The Information Machine

AI Datacenter Buildout: Cancellation Myths, Geographic Shifts, and Policy Enablement · history

Version 6

2026-06-26 08:12 UTC · 58 items

What

The US AI datacenter buildout story has four overlapping threads: a contested claim that roughly half of planned 2026 US capacity was canceled or delayed [5][6]; geographic redistribution away from Northern Virginia as FERC policy makes alternative locations viable [7][9]; a structural shift in financing from corporate balance sheets to alternative asset classes [11][12]; and SemiAnalysis's forecast, published June 25, that US grid capacity additions (~15GW/year) are far below projected datacenter power demand (~84GW added by 2030), making behind-the-meter power the structural solution for large operators — BTM projected to supply over half of new US datacenter capacity by 2028 [13].

Why it matters

If SemiAnalysis's BTM forecast is accurate, a large share of the most consequential AI infrastructure will be built outside the traditional grid interconnection framework entirely — changing regulatory exposure, equipment vendor economics, and the practical relevance of FERC's interconnection policy to actual capacity deployment.

Open questions

  • Will behind-the-meter datacenter deployment at the scale SemiAnalysis projects (~40GW+ by 2028) require new regulatory frameworks, or does it operate largely outside existing grid and FERC oversight? [13][10]

  • FERC is reported to be weighing broader federal oversight of AI datacenter grid connections [10] — would such oversight extend to BTM facilities that bypass grid interconnection entirely?

  • Where specifically is datacenter capacity relocating as Northern Virginia loses share, and what mix of power availability, FERC interconnection rules, and land cost is driving those decisions? [7]

  • As AI datacenter financing moves to broadly syndicated loans and private credit [12], does the shift change project economics enough to determine which sites get built versus shelved?

Narrative

A statistic claiming roughly half of US datacenter capacity planned for 2026 was canceled or delayed spread widely in mid-June 2026, amplified by outlets including The Register, TechSpot, 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 bottom-up model showed roughly 1% change in North America hyperscaler self-build forecasts over the prior six months, and notes the two largest hyperscalers' self-build capacity alone exceeds Sightline's entire US 'under construction' estimate. They also identify AI-generated analysis as a compounding problem: models treating press releases as ground truth produce reports that amplify the overcounting error.

Geographic redistribution and regulatory change are reshaping where capacity is actually built. 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 makes alternative locations economically viable by addressing the interconnection bottleneck that historically favored established markets. Under this framework, AI facilities 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 from an enabling rule to an active compliance mandate [9], and is reported to be weighing even broader federal oversight of AI datacenter grid connections [10].

A structural shift in AI datacenter financing is documented by Goldman Sachs and echoed by Morgan Stanley activity in capital markets [11][12]. AI infrastructure is no longer primarily funded through corporate balance sheets; it is increasingly financed through alternative assets — infrastructure funds, private equity, real estate, and private credit. Deal structures now span investment-grade bonds, project finance, high yield, and broadly syndicated loans. This institutionalization of AI infrastructure as an asset class changes the pool of capital available and the conditions under which projects proceed or stall.

SemiAnalysis published a detailed behind-the-meter (BTM) power forecast on June 25, arguing that the US grid cannot physically keep pace with datacenter demand regardless of interconnection policy improvements [13]. Their model projects US datacenter gross power demand growing from +21GW in 2026 to +84GW by 2030, while the US grid adds only approximately 15GW of net-new firm capacity annually. Available grid headroom approaches zero and turns negative by 2027. SemiAnalysis forecasts BTM power — generation installed on-site, bypassing the grid — will supply more than half of new US datacenter capacity by 2028, with the BTM equipment market exceeding 50GW/year by 2029. Developers seeking grid-connected power meanwhile face large financial commitments (letters of credit, security deposits, take-or-pay contracts) and still encounter multi-year delivery delays.

Timeline

  • 2026-06-15: SemiAnalysis announces examination of how 800VDC power distribution may change datacenter electrical infrastructure, featuring DG Matrix. [16]
  • 2026-06-17: The Register reports only half of US datacenter capacity planned for 2026 is actually under construction, amplifying Sightline Climate figures. [2]
  • 2026-06-18: SemiAnalysis publishes a rebuttal arguing the '50% delayed' figure is a methodological artifact of Sightline's data scope, not a reflection of the real hyperscaler pipeline. [5][6]
  • 2026-06-18: NVIDIA's Vladimir Troy publishes analysis of FERC's large-load interconnection framework, describing 60-day queue clearance for AI factories that self-fund grid upgrades and offer flexible load. [8]
  • 2026-06-18: Milk Road AI flags that Northern Virginia is losing datacenter market share faster than almost anyone predicted. [7]
  • 2026-06-18: NVIDIA publishes details of France's AI infrastructure buildout, including Mistral's datacenter running 18,000 GB200 systems and a French consortium bid for a European AI gigafactory. [14]
  • 2026-06-19: FERC orders six grid operators to rewrite their large-load tariffs, extending the regulatory framework from an enabling rule to an active compliance mandate. [9]
  • 2026-06-20: Empirical data covering 2015–2024 shows US household electricity prices did not rise from datacenter growth and may have slightly declined when grid capacity had room to grow. [15]
  • 2026-06-20: Goldman Sachs research documents that AI datacenter financing has shifted from corporate balance sheets to alternative assets including infrastructure funds, private equity, real estate, and private credit. [11]
  • 2026-06-21: Morgan Stanley actively pitches datacenter developers on leveraged loan structures as AI datacenter finance matures into investment-grade bonds, project finance, high yield, and broadly syndicated loans. [12]
  • 2026-06-22: Engineering News-Record reports FERC is weighing federal oversight of AI datacenter grid connections, a potential expansion beyond the tariff-rewrite order. [10]
  • 2026-06-25: SemiAnalysis publishes a BTM power forecast projecting US grid headroom turns negative by 2027, with BTM supplying over half of new US datacenter capacity by 2028 and a BTM equipment TAM exceeding 50GW/year by 2029. [13]

Perspectives

SemiAnalysis

The '50% canceled' narrative is methodologically unsound; their bottom-up model shows the actual hyperscaler build program is essentially on track. Separately, US grid capacity additions (~15GW/year) are structurally insufficient to meet datacenter power demand growth (~84GW added by 2030), making BTM power not just likely but inevitable for the largest operators.

