Version 3

Three concurrent debates shape how the US AI datacenter buildout is understood. A widely amplified claim that roughly half of 2026 US datacenter capacity has been canceled or delayed is contested by SemiAnalysis as a methodological artifact of tracking only large public announcements rather than the hyperscaler self-build pipeline [5][6]. Northern Virginia is losing datacenter market share faster than forecasters predicted [7]. FERC's large-load interconnection framework — now extended by an order requiring six grid operators to rewrite their tariffs [9] — enables AI facilities to clear interconnection queues in weeks rather than years [8].

If the cancellation narrative is accepted uncritically, investment and policy decisions are made against a distorted picture of how much AI infrastructure is actually being built. The FERC regulatory push and geographic redistribution together suggest the more meaningful question is not whether US AI datacenters are being built, but where and under what grid arrangements.

A statistic claiming roughly half of US datacenter capacity planned for 2026 has been canceled or delayed spread widely in mid-June 2026, amplified by outlets including TechSpot, The Register, and Yahoo Finance [1][2][3][4]. SemiAnalysis published a rebuttal arguing the figure originates from Sightline Climate data that systematically undercounts the real pipeline [5][6]. Sightline tracks only large publicly announced projects — the subset most prone to slippage — while excluding the hyperscaler self-build capacity that dominates actual construction. SemiAnalysis reports their own bottom-up model showed roughly 1% change in North America hyperscaler self-build forecasts over the prior six months, with co-location moving less than 5%. They note the two largest hyperscalers' self-build capacity alone exceeds Sightline's entire US 'under construction' estimate. SemiAnalysis also identifies AI-generated analysis as a compounding problem: models treat press releases as ground truth, producing reports that amplify the overcounting error.

Geographic and regulatory context complicates the picture. Northern Virginia, which concentrated datacenter infrastructure for roughly two decades, is losing market share faster than most forecasters predicted [7]. FERC's large-load interconnection framework is making alternative locations economically viable by solving the interconnection bottleneck that historically favored established markets. Under this framework, AI factories that self-fund grid upgrades and commit to flexible load operation can clear the interconnection queue in as little as 60 days rather than years [8]. FERC has since ordered six grid operators to rewrite their large-load tariffs, extending the policy reach of the new framework beyond the initial rule [9]. NVIDIA is acting on the opportunity through a partnership with Emerald AI, building AI factories designed as flexible grid assets from the ground up, with commercial deployment planned for later in 2026. NVIDIA's Vladimir Troy cites Lawrence Berkeley National Laboratory data showing roughly a 6-cent-per-kWh reduction in retail electricity prices for every 10% increase in state electricity consumption, framing the flexible-load model as simultaneously pro-growth and pro-affordability.

The sites most likely to appear in cancellation statistics — large publicly announced projects in congested markets — are precisely the projects facing the most friction, while hyperscaler self-builds and FERC-enabled flexible-load facilities in new geographies appear to be proceeding closer to schedule. European AI infrastructure expansion illustrates the same dynamic on a separate track: Mistral's first French datacenter is operational with 18,000 NVIDIA GB200 systems, a consortium of eight French companies has submitted a bid to host a European AI gigafactory, and major French enterprises have moved AI from pilot to production [10]. The US domestic story and European expansion are distinct, but together they show the overall AI infrastructure buildout redistributing rather than halting.

An emerging technical angle concerns 800VDC power distribution in datacenters, which SemiAnalysis is actively examining [11]. No detailed analysis has surfaced yet, but if 800VDC architecture changes the construction economics of AI factories — particularly for new sites designed without legacy electrical infrastructure — it could become a variable in buildout forecasts that current models do not capture.