The Data Center Supply Chain Crisis, 2024–2026

A source-cited research report on supply-constrained equipment categories and on the operators, investors, suppliers, and projects that have absorbed the financial consequences.

Prepared August 18, 2026. All figures below are drawn from pages fetched during this research session; each value is linked to the specific URL that states it. Where a value could not be confirmed from a fetched source, it is marked "n.a." rather than estimated.

Executive Framing

Between 2024 and mid-2026 the constraint on artificial-intelligence infrastructure migrated decisively away from silicon and toward the electrical and thermal plant that surrounds it. Microsoft's CFO put the inversion plainly on the fiscal Q1 2026 call: compute hardware "has not really been the constraint," the company was "short of space or power" (Data Center Dynamics on Microsoft's Q1 FY2026 results). Wood Mackenzie sizes the resulting equipment squeeze: US data center capacity is projected to expand from 24 GW to 110 GW by 2030, and the data center share of the electrical equipment market has grown from roughly 2% in 2020 to about 40% (Reuters, July 9, 2026). The same firm projects the US data center electrical equipment market expanding from roughly $20 billion "today" to $65 billion by 2030 (Data Center Knowledge, May 12, 2026).

The consequences show up in three distinct forms, all documented below: (1) lead times for long-lead electrical equipment that now exceed the useful planning horizon of a GPU generation; (2) capital that is abundant and equipment that is not, producing $50M-plus completed shells waiting on energization (Terrapin Consulting Group, 2026 lead-time survey); and (3) a reordering of project structuring around power-first sequencing and behind-the-meter generation.

PART 1, SUPPLY-CONSTRAINED CATEGORIES AND SUPPLIERS

Cross-Category Lead-Time Baseline (Q1–Q2 2026)

The single most useful anchor is a consolidated lead-time table. Two independent 2026 compilations largely corroborate each other.

Two structural indicators corroborate the table: the electrical-equipment industry backlog stood at 5.47 months versus a 3.2-month average across 2017–2019, and the ISM supplier-deliveries index registered 58.9 in July 2026 (Vawn). PJM's own data shows the downstream effect: projects entering service in 2025 spent an average of more than four years waiting to come online after receiving interconnection approval, with supply-chain delays accounting for 23% of milestone change requests and a further 28% categorized as "other" including equipment availability and EPC procurement (Data Center Knowledge).

Electronate's assessment is the sharpest statement of the structural view: the constraint is "structural, not temporary," reshoring investment will not yield meaningful new capacity until 2027 at the earliest, and general contractors are still planning some programs as if electrical equipment could be procured in 8–12 weeks (Electronate).

1. Large Power Transformers (LPTs) and GSUs

Shortage and lead times. Wood Mackenzie assesses a 30% supply deficit in power transformers, with demand up 116% since 2019 and GSU demand up 274%; unit costs have risen 77% for power transformers, 45% for GSUs, and 78–95% for distribution transformers (Wood Mackenzie). Average transformer lead times moved from roughly 50 weeks in 2021 to about 120 weeks in 2024; substation transformers averaged ~140 weeks in 2023, ~150 weeks in 2025, and exceeded 160 weeks in 2026, with some large substation and GSU units ranging 80–210 weeks (Data Center Knowledge citing Wood Mackenzie). Q2 2025 field data showed 128 weeks for standard power transformers, 144 weeks for GSUs, and some orders quoted at four years (IndustrialSage). Gordian reported some LPTs at "well over two years" in Q2 2026 (Gordian/BD+C Q2 2026).

Manufacturers and capacity. Roughly 80% of US power transformer supply is imported (Wood Mackenzie). Announced North American capacity expansions total approximately US$1.8 billion since 2023 (Wood Mackenzie), with a parallel estimate of roughly $2 billion of North American expansion landing by 2028 (IndustrialSage). Siemens Energy's Grid Technologies backlog reached a record €51 billion (~$59 billion), with Q3 FY26 grid orders up 28% to €5.4 billion, a 1.48 book-to-bill, total company backlog of €162 billion, lead times of "three years or more," and a plan to add 50% transformer and GIS capacity by 2030 (MGrid). GE Vernova's Electrification backlog reached $40.6 billion as of July 22, 2026 (Vawn), and $41 billion (+69%) per its Q2 2026 disclosure, including roughly $800 million of US transformer orders after acquiring Prolec (Utility Dive). GE Vernova's $6 billion 2025–28 capex program includes $1 billion for Prolec, $160 million at Greenville, $41 million at Schenectady, and $150 million at Prolec's Goldsboro plant to double medium power transformer output to 420 units per year by 2028 (Industrial Info). Hitachi Energy announced a $300 million transformer investment in Hefei, China on August 17, 2026, within a $9 billion global plan (Hitachi Energy).

Allocation, escalation, and input constraints. Transformer costs are expected to rise a further 4–10% over the coming year (Reuters); real transformer prices rose roughly 40% from 2020 to 2024 (Congressional Research Service R48933). Grain-oriented electrical steel is the binding input: Cleveland-Cliffs is the sole US GOES producer (IndustrialSage) and 85% of US GOES imports originate in South Korea (CRS). Tariff policy compounds cost: 50% Section 232 steel tariffs took effect June 3, 2025 and a 50% copper tariff on August 1, 2025 (CRS). Distribution transformer waits, which reached two years in February 2024, had improved to 30 weeks by Q2 2025 (CRS), the one category showing genuine relief.

Impacted operators. A 50 MW AI campus requires 4–6 LPTs (CPMPros). Roseville Electric now buys equipment five years ahead of need (Reuters). Constellation Energy told FERC (Docket ER26-2028) that restarting Three Mile Island Unit 1, contracted to supply Microsoft, requires significant transmission work and may not be "fully deliverable" until those upgrades complete (Data Center Knowledge).

2. Medium-Voltage Switchgear and Paralleling Gear

MV switchgear prices rose approximately 50% (IndustrialSage). Lead times cluster at 40–78 weeks across surveys, but data-center-specification MV switchgear is "approaching 2–3 years" at peak-demand manufacturers, with standard commercial and industrial orders competing for the manufacturing slots left over after data center allocations are filled (Electronate). Schneider Electric, Eaton, ABB, Siemens, and GE Vernova are prioritizing large data center supply agreements involving hundreds of millions of dollars over multi-year programs (Electronate).

Capacity responses: Schneider Electric invested $140 million in US switchgear manufacturing expansion and ABB doubled HV switchgear capacity at Mount Pleasant, Pennsylvania (Electronate); Schneider is also building a new Tennessee plant for custom power distribution equipment (Facilities Dive). Powell Industries is the clearest read on allocation dynamics: Q2 FY26 orders of $490 million (+97%), a 1.7x book-to-bill, and $1.8 billion backlog (Powell Industries), including a >$400 million behind-the-meter mega order from a neocloud in April 2026, the largest in company history, gigawatt-scale in phase one, with a two-year burn (Powell Q2 FY26 transcript). By August 3, 2026 Powell's backlog was reported at $2.4 billion with a 3.0 book-to-bill (Vawn).

