Energy · Annual review · 2026-08-30

The Megawatt Became the Product

The power trade moved from grid avoidance to electrical equipment, then from order books to the harder question: which collection of assets can become an energized campus on time?

Dylan Austin Bristot, Founder and publisher, AI Bottlenecks

TL;DR

The power trade moved in four steps: escape the utility calendar, build the electrical room, turn forecasts into orders, then separate backlog from campuses that can actually switch on.
A retrospective test of today's 62-name Energy Bottlenecks universe returned 53.48%, against 43.01% for XLE and 19.37% for the S&P 500. A stricter synchronized buy-and-hold version returned 36.68%. This is a research test, not a live record.
Behind-the-meter power led with 84.82%, but finished 35.43% below its peak. The three-name transformer proxy returned 61.26% and finished only 3.84% below its peak.
The grid can be physically scarce without being the cleanest equity. The utility interconnection proxy returned only 10.57% because regulation, rates and cost recovery sit between a congested queue and shareholder rent.
The next useful denominator is not requested or contracted GW. It is energized MW, followed by utilization and useful compute output per MW.

The headline arrived late

On 29 August, Elon Musk wrote that the consensus estimate was for roughly 15 GW of AI compute produced in 2027 to remain unable to switch on that year. He listed transformers, wiring, liquid cooling, chillers and networking beside the missing power. The 15 GW figure is an attributed estimate, not an audited forecast. The object is still useful: finished accelerators waiting on the campus around them. Source post

The demand is large enough to force the issue. The U.S. Department of Energy estimates that data centers consumed 176 TWh, or 4.4% of U.S. electricity, in 2023 and could reach 325 to 580 TWh by 2028. The IEA expects global data-center consumption to rise from roughly 415 TWh in 2024 to 945 TWh in 2030. DOE | IEA

The timing is much less clean. NERC raised projected North American summer peak-demand growth over the next decade to 224 GW, 69% above its prior assessment, while also noting that newer ERCOT and PJM forecasts were already trimming near-term load because projects had slipped or disappeared. Demand can be real and the announced delivery calendar can still be wrong. NERC 2025 assessment

That tension shaped the year. First the market paid for any credible route around the grid. By summer it was asking a harder question: which collection of equipment, permits and contracts could become live compute before the silicon aged?

What the tape said

The model uses today's 62-name universe, split across nine physical baskets and projected backward over the year. It equal-weights baskets daily, then names inside each basket. The universe did not exist at the start of the period, so the result carries look-ahead, survivorship and rebalancing effects.

| Market series | One-year return | | --- | ---: | | Energy Bottlenecks retrospective test | +53.48% | | XLE | +43.01% | | GRID | +26.44% | | PAVE | +19.93% | | S&P 500 | +19.37% | | XLI | +18.03% | | XLU | +4.18% |

Reasonable construction changes produced a range from 36.68% to 53.75%. The honest use of the tape is cross-sectional. It tells us what the market paid to own, not what caused every move.

The year in four turns

1. Escape the utility calendar

Behind-the-meter and campus power crossed +50% on 15 October, two days after Bloom and Brookfield announced a framework to finance up to $5 billion of Bloom-powered AI infrastructure. The framework did not guarantee Bloom revenue. It put a price on time: install generation at the site, then reduce dependence on a utility upgrade calendar. Bloom and Brookfield

Nuclear rallied at the same time, peaked, then rolled over. In hindsight the market initially bought almost every answer to the shortage before asking when each answer could arrive.

2. The electrical room gets a bid

By late February the trade had broadened into transformers, switchgear, busway, protection, backup power and cooling. Siemens Energy reported a 1.82 group book-to-bill and a record EUR146 billion backlog. Its Grid Technologies backlog later reached EUR51 billion. Siemens Energy Q1 | Siemens Energy Q3

DOE says U.S. demand for distribution transformers has risen 41% since 2019, while lead times moved from three to six months in 2019 to one to two years or longer in 2024. Large power transformers can take three to four years. More than 80,000 distribution-transformer variants sit across U.S. utilities, so even standardization was expected to improve cycle times by only 5% to 8%. This is not one shortage. It is material, design, factory-slot, test and qualification friction stacked together. DOE transformer working group | DOE large-transformer report

3. Forecasts become orders

The model gained 28.16% in twenty trading sessions from 30 March to 27 April. Almost on top of that move:

Bloom and Oracle expanded their agreement to as much as 2.8 GW, with 1.2 GW contracted and deploying. Source
GE Vernova booked $2.4 billion of data-center electrification equipment in one quarter, more than in all of 2025. Source
Powell disclosed a behind-the-meter data-center order above $400 million, then pushed backlog to $2.4 billion by the following quarter. Q2 | Q3
Vertiv reported 30% sales growth, while Flex announced plans to separate its Cloud and Power Infrastructure business. Vertiv | Flex

The shortage had moved from forecast decks into contracts, backlog, revenue and corporate structure.

4. The price paths separate

The retrospective model peaked at +82.40% on 22 June and finished at +53.48%.

!The energy baskets reached the finish line by very different paths

| Basket | Return | Drawdown from peak | | --- | ---: | ---: | | Behind-the-Meter & Campus Power | +84.82% | -35.43% | | Data Center Electrical & Thermal | +77.03% | -25.42% | | Transformers & Voltage Conversion | +61.26% | -3.84% | | Transmission & Bulk Power Delivery | +61.03% | -20.54% | | Specialized Electrical & Switchgear | +48.54% | -34.14% | | Grid Interconnection & Substations | +10.57% | -6.40% | | Nuclear & SMR Supply Chain | -6.64% | -31.92% |

The transformer proxy gave back the least. Behind-the-meter power delivered more return but also far more violence. The interconnection basket lagged despite owning a real physical problem. A regulated queue is not automatically a clean scarcity-rent mechanism.

