Somewhere in Virginia, a press release said a data center would be humming by Q3. The slab is poured. There's a chain link fence and a plywood sign with a construction company's logo. That's it. There is no building, no transformers, no switchgear, no servers. The grid connection isn't scheduled until 2029.

Multiply that scene by roughly 70.

Sightline Climate, a climate-tech research firm, just published data on 140 US data center projects that announced 2026 opening dates. Together they represent 16 gigawatts of computing capacity — roughly enough to power 12 million homes. Of that 16 GW, only 5 GW is actually under construction. The other 11 GW is stuck at the "announced" stage with zero visible construction progress, even though data centers typically need 12 to 18 months of build time. The math doesn't work. Between 30 and 50 percent of the 2026 pipeline will slip to later years or die entirely.

The problem isn't money. Hyperscalers will spend more than $650 billion on AI infrastructure this year. The problem is physics. You cannot run a GPU cluster without a transformer, and transformers are now on three-to-five-year delivery schedules. You cannot run a data center without switchgear, and switchgear is sold out through 2028. The specialized medium-voltage gear that connects a campus to the grid is so scarce that US imports from China jumped from 1,500 units in all of 2022 to more than 8,000 in the first ten months of 2025. China — the same China that Commerce keeps threatening with export controls — is now the backbone supplier for the AI build that US policy is supposedly trying to win.

And the grid itself is cooked. Total data center demand in the US is projected to nearly double from 80 GW in 2025 to 150 GW by 2028. Utilities cannot build substations fast enough. In some regions, the interconnection queue for new large loads is five to seven years. Maine just banned new data centers outright. Virginia is openly studying whether to cap them. Community opposition, per Sightline, has become "a true material driver of attrition" — meaning projects aren't just dying from supply chains. They're dying because voters, who pay the electricity bills, have started showing up to zoning meetings.

The 2027 pipeline is worse. Of 21.5 GW announced for next year, only 6.3 GW is actually being built. That's a 71 percent shortfall baked in before the year even starts.

Every AI earnings call this year has featured some version of "we are compute constrained." Every model release has come with "demand is outpacing supply." Executives have framed this as a bullish signal — as evidence that the market is voracious and needs more capacity. That framing is a choice. You could also frame it as: the entire industry has promised capacity it cannot deliver, on timelines it cannot meet, to a grid that cannot support it, using equipment it cannot source. When Oracle fires 10,000 people to free up $8 billion for data centers, and half those data centers slip by 18 months, the company paid a human cost to hit a deadline that was never physically possible.

My Opinion

I'll be blunt. The AI industry has been selling a story where the only constraint is ambition. Build bigger clusters, raise bigger rounds, announce bigger gigawatt numbers at conferences, and the infrastructure will follow. It won't. Steel doesn't care about S-1 filings. Copper doesn't care about Stargate. A transformer takes 30 months to build no matter how many times Jensen Huang says "we need more compute."

What bugs me most is the quiet disappearing act these failed projects are about to perform. In six months, nobody is going to write follow-up stories about the 11 gigawatts that never broke ground. The press release from January 2026 will stay on the corporate blog, still saying "opening Q4," long after Q4 came and went. Analysts will keep modeling 2027 and 2028 AI revenue as if capacity were a solved problem. It isn't. The whole capex-to-revenue pipeline that underwrites the trillion-dollar AI valuations has a five-year transformer order sitting in the middle of it.

The firms that survive the next two years are going to be the ones that already own their power — Google with its PPA portfolio, Microsoft with its nuclear deals, Amazon with its on-site gas. Everyone else is going to discover that "signed the lease" and "operating the cluster" are two very different things, separated by an industrial supply chain that was sized for a slower world. The bottleneck was never the model. It was the copper.

Author: Yahor Kamarou (Mark) / www.humai.blog / 19 Apr 2026