Power: The Binding Constraint

The central thesis of the datacenters beat is that the 2026 hyperscale sprint is gated by power availability, not capital or chips. Capital is abundant (the four largest hyperscalers are guiding to ~$725B of 2026 capex) and GPUs are shipping in volume — but the scarce input is energized megawatts. Per Data Center Knowledge, hyperscalers need tens of GW of aggregate new power over the next 2–3 years, against a grid that was not built for it. This reframes the whole buildout: the names that win rotate depending on whether power or GPUs is the constraint, and right now the analysis points squarely at power.

The constraint shows up as stalled projects. 36+ projects representing $162B of investment were blocked or significantly delayed — a concrete dollar figure on the bottleneck. Behind the headline, the gating mechanism is physical: high-voltage transformers and medium-voltage switchgear are long-lead items that set the energization schedule for a site, and grid strain makes "outages increasingly common." You cannot pour capital or GPUs into a campus faster than the grid can deliver power to it, and the equipment that connects a campus to the grid is itself in a multi-year queue.

The demand side keeps accelerating against that wall. Total hyperscale capacity is expected to double in ~12 quarters, with a pipeline of 770 future hyperscale facilities against 1,297 operational as of late 2025. Individual targets underline the scale: Meta is targeting >10 GW total by end-2026, with its Louisiana "Hyperion" campus alone scaling from 2 GW to 5 GW ($27B, 2.5M sq ft) and an Ohio "Prometheus" 1 GW supercluster slated to come online in 2026. Other anchor sites — Oracle/Stargate I in Abilene (1.2 GW), Microsoft's Atlanta Fairwater (operational Oct 2025) — are all measured in gigawatts, which is the unit that matters once power, not capital, is the limiting reagent.

Amendment — the constraint is now plural (2026-07-22)

The framing above ("capital abundant, GPUs shipping, power scarce") no longer survives contact with the mid-2026 supply-side evidence. TSMC reports CoWoS advanced packaging and the N3 node sold out through end-2026 with lead times into 2027, and Samsung's preliminary Q2 guidance shows memory pricing power consistent with HBM/DRAM scarcity. Power remains a binding constraint and is still the one with the longest lead times — but as of this compile the honest statement is that packaging, memory, and power bind simultaneously, and capital is the only genuinely abundant input.

This is a downgrade of the original thesis, not a reversal. Power still has the strongest evidence base (a named dollar figure of stalled projects, named long-lead equipment, and a construction-data corroboration), while the packaging and memory constraints each rest on a single analysis-tier source. What changes for positioning is that "power is THE constraint" can no longer be stated as a topic-wide answer — which constraint binds first is now a site-and-vendor-specific question the KB cannot yet resolve.

Two new pieces of evidence sharpen the power side specifically. Microsoft is reported carrying ~$80B of unfulfilled Azure backlog attributed to power constraints — the bottleneck expressed as foregone revenue rather than delayed construction. And ConstructConnect forecasts power-infrastructure construction growing +30.8% in 2026, the supply response arriving in the construction data. Meanwhile Tesla's Cortex 2 pairs a 500MW campus with on-site Megapack storage (see Vertically Integrated Compute), a live test of whether well-capitalised buyers can route around the interconnect queue rather than wait in it.

The queue, quantified — the constraint is throughput, not raw megawatts (added 2026-07-23)

The KB previously had no quantified interconnect-queue figure — the power thesis rested on stalled-project dollars and named long-lead gear, but not on the queue itself. A Carbon Direct study of the PJM and ERCOT queues (the two markets most exposed to data-center load) fills it, and in doing so reframes the constraint:

• Interconnection wait time in data-center load zones: 3 to 4 years — the headline number the thesis was missing. [analysis — Carbon Direct, as-of May 2026]
• Combined active queue: >300 GW of generation + storage across ~1,500 projects (PJM + ERCOT); renewables + storage are 77% of PJM's queue and 87% of ERCOT's.
• Natural gas clears "substantially more quickly" than renewables; post-One-Big-Beautiful-Bill-Act, monthly gas entries to ERCOT's queue rose ~150%.
• The binding constraint is queue throughput / deliverability, not a raw MW shortage. The queued generation is "enough capacity on paper to meet projected data-center demand through 2030" — the shortfall is in connecting it, not in having it.

This is the sourced mechanism behind the "$162B stalled" claim, and it sharpens frontier debate #5: behind-the-meter and on-site generation (Tesla Cortex, ERCOT gas) are attractive precisely because they skip the 3–4-year grid queue — which is why the ERCOT gas-entry surge is happening. Provenance: analysis-tier (a named research firm with a methodology, not a FERC/ISO filing); single-market GW figures vary by outlet and by what is counted, so the >300 GW / 3–4-year figures are the report's own combined-queue estimate as-of May 2026, not audited ISO output.