Nuclear split by delivery clock, but imperfectly. Rolls-Royce, Cameco and Curtiss-Wright rose while NuScale and Oklo fell sharply. Yet BWXT, Constellation and Vistra also finished negative. Existing industrial capacity generally carried a nearer clock than pre-revenue projects, but power prices, defense exposure and valuation still mattered.

The product is a synchronized campus

A campus can have generation and still lack a substation. It can own a transformer slot and wait on switchgear. It can have every major box onsite and still miss cooling, protection settings or commissioning.

!The delivered-power stack runs from generation to field synchronization

The critical path behaves like a scheduled network:

Power source: utility supply, turbines, fuel cells, nuclear and renewables carry different fuel, permit and delivery clocks.
Grid connection: studies, substations, protection and cost allocation decide whether generation can reach the site.
Campus electrical: transformers, switchgear, UPS and busway condition and distribute the load.
Rack power: Nvidia's 800 VDC design is intended to support as much as 2 MW per row in 2027. Nvidia
Thermal rail: cold plates, coolant distribution and heat rejection must scale with rack density.
Synchronization: EPC, MEP, controls, firmware, field labor and commissioning sit across every layer.

> Time to energization = slowest dependency + synchronization penalty

Higher density makes the penalty worse because power conversion, protection, backup and cooling all change at once.

A gigawatt changes meaning six times

!Requested capacity is not revenue-bearing capacity

Requested: a developer asks one or several utilities for service.
Screened: the site has credible land, studies, deposits and timing.
Contracted: equipment, power and financing commitments become firmer.
Commissionable: major systems have compatible delivery and installation dates.
Energized: power reaches the rack under real reliability and curtailment limits.
Revenue-bearing: compute runs with enough uptime and utilization to produce cash.

ERCOT once reduced new data-center load requests to 49.8% of original requests in a capacity-planning adjustment. Berkeley Lab found only 13% of generation capacity entering U.S. queues from 2000 through 2019 had reached commercial operation by the end of 2024. These are different datasets and should not be multiplied into a fake funnel. Together, they show why a queue position is weak evidence of future supply. ERCOT | Berkeley Lab

The market can be right about a huge demand wave and wrong about which projects arrive, when they arrive and who gets paid.

Five calls for the next 12 to 24 months

1. Energized MW replaces contracted GW

Backlog still matters. The harder bridge will be equipment delivered, MW commissioned, MW energized and utilization after acceptance.

Watch: installation cycle time, customer acceptance, cancellations and active MW. Wrong if: backlog grows for four quarters while weak conversion and slipping dates carry no market penalty.

2. Integration captures more rent

The buyer no longer needs five scarce products. It needs five scarce products engineered and commissioned on one date. Prefabricated power rooms, controls and field service should gain value.

Watch: integrated power-block orders, commissioning revenue, schedule guarantees and change orders. Wrong if: component lead times normalize quickly and buyers return to fragmented procurement.

3. Transformers loosen late and unevenly

Hitachi, Eaton, Siemens Energy and GE Vernova are adding capacity, but much of the qualified output arrives from 2027 through 2029. Standard distribution equipment should ease before large custom units and high-voltage components.

Watch: quoted lead times, book-to-bill, pricing and factory qualification dates. Wrong if: multiple manufacturers report shorter lead times, falling orders and price concessions before new plants ramp.

4. Behind-the-meter becomes hybrid

Onsite generation can compress the utility calendar. Fully islanded campuses will remain unusual. The practical architecture combines grid service, onsite generation, storage and controllable compute.

Watch: firm fuel, islanding capability, storage attachment and real commissioning dates. Wrong if: permits, fuel or reliability delays erase the schedule advantage.

5. The bottleneck moves into DC power, protection and heat

800 VDC reduces conversion losses but creates qualification work in breakers, busways, connectors, backup, controls and cooling. Better compute per MW can make each energized MW more valuable without reducing total demand.

Watch: production shipments, rack-density mix, liquid-cooling attachment and failure rates. Wrong if: most deployments stay below the density where the new architecture matters.

The monitoring board

| Signal | Strong evidence | Warning | | --- | --- | --- | | Energized MW | Rising commissioning and utilization | Contracted GW grows while dates slip | | Backlog conversion | Revenue and cash follow orders | Repeated change orders or cancellations | | Transformer lead time | Qualified output expands gradually | Sudden price concessions across suppliers | | Integrated systems | Shorter deployment and higher service mix | Components normalize without integration rent | | Flexible load | Faster interconnection or lower upgrade cost | Every workload still priced as fully firm | | Rack architecture | 800 VDC and liquid cooling reach production | Roadmaps move faster than field acceptance |

The read

The past year was not simply a bet on electricity demand. It was a repricing of time.

The first winners promised power sooner. The next winners owned the electrical room. Then the market found the factory slots, field labor and cooling systems that decide whether those pieces arrive together. By summer, backlog alone was no longer enough.

The scarce product is not a turbine, transformer or grid connection in isolation. It is a synchronized campus with accepted power at the rack.

The metric that joins the whole report is:

> Useful compute output per energized MW, after downtime and financing

Everything before that is potential. That conversion from announced demand to live capacity is what Energy Bottlenecks is built to follow.

Methodology

Market figures cover 28 August 2025 through 28 August 2026 and project the report-date universe backward. Prices are Yahoo adjusted closes converted to U.S. dollars. The model equal-weights available baskets, then names within them, rebalanced daily. It excludes fees and factor adjustment and is descriptive, not a live performance record.