Stop Saying Half of 2026 US Datacenter Capacity Is Canceled
Reyk Knuhtsen
· SemiAnalysis
· June 18, 2026 at 15:54
· ⏱ 33 min read
| Read on Substack ↗
Summary
The article argues that the widely circulated claim that half of 2026 US datacenter capacity will be canceled is based on flawed 'vibecoded' estimates that count speculative, early-stage announcements as real projects. SemiAnalysis's own tracking shows actual 2026 capacity forecasts have barely moved, and the cancellations are concentrated in a structurally oversupplied early-stage layer that never had equipment orders or realistic delivery timelines. For markets, this means that fears of a datacenter CapEx crash and equipment supplier order collapse are overblown, as the real pipeline continues to progress and backlogs remain locked.
•Bloomberg's April 1, 2026 article traced the 'half of 2026 capacity delayed/canceled' narrative to Sightline Climate data estimating ~12 GW expected online, with only ~5 GW under construction; SemiAnalysis claims even the 'under construction' figure is off by multiples—top-two hyperscalers alone have more than 5 GW of self-build under construction.
•SemiAnalysis's own NA hyperscaler self-build forecast for YE2026 moved only ~1% over the past 6 months, and NA colocation moved <5%, despite widespread media reports of massive delays.
•The 'vibecoded' models (e.g., Claude Code) take press releases at face value, counting speculative megaprojects from inexperienced developers (e.g., Data City/Energy Abundance, APR Energy) as real 2026 capacity, but these projects have no equipment ordered, no interconnection, and often no site control.
•Real delays are exemplified by Nebius NJ (DataOne): initial 4-month delivery target for 50 MW became 10-11 months due to equipment procurement and commissioning issues; Core Scientific Denton missed 250 MW by YE2025 target due to permitting, weather, and a transformer explosion; Oracle/STACK New Mexico campus delayed to 2029 because of gas pipeline permitting failures and NIMBY opposition.
•ERCOT's large-load interconnection queue reached >410 GW, with >87% from datacenters, but SemiAnalysis tracks 311 GW of 'phantom' demand that lacks site control or financing; Texas Senate Bill 6 filtering speculative filings.
•At least 12 US states have filed datacenter moratorium bills, but they cover areas with negligible planned capacity (e.g., Maine has <5 MW); the projects affected were already early-stage and not in 2026 delivery timelines.
•Equipment providers like Vertiv and Schneider Electric have margins above 20% due to supply-demand imbalance; critical SKUs (e.g., Reinhausen tap-changer bushings) have 3-5 year lead times, and capacity expansion (Hitachi Energy South Boston plant) won't come online until 2028.
•Hyperscalers adapt via multiple playbooks: paying directly for grid upgrades (Meta-Entergy), behind-the-meter gas generation, long-lead equipment procurement during scoping, using Chinese OEMs (e.g., Sieyuan transformers at xAI), and modular/prefabricated builds to compress timelines.
Article explicitly names Vertiv as a pure-play datacenter electrical with margins >20% and argues that 'cancellation fear' is misplaced because canceled projects never placed equipment orders, while r
Article explicitly names Vertiv as a pure-play datacenter electrical with margins >20% and argues that 'cancellation fear' is misplaced because canceled projects never placed equipment orders, while real projects have prepaid backlogs running 3-4 years deep. This supports a positive view on Vertiv's order book stability.
Risk: If the macro environment deteriorates or hyperscalers unexpectedly cut CapEx, even prepaid queues could see deferrals.
Schneider Electric is grouped with Vertiv as enjoying >20% margins on datacenter electrical equipment. The same thesis applies: the cancellation narrative does not affect their backlog because cancele
Schneider Electric is grouped with Vertiv as enjoying >20% margins on datacenter electrical equipment. The same thesis applies: the cancellation narrative does not affect their backlog because canceled projects never ordered, and deep lead times ensure slot reallocation.
Risk: Exposure to non-datacenter segments could dilute the positive impact; also, currency and geopolitical risks as a European multinational.
Article details the Oracle/STACK New Mexico campus delay to 2029 due to gas pipeline permitting failures and NIMBY opposition, despite Oracle guiding 1H2027 customer delivery. This specific project de
Article details the Oracle/STACK New Mexico campus delay to 2029 due to gas pipeline permitting failures and NIMBY opposition, despite Oracle guiding 1H2027 customer delivery. This specific project delay could negatively impact Oracle's AI infrastructure timeline and investor sentiment.
Risk: Oracle may route around the delay with other sites; the company's overall datacenter portfolio is large and diversified.
This newsletter, published June 18, 2026,
features Reyk Knuhtsen
discussing VRT, SBGSY, ORCL.
3 trade ideas extracted by AI with direction and confidence scoring.