This is part of WeGoAI's daily series on the things that actually shape our world. Figures from the Financial Times, Jeffries, MSCI, TrendForce, Bloomberg Energy, the US EIA, ERCOT, POWER Magazine and Utility Dive as of 18 October 2026. Not investment advice.

“This is part of WeGoAI's daily series on the things that actually shape our world.”

The Constraint Moved

For two years the story of the AI buildout was chips: GPUs, then HBM, then the power delivery components inside the rack. This week the reporting shifted one level up the stack, and it is a harder problem.

The Financial Times reported on 18 October that at least 47 large US data-centre projects have been delayed, withdrawn or blocked in 2026 — a count the regulatory intelligence firm that compiled it calls conservative, because it captures only the biggest sites. Analysts quoted in the same piece estimate that only about a third of the planned US pipeline is credible. More than 1,000 US communities have sought moratoria or other limits on AI infrastructure this year.

That is the difference between a capex number and a construction site. Total US AI infrastructure spending is projected at more than $10 trillion between 2025 and 2032, roughly 3.6% of GDP a year — but a dollar of intention only becomes a gigawatt of load if it clears land, permits, transformers, turbines, interconnections and, increasingly, lenders.

Announced Versus Built, in Gigawatts

The cleanest way to see the gap is Jeffries' teardown of North American hyperscaler capacity, published on 9 October. Its analyst Rex Petersen tracks what operators say they will build and what is likely to survive contact with labour, cooling equipment and grid interconnection.

The pattern is not a flat shortfall — it is a diverging one. In 2027 hyperscalers want 8.37 GW of new North American capacity and Jeffries expects 7.53 GW, a 10% haircut. By 2030 the announced 13.82 GW becomes a delivered 4.46 GW — a gap of roughly two thirds. Jeffries assigns a different binding constraint to each year: labour in 2026, cooling equipment in 2027 through 2029, and available power in 2030.

Now widen the lens to the whole industry, not just hyperscalers. Petersen's dataset tracks about 65 IT GW of data centres under construction — roughly 140% of everything currently installed — plus another around 361 GW of announced future projects, about 760% of installed capacity. The forecast for what actually energises is 12.9 GW in 2026 and 18 GW in 2027.

The Queue Is a Claim, Not a Load

ERCOT was tracking about 474.7 GW of large-load interconnection requests as of June 2026, of which 420.8 GW were data centres — more than five times the grid's all-time peak demand of 91,089 MW, set on 22 July 2026. Those numbers circulate as if they were demand. They are not: they are options, many of them duplicates, filed by developers who have not yet secured financing, land or equipment. PJM's January 2026 long-term forecast actually cut its summer-2028 peak projection by 4,414 MW (2.6%) to 165,567 MW, explicitly attributing the reduction to better vetting of large-load requests — while still projecting the summer peak to grow 3.6% a year to about 222,000 MW by 2036.

So the honest reading cuts both ways. The queues are inflated by speculative filings; the delivered load is still growing fast enough to strain the physical system. The US Energy Information Administration's October outlook has total US electricity consumption at 4,288 billion kWh in 2026 and 4,356 billion kWh in 2027, after a record 4,195 billion kWh in 2025. Commercial-sector sales — where federal statistics count most data centres — reach 1,549 billion kWh in 2026, up 3.7%, which makes commercial the largest retail electricity sector in the country for the first time, ahead of residential at 1,541 billion kWh.

Why Permits Beat Chips as the Real Constraint

The median new US data centre is now 12 times larger than the median one built a decade ago — 130 MW against 11 MW. You cannot interconnect twelve times the load in the same queue, on the same transmission corridors, with the same transformers, on the old schedule. Hence the rush to behind-the-meter generation: Bloomberg Energy's survey found expectations for fully on-site powered data centres rose 22% in six months, with over a third of data centres expected to run on 100% on-site power by 2030. That trades a permitting queue for a turbine supply chain and a gas contract — a different risk, not a smaller one.

The Financing Half of the Same Story

Physical constraints and credit constraints are the same constraint seen from two sides, because a project that will be late cannot be financed as though it will be on time. The casualties reported this week are consistent: CoreWeave's unsecured bond slid into distressed territory with the shares down almost 25% on the week; SB Energy shelved a planned $50bn listing while needing around $175bn to deliver roughly $430bn of contracted future revenue, mostly from OpenAI; Nscale, whose $45bn Anthropic deal can be cancelled if it misses delivery deadlines, had 25,000 advanced chips online in August against a goal of nearly 500,000; Nebius has financed only about a fifth of the roughly $250bn its planned capacity requires; Cipher absorbed around the first $360mn of cost overruns at a Texas site, requiring roughly $750mn of extra equity; and several Oracle sites for OpenAI's Stargate programme have slipped. OpenAI has pushed its own IPO to next year.

Contracts are adapting to this. AI labs are writing agreements that let them walk away if a partner misses financing or delivery deadlines — which is rational for the tenant and brutal for the neocloud, because it moves the delay risk from the customer's compute roadmap onto the developer's balance sheet.

What to Watch

Three signals will tell you whether this is a normal construction cycle or the start of a real repricing of announced capacity:

  • Whether hyperscaler capex guidance gets cut, rather than just re-phased. Slower capex is bearish for developers and the chip supply chain, but the cash-rich buyers are the least exposed, and some of their shareholders would welcome it.
  • Whether the interconnection wait shortens anywhere. Texas is projected to pass 40 GW of capacity by 2028 — nearly 30% of total US demand and a 142% increase in market share — while California, Iowa, Oregon and Nebraska are each expected to lose more than half of their relative share. That reallocation is the grid answering the question.
  • Whether behind-the-meter power actually arrives on schedule. If it does not, the 2029-2030 gap stops being a forecast and starts being idle silicon in a building that cannot run it.

The buildout is not collapsing. It is being filtered — and the filter is now permits, transformers, skilled labour and interest rates, none of which can be bought with a better GPU allocation.