The binding constraint on AI capacity has moved from silicon to electricity. That change rewrites who holds the advantage in the buildout, and it will not stay true forever. Scroll: this is the argument, with the evidence.
Allocation queues, export controls, the price of an H100. In 2023, silicon was the wall every AI project hit first.
Fabs expand, packaging debottlenecks, and every hardware generation loosens the last one's allocation. Electricity is different: generation, transmission and substations are built on a decade clock, by institutions that do not move at software speed.
The constraint shifted from the order book at the fab to the interconnection queue at the utility. The first question asked of a new AI campus is no longer how many GPUs it can get. It is how many firm megawatts it can deliver, and when.
Lawrence Berkeley National Laboratory has tracked the interconnection queue for two decades. Projects completed in the early 2000s waited under two years for their grid connection.1
For projects completed in 2025 the median wait passed five years, with over 2,000 gigawatts of capacity standing in line. In the primary data-center markets, developers plan on five to seven.1
That is the consequence in one line. An AI campus that begins today on an unpowered site waits out the queue before it earns a dollar. The demand curve is not waiting for it.
American industry spent a century building industrial-scale grid connections, then walked away from many of them. Homer City is the flagship: a retired 1,884-megawatt coal plant becoming a $10 billion AI campus on its existing interconnections.2
CoreWeave has contracted roughly 590 megawatts across six Core Scientific bitcoin-mining sites, about $10 billion of revenue over twelve years, on infrastructure originally built for hashrate.5
Industrial companies, miners, utilities, developers: almost none of the holders of these sites has ever run compute. The scarce asset and the capability to monetize it live in different companies, six specialist trades apart.
A creditworthy tenant and the capital structured against that contract. Without these two, nothing else on the site is financeable.
Proving the megawatts are firm, and retrofitting a shell built for another century into liquid-cooled white space.
Silicon, fabric, and the operating layer that meters and sells compute. Six trades that have rarely worked together, and no single vendor sells the assembly. The winners of this window will be the ones who get all six to land in the right order.
A creditworthy tenant on a multi-year, take-or-pay contract turns projected revenue into collateral, and collateral is what project finance runs on: debt raised against the contract rather than against the owner.
What we do about it. Neocloud Group exists for the two sides this thesis creates. If you hold a powered site, or the capital to acquire one, we take you from shell to a live, revenue-generating neocloud: feasibility, tenant, capital, build and operations, run as one plan. If you buy compute at scale, we front sites that have been qualified the way your own diligence would qualify them. The thesis is public. The execution is the engagement.
All sources accessed July 2026. Figures are as publicly reported at their dates; the thesis is re-issued annually.
Tell us what you have: the site, the load, or the mandate. We'll come back with a straight first read.
Start a conversation →