Practical allocation mechanics have emerged: a 10–20% deposit on switchgear value is now a standard way to secure a production slot, specification changes can generally be absorbed until a manufacturing freeze 4–8 weeks into the lead time, and tier-2 manufacturers can offer lead times 20–30% shorter for brand-flexible LV scope (Electronate).

Eaton's order book quantifies the demand side: Q1 2026 Electrical Americas orders +42% organic on a 12-month rolling basis, Electrical Americas backlog +44%, Electrical Global backlog +73%, book-to-bill 1.2, and $11 billion of acquisitions including Boyd Thermal (Eaton). Vertiv's backlog reached $15.0 billion as of February 11, 2026 (Vawn).

3. Diesel Generators

Caterpillar closed Q1 2026 with a record $63 billion backlog, up $28 billion or 79% year over year, with nearly $12 billion of new backlog arriving in the quarter and power generation sales up 41% to $2.8 billion, an annual run rate above $11 billion, driven primarily by large reciprocating engines and turbines for data centers (The Motley Fool). CEO Joe Creed said some customer orders extend well into 2028, and Caterpillar is expanding large reciprocating engine capacity to nearly three times its 2024 level, with most of the associated capex scheduled for 2027–2029 (The Motley Fool). A separate account puts Caterpillar's Lafayette expansion at $725 million, adding roughly 15 GW/yr of capacity by 2027, and reports Cummins large-genset allocation slipping into 2028 with Western OEM large-genset lead times of 40–107 weeks; Generac engines have begun appearing in Stargate permits (Manufacturing Magazine).

Generac confirmed the hyperscale channel on June 2, 2026, announcing a global supply agreement with an unnamed leading hyperscale data center operator for large-megawatt backup generators, awarded after multiple factory visits and quality audits, alongside an EPC Power collaboration and the Enercon acquisition (generator enclosures and switchgear), plus capacity expansion at Beaver Dam, Oshkosh, and Sussex, Wisconsin and across APAC, EMEA, and Latin America (Generac). Dollar amounts for the Generac expansions were not disclosed: n.a.

Spire's July 2026 market update puts generators at 22–90 weeks generally and 45–90 weeks above 600 kW (Spire); Linesight measured Americas generators at ~100–105 weeks versus 60–70 weeks in 2024 (Linesight). Named impacted projects for diesel specifically: n.a. beyond the Stargate permit reference.

4. Gas Turbines

This is the tightest category in the report. GE Vernova's Q1 2026 disclosed 21 GW of new turbine agreements, contracted volume rising from 83 to 100 GW, slot reservations from 43 to 56 GW, pricing up 10–20 points per kW versus Q4 2025 with a $600/kW target by end-2027 (roughly 3x 2019), only about 10 GW of capacity remaining across 2029 and 2030, and roughly three-year lead times (Power Engineering). By Q2 2026 the backlog reached 116 GW, expected to hit 125 GW by year-end 2026, with the company booking into 2031 and roughly 20% of customers being data centers (Utility Dive). The trajectory was visible early: in April 2025, GE Vernova said new orders were receiving 2028 delivery dates, was already discussing 2030 fulfillment, had added 7 GW in Q1 2025 to reach a 29 GW backlog, and reported roughly one-third of turbines destined for data centers plus 21 GW of customer reservations not yet in backlog (Energy Connects).

Siemens Energy's gas backlog reached 69 GW with capacity of 35 large-frame units (adding 15 in 2027) and 80 medium units (adding 20 by 2028), against an addressable market of 120 GW/yr, roughly half of it US (Utility Dive).

Mitsubishi Power told S&P Global on July 3, 2026 that turbine delivery now takes five years, order books are full through 2030, and it is contracting 2031–2034 deliveries against annual gas demand exceeding 100 GW versus 30 GW pre-COVID (S&P Global). MHI is doubling capacity, with manufacturing costs having nearly doubled (Data Center Dynamics). Solar Turbines, Kawasaki, Baker Hughes, and Ansaldo: specific lead-time or allocation disclosures n.a. from fetched sources.

5. Reciprocating Gas Engines and Aeroderivatives

Behind-the-meter engine plants have become the principal workaround for interconnection delay, and the order sizes are now utility-scale. Documented commitments include: Wärtsilä's 412 MW / 40× 34SG Ohio hyperscale project, part of a US data center total above 1.6 GW across four projects with delivery in early 2028; PROENERGY's 13× PE6000 units totaling 650 MW for Crusoe; an INNIO–Rehlko 1.25 GW Jenbacher framework, raised from 700 MW; and Nscale's Monarch, West Virginia 2 GW project using CAT G3500 engines, tied to a 1.35 GW Microsoft/Nvidia letter of intent (Power Engineering). Additional commitments: INNIO–VoltaGrid at 2.3 GW (92 packages × 25 MW), Caterpillar–Joule at 4 GW in Utah, Caterpillar–Hunt at 1 GW, and 2 GW of CAT G3516 engines for AIP's Monarch (Power Magazine).

Wärtsilä's capacity response is a ~€140 million investment to raise production capacity 35% at its Vaasa, Finland hub, operational by early 2028, with the CEO identifying the US as roughly 50% of the data center market and the fastest-growing (Reuters, February 4, 2026). Notably, that same release reported Wärtsilä orders missing forecasts even as it talked up data center prospects, an indication that the engine pipeline is being negotiated faster than it is being booked.

6. UPS Systems and Lithium-Ion BESS

UPS lead times run 30–48 weeks for 500–1,500 kVA units, 40–60 weeks for modular 2,000+ kVA (Terrapin), and 40–72 weeks for modular systems (CPMPros). Vendor allocation is now contractual: Schneider Electric signed a $373 million Supply Capacity Agreement with Digital Realty guaranteeing production capacity for UPS systems, LV switchgear, and prefabricated power skids, including a dedicated switchgear production line explicitly intended to help Digital Realty navigate equipment bottlenecks (Data Center Frontier). Data Center Frontier's assessment is that these multi-billion-dollar SCAs leave other buyers competing for remaining supply, with the likely result of longer lead times and higher prices, a pre-allocation dynamic explicitly compared to how GPU capacity was pre-allocated.

On the BESS side, Fluence Energy provides the cleanest documented case of production-side failure meeting data center demand. Fluence reported a record $6.4 billion backlog covering 45.6 GW and 163.7 GWh as of June 30, 2026, but simultaneously cut FY2026 revenue guidance to $2.9–3.1 billion from $3.2–3.6 billion, a $400 million midpoint reduction, because approximately $400 million of project deliveries will move into fiscal 2027 due to production problems at a new international contract manufacturing plant and construction delays at a new US contract manufacturing plant (MGrid). The Houston facility, designed for 15 GWh of annual capacity, was delayed a few months by construction and automation-equipment issues, pushing production back one quarter to full output in fiscal Q1 2027; it operated on generators while awaiting its own grid connection (Fluence Q3 FY26 transcript). Fluence's data center bookings reached ~$850 million through July 2026, including a $300 million behind-the-meter developer project that converted from lead to order in under three months and $550 million awarded by a single hyperscaler in July 2026; its data center pipeline reached 16 GWh, up more than 35% sequentially (Fluence transcript). Critically, MGrid notes that utilities and large loads pursuing the same production slots as data center customers now queue behind them, and that 2027 delivery dates should be treated as conditional on factory ramp (MGrid).