The census and the price — added 2026-09-24

Two primary sources, both from this rung, change what the page can say.

The national census of the queue. Berkeley Lab's yearly count puts 1,312 GW of generation and about 749 GW of storage in US connection queues at the end of 2025, about 8,200 projects. The whole queue shrank 10%, but gas rose 86% to 253 GW, the only large category that grew. The median project built in 2025 took 61 months from asking to connect to switching on, against 22 months in 2008, and only 13% of the capacity requested in 2000–2020 was ever built (LBNL Queued Up 2026).

This corrects two figures above, and they are kept rather than deleted. The Carbon Direct figure of "more than 300 GW across about 1,500 projects" for PJM and ERCOT combined does not match the census: LBNL counts ERCOT 408 GW (1,527 requests) and PJM 144 GW (1,134 requests) at the end of 2025, about 552 GW together. The two may count differently (storage in hybrid projects, suspended requests, date), so treat the Carbon Direct figure as its own estimate and prefer LBNL's for size. And the Carbon Direct "3–4 year" wait sits below LBNL's national median of 61 months for projects that actually got built in 2025; the two measure different things (a load-zone estimate against a completed-project median), but the census says the wait is longer, not shorter. Note also that LBNL counts only supply waiting to connect. Data-centre load queues are separate and excluded, so the census does not measure AI demand.

The first price for power on the rung. PJM's capacity auction, the annual price for power plants promising to be available at peak, has run under an agreed price collar three times, and the last two cleared at the cap: $329.17 per MW-day (2026/27), $333.44 (2027/28) and $325 (2028/29, cleared 2026-07-14). The last two were the first in PJM's history to fall short of its reliability requirement for the whole region (6,623 MW and 6,831 MW short). PJM itself attributes nearly 5,100 MW of one year's forecast growth to data centres (PJM 2028/29). Roughly, and as our arithmetic, $325 per MW-day is about $119,000 per MW-year.

What it means for the thesis. The throughput framing above survives and is strengthened: the queue is huge on paper, a small fraction gets built, and it takes about five years. What is new is that the shortfall now has a price, and the price is pinned at a cap, so the market is saying power is scarcer than the capped number shows.

Key Claims

• Power, not capital or chips, is the binding constraint on the 2026 hyperscale buildout. (Amended 2026-07-22 — see above; power is now one of three simultaneous constraints, not the sole one.) Evidence: moderate (analysis — Data Center Knowledge) (Hyperscalers in 2026)
• 36+ projects ($162B of investment) blocked or significantly delayed. Evidence: moderate (analysis — Data Center Knowledge) (Hyperscalers in 2026)
• Hyperscalers need tens of GW of aggregate new power over the next 2–3 years. Evidence: moderate (analysis — Data Center Knowledge) (Hyperscalers in 2026)
• High-voltage transformers and medium-voltage switchgear are the long-lead items that set the energization schedule; grid strain makes "outages increasingly common." Evidence: moderate (analysis — Data Center Knowledge) (Hyperscalers in 2026)
• Pipeline: 770 future hyperscale facilities vs 1,297 operational (late 2025); capacity expected to double in ~12 quarters. Evidence: moderate (analysis — Data Center Knowledge) (Hyperscalers in 2026)
• Meta targeting >10 GW total by end-2026 (Louisiana "Hyperion" $27B, 2.5M sq ft, 2→5 GW; Ohio "Prometheus" 1 GW supercluster online 2026). Evidence: moderate (analysis — Data Center Knowledge) (Hyperscalers in 2026)
• Anchor sites are gigawatt-scale: Oracle/Stargate I (Abilene, TX) 1.2 GW; Microsoft Atlanta Fairwater operational Oct 2025; Wisconsin early 2026. Evidence: moderate (analysis — Data Center Knowledge) (Hyperscalers in 2026)
• Capex confirms the priority: ~15% of the $725B 2026 capex ($105B) is earmarked for generation, PPAs, and nuclear contracts. Evidence: moderate (analysis — Value Add VC) (AI Hyperscaler Capex 2026)

