MACEDON STRATEGIC INTELLIGENCE

Strategic Intelligence Brief

The AI buildout is priced as a race for compute and capital. The binding constraint sits beneath it, in the power delivered to the site and the grid equipment that takes years to build.

01The announcement

The headline is capital. Through 2025 and into 2026, hyperscalers and the Stargate program announced data-center buildouts measured in tens of gigawatts, framed as a race to secure compute. The demand curve is real and steep: the Department of Energy, citing Lawrence Berkeley National Laboratory, projects U.S. data-center load to double or triple by 2028, and Goldman Sachs Research tracks data centers rising from about 4 percent of U.S. summer peak demand toward roughly 8 percent within two years. Read as an operating problem, the announcements do not answer the question that decides the timeline: where the power comes from, and how it reaches the building.

02The gap

The announced gigawatt is not a delivered gigawatt. Three layers the framing omits.

• Power must reach the site, and the grid is the slow step. The median time from interconnection request to commercial operation has passed five years for projects reaching operation in 2025, and in PJM, the region covering much of the data-center corridor, those projects took an average of about seven years from interconnection request to commercial operation. More than 2,000 gigawatts were actively seeking connection at the end of 2025, and of all capacity that requested interconnection from 2000 to 2020, only about 13 percent had reached operation by the end of 2025 while roughly three-quarters had been withdrawn.
• The equipment has its own backlog. Large power transformers now carry lead times of well over two years, with high-capacity units quoted around four and, in cases, five. Wood Mackenzie has modeled a shortfall of roughly 30 percent in power transformers across the national fleet. Projects executing before 2028 face that market, not the promised new capacity.
• The materials and trades run on multi-year clocks. Copper, grain-oriented electrical steel, generator step-up units, and the skilled trades that install them all carry lead times measured in years. The same constraint that gates one site gates every site bidding for the same equipment at once.

03The window

The gap does not close inside this decade by default. An interconnection that takes five to eight years, requested today, does not energize until the early 2030s, and a high-capacity transformer ordered today arrives in four to five years. Wood Mackenzie's own outlook has the power-transformer shortfall easing only gradually, from about 30 percent now toward roughly 5 percent by 2030, so the shortage is a feature of the whole back half of the decade. Demand is set to double or triple by 2028; the grid and the equipment that serve it are still working through a queue and a backlog that predate the AI wave. The years in between are where the risk sits.

04The exposure map

The exposure concentrates wherever an announced gigawatt is being treated as a delivered one. It sits with any plan or valuation that reads a data-center announcement as online capacity on the announced date; with developers whose timelines assume a grid connection they do not yet hold, against a queue where most capacity is ultimately withdrawn; in the equipment supply chain itself, where whoever can deliver a high-capacity transformer on a two-year rather than five-year horizon holds pricing power; and finally on utilities and their ratepayers, who carry the cost of a grid build-out sized to demand that may energize years later than announced. The binding constraint is grid equipment and interconnection, not chips and not financing. On-site and behind-the-meter generation is the pressure valve, and it shifts the bottleneck rather than removing it.

05The signals to watch

• Interconnection queue durations and completion rates, especially in PJM, ERCOT, and MISO.
• Large power transformer and generator step-up lead times, and whether announced North American manufacturing capacity actually shortens them.
• The share of new data-center capacity relying on on-site or behind-the-meter generation rather than grid interconnection.
• Copper and grain-oriented electrical steel pricing and lead times as a proxy for grid-build throughput.
• Slippage between announced site energization dates and actual commercial operation.

06Sources

1. Lawrence Berkeley National Laboratory, "Queued Up: 2026 Edition (As of the End of 2025)," June 2026.
2. Goldman Sachs Research, U.S. data-center power demand outlook, 2025 to 2026.
3. U.S. Department of Energy / LBNL, "United States Data Center Energy Usage Report: 2025 Update."
4. Wood Mackenzie, power and distribution transformer supply deficit analysis, 2025.
5. pv magazine USA, "U.S. transformer market faces severe supply constraints as lead times extend to four years," May 2026.
6. OpenAI, Stargate program disclosures, and related hyperscaler capacity announcements, 2025 to 2026.

Correction — July 21, 2026: An earlier version stated that PJM projects reaching operation in 2025 took an average of about eight years from interconnection request to commercial operation. Analysis of Lawrence Berkeley National Laboratory’s 2026 interconnection-queue dataset shows an average of about seven years. The LBNL report citation was also corrected from the 2025 Edition to the 2026 Edition. These corrections do not change the article’s conclusion.