Tesla's Megapack has become a documented data center product: Tesla sold $430 million of Megapacks to xAI in 2025, about 3.4% of its energy revenue, which rose 27% to $12.8 billion; the newer Megablock configuration bundles four Megapacks to a single transformer (CNBC). Powin, Piller, Huawei, and Mitsubishi UPS-specific allocation data: n.a.

7. Fuel Cells

Bloom Energy is the standout. FY2025 revenue was $2.02 billion with product backlog of approximately $6 billion (up 2.5x) and total backlog of roughly $20 billion (Bloom Energy). A secondary account puts the $20 billion backlog at +250% year over year, with capacity rising from 1,180 MW to 2,000 MW in 2026, roughly 90-day deployment timelines, a 2.8 GW partnership, and Project Jupiter in New Mexico (Ad-hoc News).

FuelCell Energy is repositioning around the same demand but from a much weaker base. Its Q2 FY2026 sales pipeline reached 4 GW, up 267% from Q1 2026, but explicitly not signed agreements, while total backlog fell 9.9% year over year to $1.135 billion as of April 30, 2026, with product backlog collapsing to $36.1 million from $98.2 million (FuelCell Energy). The company introduced a standardized 12.5 MW Energy Block to shorten time-to-power in grid-constrained markets and raised its Torrington, Connecticut expansion target to 500 MW annualized from 350 MW, at an estimated cost of $200–275 million over twenty-four months (FuelCell Energy). Plug Power and Ballard data center specifics: n.a. from fetched sources.

8. Wire and Cable

Copper is the binding input. Gordian's Q2 2026 index shows copper wire up 5.30% quarter over quarter and 18.42% year over year, and conduit up 5.22% QoQ and 7.20% YoY, describing copper markets as facing their "largest supply deficit since 2004" (Gordian/BD+C). More than half of all mined copper ends in electrical applications, and US tariffs on steel and aluminum mill products are pushing producer price indices higher (ConstructConnect). A 50% copper tariff took effect August 1, 2025 (CRS).

Prysmian, having acquired Encore Wire, reported at Q2 2026 that its European and North American power grid capacity is full, that it is signing take-or-pay contracts with guaranteed margins, and that it expects €10 billion of incremental optical revenue over ten years, including a €5.5 billion Molex agreement with a €550 million down payment (Prysmian Q2 2026 transcript). Encore Wire unveiled a new 340,800 sq ft copper building-wire plant plus a 1-million-sq-ft service center in McKinney, Texas on April 13, 2026 (Prysmian). Prysmian also agreed to acquire Atkore for $3.8 billion; Atkore derives 10–15% of revenue from data centers across 30 factories (MarketBeat). MV cable lead times run 12–30 weeks (Spire). Nexans, Southwire, and LS Cable specifics: n.a.

9. Busway

Busway/bus duct lead times span 30–52 weeks (CPMPros), 16–40 weeks (Spire), 20–36 weeks depending on rating and manufacturer (Electronate), and 27 weeks for bus duct specifically (Vawn). This is the least-stressed of the major electrical categories in relative terms. Vendor-specific allocation for Eaton, Siemens, Schneider, Legrand/Starline, Anord Mardix, and Vertiv busway: n.a.

10. Conduit and Cable Tray

Conduit prices rose 5.22% quarter over quarter and 7.20% year over year in Q2 2026 (Gordian/BD+C), and lighting is seeing 5–20% tariff pass-through as an indicator of the broader commodity-electrical pass-through environment (Spire). Atkore reported double-digit data center growth in Q2 FY26 (Atkore transcript) and is being acquired by Prysmian for $3.8 billion, with 10–15% of revenue from data centers (MarketBeat). Nucor, Wheatland, Eaton B-Line, and Legrand/Cablofil specifics: n.a.

11. ATS and STS

ATS/STS units above 2,000 A carry 20–36 week lead times (CPMPros); ATS generally runs 24–40 weeks (Terrapin). Schneider Electric launched the ASCO 300 Series multivoltage transfer switch for commercial backup power in June 2026 (Schneider Electric press release). Russelectric, Cyberex/Layer Zero, Cummins, and Kohler ATS allocation situations: n.a.

12. Networking (400G/800G)

Arista Networks' Q1 FY26 call is the most explicit vendor statement of multi-input scarcity: industry-wide shortages across wafers, silicon, CPUs, optics, and memory; demand outstripping supply; purchase commitments increased from $6.8 billion to $8.9 billion; inventory at $2.38 billion; management describing the situation as a "one or two year phenomena"; and qualification cycles extending from 2–4 quarters to 6–8 quarters, against more than 100 cumulative 800G customers (MarketBeat). The lengthening of qualification cycles is the most consequential detail, it means the switch to alternate suppliers is itself a multi-quarter project.

On the HPE/Juniper side, HPE's Q1 FY26 networking revenue reached $2.7 billion, up 152% reported (7% normalized ex-Juniper), with AI backlog above $5 billion, 64% of cumulative AI orders from enterprise and sovereign rather than hyperscale customers, and strong demand for Juniper QFX data center switches and PTX/MX routers (Investing.com). HPE raised FY2026 networking growth guidance to 68–73%, said outright that it "does not have enough supply to meet current demand," expects elevated prices to persist through 2027, shortened quoting cycles, retained the ability to adjust pricing between order and shipment, and, critically, said its strategy for the remainder of the year prioritizes higher-margin orders, which "may have an impact on our AI systems revenue growth" (Reuters, March 9, 2026). That is a vendor explicitly choosing to forgo revenue rather than fill constrained demand.

13. Optics, DAC, and AEC

The optics shortage is quantified and durable: 800G production is running 40–60% short of demand through 2027, and 1.6T is projected 30–40% short through 2029, against a market moving from $16.5 billion in 2025 to $26 billion in 2026 and unit volumes from 24 million to 63 million (TechTimes). Nvidia's response was vertical capture: $2 billion into Lumentum and $2 billion into Coherent announced March 2, 2026, locking up EML supply and pushing non-Nvidia customers' lead times beyond 2027 (TechTimes).

Concentration risk is severe and now regulatory: Chinese suppliers account for roughly 60% of datacom optics revenue, Innolight alone posted $2.6 billion in 1Q26 for a 34% share, and Cignal AI reports FCC preparation for a ban on new Chinese optical modules alongside shortages at Fabrinet (Cignal AI). A supply-side ban layered on a 40–60% shortfall is the single largest unhedged risk in this category.

