The AI boom is a delivery problem

For a decade, the story about AI was silicon, models, and money. In 2025 and 2026, that story flipped. The people racing to train and serve models still need chips. They still have capital. What they cannot buy on a quarterly schedule is a studied interconnection, a free bay at a substation, a crew that can pour a foundation next month, or a gas turbine that arrives when the lease starts.

The thesis is simple: the AI buildout is now a delivery problem. Whoever can separate real load from phantom load, clear interconnection, and energize firm megawatts on hyperscaler timelines captures the physical margin of this boom. Everyone else waits in a queue that looks like demand and behaves like speculation.

That is not a slogan. It is what the public record now shows.

Capex moved upstream of the grid

Nat Bullard’s 2026 annual presentation, and the Machine Readable note that introduced it, put the scale in historical company. Tech capex—mostly, but not entirely, AI-driven—reached just under 2% of U.S. GDP last year, larger than the peak of the late-1990s broadband buildout and far larger than the interstate highways or Apollo at their peaks. Bullard is careful about what that number includes: the compute, the buildings that hold it, and the power gear inside the fence. It does not include the transmission, generation, and water systems that have to catch up.

On Catalyst with Shayle Kann, Bullard made the same point in plainer language: the prime mover of new U.S. infrastructure demand is energizing data centers, and the utility-side capital that rides with it is measured in tens to hundreds of billions.

Capital found the compute. Delivery has to find the electrons and the wire.

The pipeline looks like another country

How big is the asked-for load? Latitude Media’s reporting from its Transition-AI 2026 conference put one number on the table that refuses to leave the room: about 780 GW of announced U.S. data-center projects, against a national peak load near 759 GW. Aurora Energy Research’s Oliver Kerr called it “an additional U.S. of power by 2030.”

Operating reality is smaller, and still growing fast. FERC’s 2025 State of the Markets report, as covered by Utility Dive, put more than 50 GW of data-center capacity in service at the end of 2025—about 24% compound annual growth since 2020—with average facility size rising from roughly 25 MW in 2020 to nearly 80 MW for units that came online last year. Facilities that large force new generation or transmission into the study. Planners cannot treat them like ordinary commercial load.

The gap between announcements and steel is the story. It is also the opportunity for builders who can close it.

Queues are full. Certainty is scarce.

Interconnection queues now carry the speculative habits that generation queues have always had, and the industry is still learning how to discount demand-side requests the way it discounts wind and solar options.

Utility Dive has tracked the response across 2025 and 2026: large-load tariffs with deposits, minimum demand, collateral, exit fees, and ramp schedules; state rules that force site control and curtailment commitments; cluster studies instead of endless sequential reviews. Texas is the clearest stress test. The same reporting describes a large-load queue on the order of hundreds of gigawatts—several times recent peak demand—paired with an interconnection pause and audit while operators try to tell real projects from tire-kickers. BloombergNEF’s public estimate, cited in that coverage, is that the pause alone could delay roughly a fifth of the U.S. data-center pipeline if it stretches.

Bullard and Kann walked the same ground on Catalyst: ERCOT’s large-load requests exploded from the low tens of gigawatts in early 2024 into the low hundreds by late 2025, while grid-operator and transmission-provider forecasts diverged by hundreds of terawatt-hours. Different incentives produce different “demand.” As Bullard put it in the episode framing, this is a “no one knows anything” slide for a reason.

For operators and founders, the practical reading is not “the boom is fake.” It is that certainty became the scarce product. EPC firms told Latitude’s conference audience they can now choose among customers; one contractor estimated that maybe one in four queue projects will actually move. Hyperscaler cash still punches through. Speculative colo that cannot lock offtake does not.

The physical stack is sold out

Delivery fails for boring reasons: turbines, transformers, crews, and study bandwidth.

On Catalyst, Bullard described a gas-turbine order book running tens of gigawatts above annual production capacity, with aero-derivative and jet-engine makers pulled into base-load duty they were not originally designed for. Halcyon’s public write-up of the annual deck notes gas-plant costs rising across a tracked set of U.S. projects under development, with no sign of a near-term reset. Latitude’s conference reporting added the construction side: labor and materials are scarce; break-ground timelines that used to run about six months are stretching toward eighteen; EPCs prioritize counterparties who can prove the load is real.

None of that is a software problem. It is a schedule problem. The companies that win treat schedule as the product.

Capital is already pricing the wire

The money is not only in racks. BloombergNEF’s public press summary of Energy Transition Investment Trends 2026 puts global energy-transition investment at a record $2.3 trillion in 2025, with power grids alone at $483 billion. BNEF also estimated data-center investment around half a trillion dollars in 2025—ahead of solar, behind electrified transport—and flagged data-center buildouts as a continued driver of clean-power and buildings M&A. (Figures above are from BNEF’s public press release and free summary materials only; we do not reproduce subscriber research.)

New Project Media sells deeper pipeline, interconnection, and financing coverage across power, renewables, storage, and data centers. We cite only its public pages—starting with NPM Free Intel—and never subscriber copy. Even at the public layer, the product map is the point: the market now prices project status, queue position, and permit path as tradable intelligence. That is what a delivery boom looks like from the research desk.

What this means for people building companies

If you run plants, sites, fleets, or crews—or you are founding a company that will—three implications follow from the sources above.

1. Treat load as an operating system, not a customer. The winners in 2026 are not the ones with the loudest load forecast. They are the ones who can prove site control, study readiness, offtake, and a ramp schedule that utilities and ISOs will underwrite. Large-load tariffs and cluster studies are turning “we might build here” into a product with fees and gates. Learn those gates before you sell past them.

2. Own a scarce step in the path from request to energization. Generation, interconnection studies, substations, transmission rights-of-way, turbine and transformer lead times, EPC capacity, and commissioning talent are all bottlenecks. A company that shortens any one of those steps—honestly, with primary data—has pricing power. A company that only adds another speculative request to the queue does not.

3. Put AI inside the work that delivery actually is. Estimating, procurement, project controls, outage planning, and field ops are where schedule slip dies or lives. The same AI wave that is creating the load is also the tool set that can compress the paperwork and the change orders around it. Use AI there. Do not confuse a chatbot demo with a plant that can start on the day the lease starts.

The builders who move first on those three points set the terms for everyone else. The ones who wait will spend the next decade buying access from the people who did not.

What Driftwood Club is watching

We will keep this thesis honest against the public record:

• Queue reforms and large-load tariffs, as Utility Dive and the primary dockets keep filing them.
• Bullard’s annual deck and Machine Readable for the macro frame, and Catalyst for the conversations that stress-test it.
• Public BNEF and NPM material only, clearly labeled, never paywalled text.
• Primary agency and ISO releases that settle arguments the trade press can only preview.

If you are building companies in energy, power, grid, construction, or infrastructure—and using AI to build them better—this is the conversation the room is for. Join Driftwood Club. Tell us what you run or what you are building. We introduce members working on the same problems, one at a time, and we send the sourced reading each week.

The boom is real. The pipeline is noisy. Delivery is the work.