Added 2026-07-22

• Power is no longer the sole constraint — CoWoS packaging and HBM/DRAM are also constrained, so the binding input is site- and vendor-specific rather than topic-wide. Evidence: moderate (converging analysis across three sources) (TSMC Q2 2026); (Hyperscalers in 2026)
• [UNSUPPORTED — checked against primary 2026-07-22] Microsoft carries ~$80B of unfulfilled Azure backlog attributed to power constraints — a dedicated primary sweep of Microsoft's FY2025 10-K and both FY26 10-Qs found no corroboration for this figure anywhere. Microsoft discloses total RPO ($633B) and commercial RPO ($627B) but no Azure-only RPO at all, and management is only qualitatively supply-constrained. Treat as reported-not-filed; do not use as the power thesis's quantified leg — the bottleneck stated as foregone revenue, and the most concrete revenue-side estimate of the power gap in this KB. Evidence: weak (single analysis source) (Futurum AI Capex 2026)
• Independent confirmation that power is the named constraint: a third analyst source flags power infrastructure as the binding constraint on deployment pace, matching the Data Center Knowledge finding. Evidence: moderate (two independent analysis sources agree) (Futurum AI Capex 2026)
• [FORECAST] Power-infrastructure construction projected +30.8% in 2026 vs 2025 — the supply response showing up in construction-starts data. Evidence: weak (single analysis forecast) (ConstructConnect July 2026)
• Geographic concentration tracks power availability: VA, TX, LA, IL and NC capture ~60% of new data-center starts spending (trailing 12 months); New England plus the entire Western US account for ~6%. Evidence: moderate (analysis — ConstructConnect, industry-primary data) (ConstructConnect July 2026)
• Behind-the-meter is being tested in practice: Tesla's 500MW Cortex 2 pairs compute with on-site Megapack storage rather than relying solely on grid supply — evidence bearing on frontier debate #5. Evidence: weak (single analysis source; storage is not generation, and the grid/behind-meter split is undisclosed) (Tesla Cortex 2)
• Meta's Ohio (1GW) and Louisiana (toward 5GW) sites are independently confirmed by a second source, corroborating the gigawatt-scale anchor-site claims already on this page. Evidence: moderate (two independent analysis sources agree) (Futurum AI Capex 2026); (Hyperscalers in 2026)

Added 2026-09-24 (census and capacity price)

• US connection queue, end-2025: 1,312 GW generation + about 749 GW storage, about 8,200 projects; total down 10%, gas up 86% to 253 GW. Evidence: strong (government-lab census) (LBNL Queued Up 2026)
• Median time from request to operation for projects built in 2025: 61 months (22 months in 2008). Only 2% of the 2024 queue reached operation in 2025. Evidence: strong (LBNL Queued Up 2026)
• ERCOT 408 GW and PJM 144 GW of active queue at end-2025 — supersedes the Carbon Direct combined ">300 GW" for size. Evidence: strong (LBNL Queued Up 2026)
• PJM capacity price, three collared auctions, the last two at the cap: $329.17 → $333.44 → $325 per MW-day (2026/27 → 2028/29); the last two short of the reliability requirement by 6,623 MW and 6,831 MW. Evidence: strong (market operator's own releases; earliest price from a Talen 8-K) (PJM 2028/29)
• Data centres named as the driver of PJM demand growth: nearly 5,100 MW of the 5,250 MW rise in the 2027/28 peak forecast. Evidence: strong (PJM release) (PJM 2028/29)

Added 2026-07-23 (grid-interconnect queue, quantified)

• Interconnection wait time in data-center load zones: 3–4 years — the queue figure the thesis previously lacked. Evidence: moderate (analysis — Carbon Direct, named research firm, as-of May 2026) (Carbon Direct)

• Combined PJM + ERCOT active queue: >300 GW across ~1,500 projects — superseded 2026-09-24 by the LBNL census (ERCOT 408 GW + PJM 144 GW at end-2025); kept as Carbon Direct's own estimate (renewables + storage = 77% of PJM / 87% of ERCOT). Evidence: moderate (analysis — Carbon Direct) (Carbon Direct)

• The constraint is queue throughput / deliverability, NOT a raw MW shortage — the queue holds "enough capacity on paper to meet projected data-center demand through 2030." Evidence: moderate (analysis — Carbon Direct) (Carbon Direct)

• Gas clears the queue faster than renewables; ERCOT monthly gas-queue entries rose ~150% post-OBBBA — the mechanism behind the behind-the-meter/on-site-gas pull. Evidence: moderate (analysis — Carbon Direct) (Carbon Direct)

• On-site generation is attractive because it skips the 3–4-year queue — gives the Tesla Cortex behind-the-meter test (and the ERCOT gas surge) an explicit mechanism. Evidence: moderate (Carbon Direct reframing + TSLA print) (Carbon Direct); (TSLA Q2 2026)