14. Fiber Cable and Connectors

Fiber has moved from commodity to allocated good. Data center fiber demand rose 76% in 2025 and is projected to reach 30% of global fiber demand by 2027, versus under 5% in 2024; fiber prices rose roughly 70% from $3.70 to about $6.30 per fiber-km; lead times run 20 weeks for large buyers and up to one year for small ones; and North American demand is growing 22–25% in 2026 against supply growth of 12–19% (Tom's Hardware). Hengtong and FiberHome are booked into early 2027 (Tom's Hardware). Incab America reported in August 2025 that lead times had stretched to a year against a normal 8–12 weeks, that one US glass maker was sold out through 2026, that Corning was "quite tight," and that 80% of germanium supply originates in China (Light Reading).

The hyperscaler response has been to buy the factory output in advance. Corning and Meta announced a multiyear agreement valued at up to $6 billion on January 27, 2026, under which Meta becomes anchor customer for Corning's Hickory, North Carolina optical cable facility, with a new Hickory plant plus expanded North Carolina capacity and 15–20% projected employment growth on a base of more than 5,000 state employees (Corning). Nvidia and Corning announced a long-term partnership on May 6, 2026 to increase US optical connectivity manufacturing capacity 10x and US fiber production capacity by more than 50%, via three new facilities in North Carolina and Texas creating more than 3,000 jobs (Corning). Amazon separately struck a multi-billion-dollar North Carolina fiber deal with Corning, and Microsoft, Corning, and Heraeus announced hollow-core work in September 2025 (Tech Blog / ComSoc).

The pattern is unmistakable: four of the largest buyers have each pre-purchased dedicated fiber capacity, which by construction leaves the residual market, telecom carriers, rural broadband, smaller colocation, competing for what is left.

15. Structured Cabling (MPO/MTP)

No fetched source provided category-specific lead times, allocation status, or price escalation for MPO/MTP structured cabling from Corning, Panduit, Belden, Leviton, or Legrand: n.a. The closest available evidence is the upstream fiber and connectivity constraint documented in section 14, and Prysmian's €5.5 billion Molex optical agreement with a €550 million down payment (Prysmian transcript), which indicates connectivity-level capacity is being contracted the same way fiber is.

16. Cooling (Air and Liquid)

Liquid cooling has its own distinct shortage map. As of July 2026: CDUs and chillers at 26–40 weeks; new 1–3 MW CDU orders at 6–10 months; brazed-plate heat exchangers at 30–38 weeks; quick-disconnect couplings at 18–24 weeks; manifolds at 18–24 weeks; and dielectric coolant being rationed (SourceBySpec). Chillers above 500 tons run 40–60 weeks (CPMPros), and mechanical equipment globally runs 20–43 weeks (Linesight).

The supply base is consolidating fast and passing through cost. Ecolab closed its $4.75 billion acquisition of CoolIT in March 2026 and implemented a 10–14% energy surcharge from April 1, 2026, against a tariff backdrop of 25% on Canada, 20% on the EU, and roughly 145% on China; Google has qualified Envicool CDUs, and a $150,000 CDU involves roughly 15 tier-1 suppliers (The Cooling Report). On the acquisition side, Eaton agreed to buy Boyd's thermal business for $9.5 billion, Vertiv acquired CoolTera in 2023 and has a planned $1.0–1.25 billion acquisition of PurgeRite (liquid-cooling flushing and filtration services), and Schneider acquired a controlling stake in Motivair (Data Center Frontier; Facilities Dive).

Vertiv's Q4 2025 quantifies the demand shock: orders +252%, backlog $15 billion (+109%), book-to-bill 2.9x, roughly $1 billion of acquisitions, and capex rising to 3–4% of sales, with the CEO describing liquid cooling capacity growing "really, really, really rapidly" (24/7 Wall St). But Vertiv's Q2 2026 shows the execution strain: management disclosed "minor timing shifts in Q2 revenue" from multiphase project execution and "temporary supply chain dynamics," expected Americas congestion to resolve in 2H26, and carried inventory of $2.523 billion versus $1.457 billion at year-end 2025, with deferred revenue of $3.634 billion and capex at 4% of sales (Vertiv Q2 2026 transcript). Schneider's largest cooling engagement to date sits inside a $1.9 billion Supply Capacity Agreement with Switch covering prefabricated power modules, the largest data center cooling project in North America, and the first US deployment of Uniflair chillers (Data Center Frontier).

Specific allocation status for Stulz, Trane, Carrier, JCI, Munters, Airedale, JetCool, Asetek, GRC, Submer, LiquidStack, and Iceotope: n.a.

17. Piping and Valves

No fetched source provided lead times, allocation status, or escalation specific to Parker, Swagelok, or Danfoss. The closest confirmed data points are quick-disconnect couplings at 18–24 weeks and manifolds at 18–24 weeks, with dielectric coolant rationed (SourceBySpec). Vertiv's planned acquisition of PurgeRite for $1.0–1.25 billion signals that the fluid-side services layer is now considered strategic (Data Center Frontier). Vendor-specific figures: n.a.

18. Structural Steel, Precast, and Raised Floor

Steel prices began rising again in 2025 after retreating from 2021–2022 highs, as trade frictions, energy costs, and freight rates constrained supply; aluminum and copper face pressure from electrification demand absorbing global capacity; and steel, aluminum, and copper are all identified as long-lead materials (ConstructConnect). The demand pull is extreme: year-to-date data center starts through November 2025 reached $53.7 billion, up 138.6% year over year, with 22 projects breaking ground in November 2025 alone representing more than $9.8 billion, nearly four times November 2024 (ConstructConnect). Mid-2025 saw $14 billion of data center construction starts in a single month (Electronate).

Labor, not material, is now the marginal constraint on the construction side. Equinix identified skilled trades, electricians and plumbers specifically, as a major constraint, cited greater Chicago as having substantial planned activity and insufficient labor, noted electricians earning $150 per hour in some markets, and said manufacturing-side cost pressure has moderated while on-site labor costs remain elevated, driving increased interest in prefabrication and off-site completion (MarketBeat on Equinix). Digital Realty's cost structure gives the shell/fit-out split: fully outfitted US data centers average approximately $13.3 million per megawatt, with roughly 20–25% of final construction cost spent on the shell before the building is prepared for occupancy (Los Angeles Times). Raised-floor-specific lead times: n.a.

19. Semiconductors, HBM, and CoWoS

Advanced packaging is sold out. CoWoS is sold out through 2026 with lead times exceeding 50 weeks; TSMC Arizona is booked through 2027; N3 runs 52–78 weeks and N2/A16 78–104 weeks; CoWoS capacity is scaling from 13,000 wafers per month at end-2023 toward a 130,000 wpm target by end-2026; Nvidia took roughly 70% of CoWoS-L in 2025 and 595,000 wafers in 2026 (about 60% of demand), versus Broadcom at 150,000 and AMD at 105,000; and HBM3E is fully allocated through 2026 with prices up 15–22% year over year and SK hynix shortage conditions extending into late 2027 (AI Expert News on TSMC).