Evolution: Expanded from cancellation rebuttal to a broader grid-constraint thesis with BTM as the structural conclusion.

[5][6][13]

Sightline Climate (implied source of disputed figure)

Tracks publicly announced large-scale datacenter projects; the 50% delayed figure flows from this dataset.

Evolution: No direct response to SemiAnalysis's critique is available in current items.

[5][1][2]

NVIDIA / Vladimir Troy

FERC's large-load interconnection framework is pro-growth and pro-affordability; NVIDIA is acting on it through the Emerald AI partnership and positions itself as the primary infrastructure engine for both US and European AI buildout.

Evolution: Consistent commercial advocacy; empirical 2015–2024 electricity pricing data adds external support to the ratepayer-benefit argument.

[8][14][15]

FERC

Large AI loads can and should be integrated through flexible interconnection arrangements; the tariff-rewrite order directed at six grid operators extends this policy from an enabling framework to an active compliance mandate, and federal oversight of datacenter grid connections is under consideration.

Evolution: Moved from enabling framework to active compliance mandate, and now potentially toward direct federal oversight of AI datacenter connections.

[8][9][10]

Goldman Sachs / Morgan Stanley

AI datacenter financing has structurally shifted from corporate balance sheets to alternative asset classes; deal structures now span investment-grade bonds, project finance, private credit, high yield, and broadly syndicated loans.

Evolution: Consistent; no new positions this pass.

[11][12]

Milk Road AI

Northern Virginia's loss of datacenter market share is faster and more significant than forecasts suggested.

Evolution: No prior stance; early-stage observation without deep analytical support.

[7]

Tensions

  • SemiAnalysis argues the '50% delayed' figure misrepresents the US datacenter pipeline by using a data source that undercounts construction by multiples; Sightline Climate's figures and their amplifiers treat it as a valid market-wide statistic. [5][6][1][2][3][4]
  • SemiAnalysis argues AI-generated market analysis treats press releases as ground truth and produces systematically inflated delay estimates, meaning a significant share of the cancellation narrative is AI-amplified error rather than observed fact. [5]
  • FERC's large-load interconnection framework assumes grid-connected power as the primary solution for AI datacenters; SemiAnalysis's BTM forecast argues the grid cannot add capacity fast enough to meet demand regardless of interconnection improvements, making BTM the structural outcome for large operators by 2028. [9][13]
  • NVIDIA and empirical 2015–2024 data argue large datacenter loads spread fixed grid costs and can lower retail electricity prices; critics of large-load growth argue such facilities strain grids and raise rates for residential customers in constrained regions. [8][15]

Sources

  1. [1] Nearly half of US data centers planned for 2026 are facing delays or cancellation | TechSpot — reactive:ai-datacenter-buildout-geography
  2. [2] Only half of US datacenter capacity planned for 2026 is actually ... — reactive:ai-datacenter-buildout-geography
  3. [3] Half of planned US data center builds have been delayed or ... — reactive:ai-datacenter-buildout-geography
  4. [4] Half of Planned US Data Center Builds Face Delays or Cancellations — reactive:ai-datacenter-buildout-geography
  5. [5] Stop Saying Half of 2026 US Datacenter Capacity Is Canceled: — SemiAnalysis Twitter (2026-06-18)
  6. [6] Stop Saying Half of 2026 US Datacenter Capacity Is Canceled — reactive:ai-datacenter-buildout-geography
  7. [7] This is WILD! — Milk Road AI Twitter (2026-06-18)
  8. [8] How FERC’s Large-Load Interconnection Actions Help Address Grid Stress, Improve Affordability — NVIDIA Blog (2026-06-18)
  9. [9] AI Power / FERC: Six Grid Operators Told To Rewrite Large-Load Tariffs — reactive:ai-datacenter-buildout-geography (2026-06-19)
  10. [10] FERC Weighs Federal Oversight of AI Data Center Grid Connections | Engineering News-Record — reactive:ai-datacenter-buildout-geography
  11. [11] Goldman Sachs Research: AI data centers are no longer being financed mainly by traditional corporate balance sheets. — Rohan Paul Twitter (2026-06-20)
  12. [12] AI data center finance is becoming its own serious asset class. — Rohan Paul Twitter (2026-06-21)
  13. [13] US Grid Constraints: Towards 40GW+ of Behind-The-Meter Datacenter by 2028? — SemiAnalysis Twitter (2026-06-25)
  14. [14] France Advances Europe’s AI Future With NVIDIA Technologies — NVIDIA Blog (2026-06-18)
  15. [15] That’s a pretty unusual finding. — Rohan Paul Twitter (2026-06-20)
  16. [16] Haroon from DG Matrix stops by this week to answer the teams questions about how 800VDC is about to change the electrica… — SemiAnalysis Twitter (2026-06-15)