• Stalled pipeline: 36+ projects, $162B blocked/delayed

• Aggregate new power needed: tens of GW over 2–3 years

• Pipeline vs operational: 770 future / 1,297 operational (late 2025)

• Capacity doubling: ~12 quarters

• Meta target: >10 GW total by end-2026

• Meta Hyperion (Louisiana): $27B, 2.5M sq ft, 2→5 GW

• Meta Prometheus (Ohio): 1 GW supercluster, online 2026

• Oracle/Stargate I (Abilene): 1.2 GW

• Power slice of 2026 capex: 15% ($105B)

Added 2026-07-22:

• Microsoft unfulfilled Azure backlog attributed to power: ~$80B
• [FORECAST] Power-infrastructure construction growth 2026: +30.8%
• Geographic concentration of new DC starts (T12M): VA + TX + LA + IL + NC ≈ 60%; New England + Western US ≈ 6%
• Tesla Cortex 2 behind-the-meter-adjacent storage: 500MW campus, [UNVERIFIED] ~250 Megapacks (~975 MWh)
• Grid-interconnect queue wait time: 3–4 years in DC load zones — filled 2026-07-23 (Carbon Direct); >300 GW combined PJM+ERCOT queue / ~1,500 projects; constraint is throughput, not raw MW
• US capacity shortfall trajectory (today vs 2028): GAP — still no deliverable-shortfall figure (Carbon Direct says the queue holds enough "on paper" through 2030, but not the deliverable path)

Open Questions

• Is power or GPUs the binding constraint in 2026–27? The answer rotates which names win. (Frontier debate #1 — reframed 2026-07-22: the evidence now says packaging, memory and power bind simultaneously; the useful question is which binds first for a given project, which no source answers.)
• How much revenue is the power constraint actually costing? Microsoft's ~$80B unfulfilled backlog is the only quantification in the KB; the industry-wide figure is unknown.
• Does storage (Megapacks) actually relieve the interconnect constraint, or only buffer it? Storage shifts load in time; it does not create generation.
• Who owns the power — utilities, IPPs, or hyperscalers building behind-the-meter? (Frontier debate #5; cross-link to the energy beat.)
• What is the actual grid-interconnect queue wait time? ANSWERED 2026-07-23: 3–4 years in DC load zones (Carbon Direct). The refined open question: how much of the tens-of-GW need is stuck in the >300 GW queue, and how fast can queue throughput (not raw MW) scale — the reframed binding constraint.
• How quickly can transformer and switchgear supply scale to relieve the long-lead bottleneck?
• What is the US capacity shortfall trajectory (today vs 2028)? (Gap flagged in commercial/bottlenecks.md.)

Related Concepts

• Hyperscaler Capex Cycle — capital is abundant; this concept explains why that capital still can't buy gigawatts on demand
• Rack Power Density — denser racks concentrate the power problem (and the cooling problem) per square foot
• Advanced Packaging Bottleneck — the second simultaneous constraint (CoWoS/N3 sold out through end-2026)
• HBM & DRAM Supply Tightness — the third simultaneous constraint (memory)
• Vertically Integrated Compute — the live test of whether buyers can route around the interconnect queue

Changelog

• 2026-06-24 — Created from 2 sources (hyperscaler-power-bottleneck-2026, ai-hyperscaler-capex-2026)
• 2026-07-22 — Amended the core thesis: power is now one of three simultaneous constraints (with CoWoS packaging and HBM memory), not the sole one — a downgrade, not a reversal. Added Microsoft's ~$80B power-attributed unfulfilled backlog, ConstructConnect's +30.8% power-construction forecast, the five-state ~60% geographic concentration, and Tesla's behind-the-meter test. Reframed frontier debate #1.
• 2026-07-23 — Filled the grid-interconnect-queue gap (Carbon Direct): 3–4-year waits in DC load zones, >300 GW combined PJM+ERCOT queue across ~1,500 projects. Key reframing: the binding constraint is queue throughput/deliverability, not a raw MW shortage (queue holds enough on paper through 2030) — which gives the behind-the-meter/on-site-gas pull (Tesla Cortex, ERCOT gas surge) an explicit mechanism. Analysis-tier; a FERC/ISO-primary single-market number stays open.
• 2026-09-24 — Moved to the energy rung's own sources. Added the LBNL Queued Up 2026 census (1,312 GW generation + ~749 GW storage; gas +86%; 61-month median) and the PJM capacity-price series ($329.17 → $333.44 → $325/MW-day; the last two at the cap). Corrected the Carbon Direct PJM+ERCOT queue size against the census, and noted its 3–4 year wait sits below LBNL's 61-month national median. Repointed links to sources held on datacenters.