Commodity memory is worse. TrendForce raised its 1Q26 conventional DRAM contract-price forecast to +90–95% quarter over quarter from +55–60%, and NAND to +55–60% from +33–38%, with server DRAM specifically expected to rise around 90% QoQ, the largest quarterly increase on record, while suppliers allocate capacity across strategic accounts and reallocate NAND lines to DRAM; enterprise SSD prices were expected to rise 53–58% QoQ (TrendForce, February 2, 2026). An earlier reading had DRAM up 50–55% QoQ with a 3:1 HBM wafer tradeoff and Micron exiting the consumer PC segment (CNBC); Samsung, SK hynix, and Micron are reported sold out for 2027, with customers allocated only 60–70% of requested volumes (Seeking Alpha).

The financial pass-through is now visible at hyperscaler scale. Amazon raised expected 2026 capex to approximately $220 billion from about $200 billion, attributing the increase primarily to higher memory costs rather than a broader construction push, and named "resource and supply volatility, including for memory chips" as a business risk (Data Center Knowledge). Microsoft's FY2026 spending is expected to reach $190 billion partly due to "$25bn higher component pricing," driven particularly by memory (Data Center Dynamics). Dell flagged atypically rapid cost increases across DRAM, NAND, HDD, and leading-edge nodes, adopted a "supply-first stance," and said it would manage impact through configuration flexibility, availability management, and repricing where necessary (Futurum Group).

20. GPUs and AI Accelerators

Vendor backlogs are the proxy for scarcity. Dell reported $12.3 billion of AI server orders in Q3 FY2026 with $30.0 billion year to date, $5.6 billion shipped in the quarter and $15.6 billion YTD, and a record $18.4 billion AI server backlog, with a five-quarter pipeline described as "multiples of backlog" and mix shifting toward GB300 (Futurum Group). At the March 2026 reporting point Dell's AI backlog was $15.2 billion, up from $2.9 billion in mid-2024, with GPU and liquid-cooling component constraints explicitly identified as pacing industry growth (Financial Content).

The Blackwell ramp was not smooth. In January 2025, Nvidia's largest customers, Microsoft, AWS, Google, and Meta, each with GB200 rack orders worth $10 billion or more, were reported to be delaying or cutting orders because first shipments were overheating and showed chip-interconnect glitches; some customers waited for a later revision or bought older Hopper parts, and OpenAI asked Microsoft for Hopper chips after delays emerged, while Microsoft had planned to install GB200 racks with at least 50,000 Blackwell chips at a Phoenix facility (Reuters).

Alternative silicon is scaling. Amazon's Project Rainier in Indiana was filled wall-to-wall with Trainium2 with no Nvidia GPUs, characterized as the largest known global deployment of non-Nvidia computing, expected to reach one million chips by year-end, with Amazon having doubled its Trainium order (CNBC). Amazon's AI and custom-silicon businesses each surpassed a $25 billion annual run rate, with Anthropic and OpenAI making multi-year, multi-gigawatt Trainium commitments (Data Center Knowledge). On AMD, Oracle announced OCI availability of Instinct MI355X scalable to 131,072 GPUs, building on prior plans for roughly 30,000 units (Data Center Dynamics). Explicit AMD, Groq, and Cerebras allocation or sold-out statements: n.a.

Perhaps the most important supply-chain fact in this category is generational obsolescence risk. Nvidia's data center processors moved from a two-year to a one-year release cycle, Vera Rubin was unveiled at CES in January 2026 and already in production delivering five times Blackwell inference performance, and securing a site, connecting power, and standing up a facility takes at least 12–24 months (CNBC). When accelerator cycles are shorter than power-delivery cycles, every long-lead procurement decision carries obsolescence risk.

PART 2, LOSSES, DELAYS WITH MONETARY CONSEQUENCES, AND CLOSURES (2023–2026)

A. Hyperscalers

Narrative. The hyperscaler record between early 2025 and mid-2026 divides into two phases that are easily confused. The first, Q1 2025, looked like oversupply: Microsoft canceled roughly 200 MW of leases and paused SOQ conversions, and TD Cowen's follow-up note put the total at up to 2 GW across the US and Europe. But the reported cause was never demand weakness. Microsoft's own framing was that it was "short power and space" (Reuters; DCD), and the abandoned European leases were absorbed by Google and Meta almost immediately (DCD), indeed "to an extent not previously reported" (DCD). What Microsoft was actually doing was re-optimizing a portfolio against a $80 billion budget it had begun to exceed at the end of 2024 (DCD).

The second phase, from late 2025 through mid-2026, is unambiguous scarcity. Microsoft spent $11.1 billion on data center leases in a single quarter (Q1 FY2026) on total capex of $34.9 billion, with CFO Amy Hood stating that compute hardware "has not really been the constraint" (DCD). Microsoft signed roughly $33 billion of neocloud compute deals, $14 billion with Nscale, $19.4 billion with Nebius, plus CoreWeave and Lambda, precisely because it could not build fast enough (DCD). Amazon's Jassy said the company would "still not have enough capacity to meet all the demand we have in 2026," expected the same in 2027, and noted that much of planned 2027 capacity was already reserved and that "the demand we already have for 2028 is striking," on a $496 billion AWS backlog (Data Center Knowledge). Google's $462 billion backlog doubled while it described itself as compute constrained (TechCrunch). Meta went from $72.22 billion of 2025 capex to $115–135 billion guided for 2026 while renting from three competitors (Reuters).

The Abilene episode is the clearest single illustration of how supply-chain and generational risks compound. Oracle had secured the site, ordered hardware, and spent billions on construction and staff, but power was not projected online for a year, and OpenAI wanted clusters built on Vera Rubin rather than Blackwell (CNBC). Beyond the newer-silicon motive, DCD reported that a multi-day winter-weather liquid-cooling outage damaged the OpenAI–Crusoe relationship, that Nvidia had paid Crusoe a $150 million deposit, that OpenAI's compute commitment had been recut to roughly $600 billion through 2030 from $1.4 trillion by 2033, and that SoftBank was seeking $40 billion in loans; Meta was in talks to pick up the Crusoe capacity with Nvidia's help (DCD). Separately, Stargate's broader 10 GW program was reported delayed by control disputes among OpenAI, Oracle, and SoftBank (Tom's Hardware).

CoreWeave's Denton episode is the single largest disclosed monetary consequence of a supply-chain delay in this report: a third-party developer's construction delay triggered a guidance cut and a securities class action alleging a $14 billion market-cap loss (GlobeNewswire). The site did recover to more than 16,000 GPUs by end-December, on a $6.1 billion Core Scientific conversion supporting $22.4 billion of OpenAI contracts (DCD), and Core Scientific went on to reject CoreWeave's acquisition offer while retaining a 200 MW hosting deal (DCD).

