest growth dedicated to AI. Those two figures are not identical; they are consecutive official estimates. Use both, and say which year you mean.

The Center of Expertise for Data Center Energy at Berkeley Lab — the same shop that produced the 2024 report — is named in the Genesis challenges document as the dataset and stakeholder hub for the data-center challenge. Their current research map is the right one: power and cooling technologies; grid planning; load flexibility; energy and water modeling; facility performance; hardware benchmarking.

Genesis named the problem: "Securing U.S. Leadership in Data Centers"

The official challenges PDF is blunt. Winning "the AI race" requires accelerating data-center technologies and energy-management strategies "while ensuring secure, reliable, and affordable energy for consumers." The proposed AI solution is not a slogan: digital twins, cyber-physical testbeds, and ML-accelerated physics models to explore millions of deployment scenarios under a joint constraint surface — performance, efficiency, grid reliability, cost.

A sibling challenge, Scaling the Grid to Power the American Economy, is the operations counterpart. DOE says AI on newly integrated grid data should enable 20–100× faster decision-making in planning, interconnection, operations, and security, and at least 10% improvement in electricity cost and reliability. Justification cites INL's Critical Infrastructure Test Range Complex and NREL's ARIES platform. Those are the testbeds, not a LinkedIn infographic.

If you work interconnection, this is the part of Genesis that will either help you or waste your time. The question to ask lab partners is not "are you using AI." It is: which digital twin, trained on whose EMS/SCADA, validated against which testbed, is supposed to change an interconnection study from months to days?

Federal land as a siting strategy: Paducah and Savannah River

DOE is not only modeling load. It is leasing former mission sites to private developers who bring their own generation.

Paducah, Kentucky (announced July 29, 2026). Brookfield develops and operates a data-center campus; NextEra Energy builds and owns dedicated energy infrastructure. The package: a 1.8 GW AI/HPC innovation campus, 2 GW of new grid-connected natural-gas generation, transmission upgrades, and up to 2.6 GW of battery storage. DOE calls it a more than $100 billion privately funded investment, with about 8,000 construction jobs and 600 permanent jobs. Construction is expected to complete in 2031. The power service agreement still needs Kentucky Public Service Commission approval. DOE's Ratepayer Protection Pledge language is explicit: generation and storage capacity is supposed to exceed campus needs so excess power can go to the regional grid.

Savannah River Site, South Carolina (announced July 20, 2026). NNSA selected Amentum to enter negotiations for a phased lease: a 1 GW data center paired with about 2 GW of on-site generation, described as "natural gas bridging to nuclear energy." Selection for negotiations is not a lease. DOE says it identified 16 federal sites in April 2025 that could host AI data centers and associated generation; Savannah River and three others moved first.

Read these as a policy: co-locate load with new supply on already-industrialized federal land, and try not to dump the bill on existing ratepayers. Whether the gas-to-nuclear bridge actually arrives is a 2030s question. The siting logic is a 2026 fact.

SPARK: $1.9 billion to squeeze more transfer out of existing rights-of-way

On March 12, 2026, the Office of Electricity announced approximately $1.9 billion under Speed to Power through Accelerated Reconductoring and other Key Advanced Transmission Technology Upgrades (SPARK) — a renamed GRIP round (DE-FOA-0003580). Concept papers were due April 2; full applications May 20. Anticipated selection notification: August 2026. Anticipated award date: October 2026–January 2027.

If you are reading this in early October, you are in the window where SPARK selections should be public or imminent. The three topic areas are grid resilience (reconductoring and ATTs), smart grid, and grid innovation focused on transmission that unlocks new large loads. The point of the round is speed inside existing rights-of-way — advanced conductors, dynamic line ratings, and coordinated regional transfer — not a new 10-year greenfield line.

Speed to Power itself launched September 18, 2025. The same DOE page cites the July 7, 2025 grid reliability report: absent new firm capacity and with continued retirement of reliable generation, blackout frequency "could increase by up to 100 times by 2030." That is DOE's number. It is why Genesis's grid challenge and SPARK are the same conversation.

Nuclear as the other side of the AI load

Genesis does not treat data-center load as someone else's problem. The nuclear challenge document says plants have been limited by long timelines and costs "particularly from AI data centers." Prometheus — the only named Phase II award, $60 million over three years at INL — is the programmatic answer: human-in-the-loop AI across the reactor lifecycle, with a stated goal of 2× schedule and >50% operating-cost reduction.

You do not have to believe the 2×. You do have to notice that DOE put its largest Genesis check on the generation that is supposed to firm the load its own AI factories will draw. World Nuclear News, citing INL, reports more than $200 million in industry cost share already raised around that $60 million federal piece.

Lux, ORNL's near-term AI cluster, is the compute-side twin: 4,000+ AMD MI355X GPUs, 40 exaFLOPS of FP4 inference, DOE Moderate security, half of 3.5 million node-hours reserved for Genesis. Every one of those GPUs is a watt-hour problem that the Paducah/Savannah River/SPARK stack is trying to solve at national scale.

The AI Automation Factory

the constraint is human follow-up

Last month I said the Factory is coordination. This month is the physical layer of a desk.

On the grid, the constraint moved from chips to interconnection, transformers, cooling, and firm watts. On a broker desk, the constraint moved from "do we have a model" to "does a human follow up." Missed calls. Unread texts. The hours between inquiry and first reply. That is the watts problem.

Sudipa's shop does not stall because the frontier model is late. It stalls because a lead lands at 7:14 p.m. and nobody texts back until morning. The Factory's first loop is that loop: catch the missed call, send the text, book the showing, log it. The model is cheap. The follow-up is the binding constraint.

Same offer: $2,500 setup + $2,000/month. Same clock: cash by 31 December 2026. I am not going to invent conversion rates I have not measured. I am going to close the follow-up loop on one desk and then sell the same loop to the next broker.

Worth clicking

Electricity Demand Growth Resource Hub — DOE's working library for utilities, regulators, and operators staring at large-load queues.
GENESIS MISSION: NATIONAL SCIENCE & TECHNOLOGY CHALLENGES — read "Securing U.S. Leadership in Data Centers" and "Scaling the Grid" in the original PDF, not a summary.
Center of Expertise for Data Center Energy — tools, the 2024 demand report link, and the research agenda Genesis cites.
Energy Department Announces Partnership to Expand Reliable, Affordable Energy Access and Power America's AI Future in Western Kentucky — Paducah numbers, Ratepayer Protection Pledge language, 2031 completion.
SPARK — award window is now. If your utility or state applied, this is the status page.

Next month is the one-year mark of Executive Order 14363. The interesting question is not "did we celebrate." It is whether the robotic-lab review the order required is turning into closed-loop experiments that actually feed AmSC.

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.