SemiAnalysis: U.S. Data Center Capacity Downward Revision for 2026 Is Less Than 5%; "Cut in Half" Narrative Uses Wrong Denominator
Claire Weston
Chip research firm SemiAnalysis says its 2026 North American data center supply forecast was revised down by under 5% over the past six months — nowhere near the "half cancelled" narrative circulating in markets, which it traces to a flawed denominator.
Where does the "halved" claim go wrong?
Bloomberg cited Sightline Climate data showing the U.S. expects roughly 12 GW online by 2026, with only about 5 GW under construction — implying roughly half faces delay risk.
This means → the denominator — 5 GW under construction — was drastically understated. SemiAnalysis says the top two hyperscalers' self-built capacity under construction already exceeds 5 GW on its own.
In plain terms = Sightline counted only large, publicly announced projects. A huge volume of real construction was missed, deflating the "under construction" base by multiples and inflating the implied cancellation rate.
How did this narrative snowball?
SemiAnalysis traced the chain: Bloomberg published a report on April 1 about U.S. AI buildout reliance on Chinese electrical components. TechRadar, Tom's Hardware, and The Register then amplified the story in successive rounds.
By the time it reached financial and social media, the narrative had morphed into "half of capacity cancelled" — a supply-chain risk story had become a capacity-collapse conclusion.
This reflects a familiar amplification pattern: one sourced report on supply-chain dependency, re-told enough times, becomes an entirely different claim.
What is actually happening with "delayed" projects?
SemiAnalysis sorted real delays into three buckets. First: newcomers with overly aggressive timelines. Energy Abundance's Data City claimed 300 MW would go live in 2026, but satellite imagery shows no physical construction; APR Energy announced a 400 MW Texas campus for early 2026 while still looking for customers.
Second: underestimated construction cycles. Nebius's 300 MW flagship campus in New Jersey was supposed to deliver 50 MW in four months; it stretched past six. Core Scientific's Denton campus planned 250 MW to CoreWeave by end of 2025, but GB200 design changes forced infrastructure rework, making the deadline very unlikely.
Third: permitting and infrastructure bottlenecks. Oracle and STACK's Project Jupiter in New Mexico tried to skirt major-source emissions thresholds via two micro-grids, but regulators flagged east-side NOx emissions at up to 521 tons/year — far above the roughly 249 tons/year application cap. The project has been pushed to 2029.
How much do these delays really dent deliverable capacity?
SemiAnalysis's core judgment: most of these "delayed" projects were never on track to deliver in 2026 — they represent announcement-level froth, not real capacity loss.
In plain terms = many projects never left the PowerPoint stage. Counting them as "cancelled capacity" is like counting items you never bought as "returns."
Over the past six months, the firm revised its 2026 North American hyperscaler self-built supply forecast down by roughly 1%, and its colocation forecast by under 5%.
What are the leading operators doing to work around bottlenecks?
SemiAnalysis notes that top operators are not waiting passively. They are locking in long-lead equipment early, investing directly in grid infrastructure or power plants, buying powered land, deploying behind-the-meter generation — power sourced on-site, bypassing the grid — and tapping Chinese OEMs to fill equipment gaps.
They are also pushing modular and prefabricated designs to cut construction complexity.
This means → the leading players are using capital and supply-chain muscle to route around traditional grid queues and construction bottlenecks, not sitting idle.
Is the "cancelled equipment orders" narrative credible?
SemiAnalysis says no: early-stage "air projects" never placed equipment orders in the first place — there is nothing to cancel.
Projects that did enter OEM backlogs typically locked their slots with prepayments. Even when a genuine order is cancelled, downstream queue demand absorbs it quickly.
In plain terms = the U.S. large-load interconnection queue exceeds 1 terawatt, but most of that is speculative load applications. There is a massive asymmetry between low-cost announcements and high-difficulty construction — markets should not confuse the clearing of announcement-level froth with a collapse in deliverable capacity.
Content is for reference only, not financial advice.