B. Colocation and Wholesale

Narrative. The colocation sector's defining 2026 characteristic is stranded capital: buildings that exist and cannot be energized. Two Santa Clara facilities, Digital Realty's 430,000 sq ft shell, applied for in 2019, and Stack's 551,000 sq ft 48 MW building, which sought approval in 2021, sat empty awaiting grid connections, with Silicon Valley Power's $450 million system upgrade not scheduled for completion until 2028 (DCD; Los Angeles Times). At Digital Realty's disclosed ~$13.3 million per MW for fully outfitted capacity and 20–25% of that spent on shell before occupancy readiness, the two idle shells represent a large, precisely quantifiable carry cost with no revenue against it (Los Angeles Times). Terrapin generalizes the phenomenon: "$50M+ completed shells waiting on energization" (Terrapin).

The operator response has been to convert balance-sheet strength into supply-chain priority. Equinix is now pre-purchasing mechanical and electrical equipment with its own balance sheet, gating new projects on power/energization/permitting, and making "advance manufacturing purchases" for deliveries in 2028 and 2029, while holding 3 GW of land against only about 700 MW under construction (Equinix Q2 2026 transcript; MarketBeat). The 3-GW-to-700-MW ratio is the sector's core mismatch in a single statistic: land is not the constraint, and neither is demand. Equinix did manage to accelerate 7,000 cabinets from 2027 into Q4 2026 by managing project risk and scheduling closely, evidence that schedule advantage is now a procurement competency, not a construction one (Equinix).

Switch and Digital Realty took the other route: locking dedicated factory capacity. Schneider Electric's $1.9 billion Switch agreement and $373 million Digital Realty agreement together represent roughly $2.27 billion of pre-allocated US power and cooling equipment, with Digital Realty's including a dedicated switchgear line explicitly to navigate bottlenecks (Data Center Frontier). Data Center Frontier's read, that this leaves other buyers competing for what remains, is the sector's central second-order effect.

C. Neoclouds

Narrative. The neocloud cohort has the most leveraged exposure to supply-chain slippage, because its revenue is contractually tied to delivery dates it does not control. CoreWeave is the archetype: a third-party developer's construction delay converted directly into a guidance cut and litigation alleging a $14 billion market-cap loss (GlobeNewswire). Crusoe demonstrates the reverse exposure, a 1.8 GW project paused not by equipment but by a customer's change of mind, after Nvidia had already put a $150 million deposit into the relationship (DCD; DCD).

The financing structures that have emerged are direct responses to this counterparty fragility. Google backstopping $1.8 billion of Fluidstack's lease obligations to TeraWulf, and taking roughly 8% of TeraWulf's equity in warrants for doing so, is a hyperscaler credit-wrapping a startup tenant so that a miner-turned-operator can raise project debt (DCD). Google did the same at Cipher, backing the Fluidstack lease and agreeing to assume it if the startup fails, plus taking a stake (DCD). Applied Digital shows the debt-market version: $2.15 billion of 6.750% secured notes issued at 98 and $1.59 billion of 7.000% notes at par, funding specific building-level capacity against 15-year take-or-pay leases worth $7.5 billion, $7.5 billion, and $5.2 billion respectively from a single investment-grade hyperscaler (Applied Digital). Coupons in the high-6% to 7% range on secured, contracted, investment-grade-anchored paper indicate how the market prices execution risk in this segment.

D. Crypto-to-AI Conversions

Narrative. The crypto-to-AI cohort has become the marginal supplier of energized megawatts in North America, and its disclosures are consequently the best available window into what power access is worth. TeraWulf's assessment, that "access to power had become the primary constraint across the industry" and that utilities are increasingly favoring experienced, well-capitalized, creditworthy counterparties, is the sector's thesis stated by a participant (TeraWulf). The corollary is that these firms' most valuable asset is a pre-existing interconnection, which is precisely why TeraWulf acquired Hawesville, Kentucky for its 480 MW of immediate grid-connected power and is pursuing Chesapeake Data's ~210 MW expandable to 1 GW (TeraWulf).

Contract values per megawatt are extraordinary and reveal the scarcity premium: Riot's 191 MW to Anthropic at $9.1 billion base and up to $16.1 billion (CNBC); Cipher's 300 MW to AWS at ~$5.5 billion and 168 MW to Fluidstack at $3 billion (DCD); Hut 8's 949 MW at ~$26.6 billion of base-term value with >$1.75 billion expected average annual NOI and $7.5 billion of investment-grade project financing already secured (Hut 8); Applied Digital's 1,410 MW at ~$36 billion (Applied Digital).

But the same disclosures show the constraint reasserting itself inside the contracts. Hut 8 explicitly excludes 1,000 MW of River Bend expansion capacity as "subject to the expansion of power at the site" (Hut 8). MARA's 2 GW Matagorda County site is conditioned on ERCOT and interconnection approvals, while its hosting agreements begin expiring in Q3 2027, a timing squeeze between contract roll-off and interconnection (MarketBeat). Cipher's Colchis 1 GW carries a fully executed AEP direct-connect agreement but a 2028 energization target (DCD). IREN's AI cloud revenue ramp slipped through calendar 2026 with analyst estimate cuts following (Investing.com), and Hut 8's Q1 2026 progress was reported explicitly "amid losses" (Investing.com).

E. Private Equity and Infrastructure Owners

Narrative. The private-capital side of this market has no capital problem, and that is precisely the source of the distortion. Investment Magazine's framing of the AirTrunk trade, Macquarie's $3 billion 2020 entry becoming a $24 billion enterprise value on exit to Blackstone and CPPIB, against a roughly $110 billion expansion pipeline, captures both the return available and the scale of committed follow-on capital (Investment Magazine). Aligned at $40 billion, acquired by AIP, MGX, and BlackRock's GIP, set a new record and closed in July 2026, on a platform that grew from two locations to fifty in seven years (DCD).

Blue Owl's role is the most instructive because it appears on both sides of the ledger. It announced more than $50 billion of data center investment in September–October 2025 alone, including $30 billion for Meta's Louisiana campus and more than $20 billion with Oracle in New Mexico, and had 1,000 people at Stack doing design, build, and operations (Los Angeles Times), and then declined to fund an additional Oracle facility by March 2026 (CNBC). The abundance of capital is conditional, and the condition being applied is asset-level durability against a one-year GPU cycle and multi-year power delivery. SoftBank's abandoned ~$50 billion Switch acquisition points the same direction (DCD).

F. Equipment Suppliers

Narrative. The supplier record contains a pattern that is easy to miss: record backlogs and revenue misses are occurring in the same companies at the same time. Fluence booked a record $6.4 billion backlog and cut FY2026 revenue guidance by $400 million in the same release, because its own new factories were delayed, one of them running on generators while it awaited a grid connection (MGrid; Fluence transcript). Vertiv, with a $15 billion backlog and 252% order growth, reported "minor timing shifts" from "temporary supply chain dynamics" and carried inventory that had risen from $1.457 billion to $2.523 billion (24/7 Wall St; Vertiv transcript). Supermicro shifted roughly $1.5 billion of revenue between quarters over GPU-rack configuration changes while holding a $13 billion GB300 backlog (Reuters). Wärtsilä missed order forecasts while describing data center prospects as significant (Reuters). The bottleneck has moved inside the suppliers themselves.

FuelCell Energy is the counterexample that proves demand alone is insufficient: a 267% pipeline increase to 4 GW alongside a 9.9% backlog decline and a product backlog that fell from $98.2 million to $36.1 million (FuelCell Energy). Pipeline is not backlog, and in a market where buyers demand delivery certainty, only firms that can commit dates convert.

The most consequential supplier behavior is deliberate demand rationing. HPE said outright that it does not have enough supply and is prioritizing higher-margin orders even at the cost of AI systems revenue growth (Reuters). Nvidia bought $4 billion of laser supply (Lumentum and Coherent) to secure EMLs, pushing rivals past 2027 (TechTimes). Prysmian is signing take-or-pay contracts with guaranteed margins because its European and North American grid capacity is full (Prysmian transcript). Suppliers now select customers.

G. Projects Delayed, Downsized, or Canceled

Narrative. Read together, the project record shows very few genuine cancellations for lack of demand and a great many pauses, re-phasings, and permit-driven stops. Meta's Hyperion moved in the opposite direction entirely, from 2 GW and $10 billion at its December 2024 groundbreaking to 5 GW and more than $50 billion by July 2026, with 2 GW targeted by 2030 and full buildout around 2032, financed in part through a $27 billion Blue Owl joint venture (CNBC). Microsoft's Wisconsin pause coexisted with a $7.3 billion Fairwater announcement in the same state (WPR; DCD).

The xAI Memphis and Southaven episodes are the clearest illustration of behind-the-meter generation outrunning its regulatory envelope. The Memphis facility ran more than 30 gas turbines for over a year before receiving a July 2025 permit; Colossus 2 in Southaven ran roughly 20 unpermitted turbines before Mississippi regulators issued a permit for 41 turbines totaling 1.2 GW, with the Southern Environmental Law Center and NAACP threatening or pursuing litigation at both sites and community comments alleging the draft permit understated emissions (E&E News). Turbine supplier Solaris Energy Infrastructure told CNBC that pollution controls were not included in the "temporary" turbines, though xAI had told Memphis authorities they would include selective catalytic reduction (CNBC). This is the compliance cost of the speed-to-power workaround.

Political risk is now a supply-chain variable in its own right. New York imposed the first statewide moratorium on new hyperscale data centers in July 2026, pausing state environmental permits for up to a year for applications not already deemed complete, and Caterpillar's own analysts flagged that if other states follow, portions of the pipeline underpinning 2027 and 2028 equipment orders could shrink or slip (The Motley Fool). That is the first documented link running backward from siting policy to OEM backlog quality.

H. Grid and Utility Delays

Narrative. The grid data is the report's most quantitatively decisive. PJM's 2026/27 capacity price of $329.17/MW-day against $28.92 in 2024/25 is an 11x escalation in two years, and data centers accounted for $6.5 billion, 40%, of the $16.4 billion in December 2025 auction costs, with roughly $6.2 billion of that attributable to data centers not yet constructed but projected in service by the 2027/28 delivery year (MGrid). Ratepayers are paying today for load that does not yet exist, which is why siting politics have turned.

The 2027/28 auction produced PJM's first system-wide reliability shortfall, 6,623 MW short of requirement at a 14.8% reserve margin against a 20% standard, with nearly 5,100 MW of the year-over-year peak forecast increase attributable to data centers (MGrid). The arithmetic behind that outcome is simple and durable: 5–7 GW of annual data center load additions through 2032 against 2–3 GW of annual supply additions, with utilities having doubled their large-load forecasts to 12–14 GW per year (MGrid).

Dominion's disclosure is the sharpest single constraint statement in the report: roughly 70 GW in queue requiring new generation before connection, capacity to connect approximately ten large-load customers per year, and an average data center grid-connection wait of seven years (Virginia Mercury). Ten customers per year against 70 GW is a rationing problem, not a queueing problem, and it explains the entire behind-the-meter turbine phenomenon. Virginia Mercury also documents the regulatory arbitrage that follows: a data center may place 25 MW of gas generation across four or five separate turbines, with individual units potentially remaining below the stationary-source threshold, while units above 25 MW must participate in RGGI's carbon-allowance auction (Virginia Mercury).

Europe is worse on the interconnection dimension and better on nothing. AWS's EMEA energy regulation lead disclosed that securing power can take up to seven years against roughly two years to build the facility, with IEA figures putting Frankfurt, London, Amsterdam, Paris, and Dublin connection waits as high as ten years (Reuters; The Register). Speculative-project backlogs under first-come-first-served rules are the specific mechanism in Italy and Spain (Reuters). Amazon's response, an internal plan to cut its Dublin e-commerce footprint 40%, exit Dublin entirely by end-2026, and leave Northern Virginia and Oregon AWS regions by 2029, at roughly $90 million of one-time cost and 10–15% higher hosting costs for some migrated services, is the first documented case of a hyperscaler relocating existing production workloads because of power scarcity rather than cost or latency (Business Insider). Even so, Zaragoza's own capacity limits meant only 65% of the remaining Dublin footprint could move there (Business Insider).

One reliability data point deserves separate emphasis because it changes the regulatory argument: in May 2025, a single remote fault in Ireland caused 387 MW of data center load, 52% of all data center demand, to drop simultaneously, because UPS systems switch to battery instantly on transient faults, producing collective demand drops of hundreds of megawatts within milliseconds (MGrid). This makes data centers a grid-stability concern independent of their energy consumption, and it strengthens the case utilities make for interruptible tariffs and curtailment authority.

Key Themes and Cross-Cutting Bottlenecks

1. Transformers, gas turbines, and grid interconnection are the three tightest constraints, and they are sequentially dependent. Nothing else in the stack matters if any one of them fails. Transformers: a 30% supply deficit, GSU lead times above 160 weeks versus a 143-week 2024 average, and custom units at 3–5 years, against roughly 80% import dependence and a single domestic GOES producer (Wood Mackenzie; Reuters; IndustrialSage). Gas turbines: GE Vernova booking into 2031 with roughly 10 GW left across 2029–2030 and pricing up 10–20 points per kW; Mitsubishi at five-year deliveries with order books full through 2030 and contracting 2031–2034 (Power Engineering; S&P Global). Interconnection: Dominion connecting ~10 large-load customers per year against ~70 GW in queue, seven-year average waits, and PJM's 1,200 GW queue averaging five years (Virginia Mercury; MGrid). Every other category in Part 1, including GPUs, is downstream of these three.

2. Power-first project structuring has replaced site-first structuring. Equinix now gates new projects on power, energization, and permitting, pre-purchases mechanical and electrical equipment with its own balance sheet, and makes advance manufacturing purchases for 2028–2029 deliveries, while holding 3 GW of land and building only ~700 MW (Equinix; MarketBeat). Hut 8 describes a "power-first execution model" and excludes 1,000 MW of expansion from its contracted figures because it is subject to site power expansion (Hut 8). TeraWulf buys sites for their existing interconnections, 480 MW at Hawesville, rather than their land (TeraWulf). Roseville Electric procures equipment five years ahead (Reuters). Procurement has been promoted from a construction function to a strategic one; the 10–20% switchgear deposit securing a production slot is its most literal expression (Electronate).

3. Behind-the-meter gas is the dominant workaround, and it is now measured in gigawatts. Documented commitments include INNIO–VoltaGrid at 2.3 GW, Caterpillar–Joule at 4 GW, Caterpillar–Hunt at 1 GW, 2 GW of CAT G3516 for AIP Monarch, Wärtsilä at 412 MW in Ohio within a >1.6 GW US pipeline, PROENERGY's 650 MW for Crusoe, INNIO–Rehlko's 1.25 GW framework, and Nscale's 2 GW Monarch project (Power Magazine; Power Engineering). Powell's largest-ever order, over $400 million, gigawatt-scale phase one, came from a neocloud building behind the meter (Powell transcript). Fluence signed its first large behind-the-meter order in company history in the June 2026 quarter, a $300 million developer project that converted from lead to contract in under three months, versus a 12–18-month norm (Fluence transcript). The workaround carries its own costs: xAI's unpermitted turbine operations at two sites, litigation from SELC and the NAACP, and Virginia legislation targeting behind-the-meter gas (E&E News; Virginia Mercury). And it does not escape the turbine queue, it merely moves from utility-scale frames to reciprocating engines and aeroderivatives, which are themselves now being allocated into 2028 (Manufacturing Magazine).

4. Capital is abundant and equipment is scarce, and the mismatch produces stranded assets, not faster delivery. The capital side shows $40 billion for Aligned, $24 billion enterprise value for AirTrunk against a $3 billion 2020 basis, more than $50 billion of Blue Owl commitments in two months, a $27 billion Meta JV, Applied Digital raising $3.74 billion of secured notes in two tranches, and Hut 8 securing $7.5 billion of investment-grade project financing (DCD; Investment Magazine; Los Angeles Times; CNBC; Applied Digital; Hut 8). The equipment side shows two finished Santa Clara buildings sitting empty pending a $450 million utility upgrade that completes in 2028 (DCD) and, generically, "$50M+ completed shells waiting on energization" (Terrapin). Money cannot compress a 160-week transformer lead time; it can only reprice the queue position, which is what take-or-pay contracts with guaranteed margins (Prysmian), $2.27 billion of Schneider Supply Capacity Agreements (Data Center Frontier), and Nvidia's $4 billion laser lockup (TechTimes) all accomplish.

5. Pre-allocation by the largest buyers has become the defining market structure, and it externalizes scarcity onto everyone else. Meta anchoring Corning's Hickory plant at up to $6 billion; Nvidia funding a 10x expansion of Corning's US optical connectivity capacity; Amazon's multi-billion North Carolina fiber deal; Switch's $1.9 billion and Digital Realty's $373 million Schneider agreements including a dedicated switchgear line; Prysmian's €5.5 billion Molex contract with a €550 million down payment (Corning/Meta; Corning/Nvidia; Tech Blog; Data Center Frontier; Prysmian). The documented consequences: standard commercial and industrial switchgear orders competing for slots left over after data center allocations (Electronate); utilities and large loads queueing behind data center customers for Fluence production slots (MGrid); non-Nvidia optics buyers pushed past 2027 (TechTimes); small fiber buyers at one-year lead times versus 20 weeks for large ones (Tom's Hardware); memory customers allocated 60–70% of requested volumes (Seeking Alpha). US broadband deployment targets are among the collateral casualties (Light Reading).

6. Suppliers' own factories are now the bottleneck, which breaks the assumption that capacity announcements resolve shortages. Fluence's Houston plant, 15 GWh of intended capacity, was delayed by construction and automation problems and ran on generators while awaiting its own grid connection, contributing to a $400 million guidance cut (Fluence transcript). Vertiv cited "temporary supply chain dynamics" and near-doubled inventory (Vertiv transcript). Arista reported shortages across five distinct input categories simultaneously with qualification cycles tripling to 6–8 quarters (MarketBeat). Electronate's judgment that reshoring investment will not produce meaningful capacity before 2027–2028, and that announced expansions take 2–3 years to yield output, should be read against the $1.8–2.0 billion of announced North American transformer expansion (Electronate; Wood Mackenzie).

7. Memory has become the cost shock of 2026, displacing GPUs as the marginal price driver. Server DRAM rising ~90% quarter over quarter in 1Q26, the largest quarterly increase on record, plus NAND at +55–60% and enterprise SSD at +53–58% (TrendForce) shows up directly in hyperscaler capex: Amazon at +$20 billion primarily on memory (Data Center Knowledge) and Microsoft at +$25 billion of component pricing (DCD). OEMs are responding with contractual price adjustability: Dell repricing where necessary (Futurum) and HPE shortening quoting cycles while retaining the ability to reprice between order and shipment (Reuters). Fixed-price server procurement is effectively over.

8. Compressed silicon cycles are creating obsolescence risk inside long-lead procurement. Nvidia moved to annual data center processor releases; Vera Rubin was in production by January 2026 at five times Blackwell inference performance; and standing up a site takes at least 12–24 months (CNBC). Abilene is the case study: hardware ordered, billions spent, power a year away, and the anchor tenant declining to expand because it wanted newer silicon (Reuters; CNBC). Microsoft's stated answer is to buy and replace annually rather than standardize on one generation (DCD). For debt-financed developers with 15-year take-or-pay leases underwriting 6.75–7.0% secured notes (Applied Digital), the mismatch between a one-year silicon cycle, a two-to-four-year build cycle, a seven-year interconnection cycle, and a fifteen-year lease is the central unresolved structural risk in this market.

9. Siting politics has become a supply-chain input. New York's statewide hyperscale moratorium in July 2026 pauses state environmental permits for up to a year, with analysts flagging that broader adoption could shrink or slip the 2027–2028 project pipeline underpinning OEM order books (The Motley Fool). PJM's disclosure that data centers accounted for $6.5 billion, or 40%, of December 2025 auction costs, roughly $6.2 billion of it for facilities not yet built (MGrid), and reporting that monthly electricity bills near new facilities ran 267% higher than five years earlier (CNBC) explain why. Virginia's new $1.2 billion biennial data center tax, the Data Center Coalition's SCC challenge to interconnection-queue transparency, and FERC's directive that RTOs justify or reform large-load interconnection rules (Virginia Mercury) indicate the regulatory response is still being written. Any 2027–2029 equipment backlog should be discounted for the possibility that its underlying projects lose their permits.

Methodology and Limitations

Every value in this report is linked to a page retrieved during this research session; approximately 160 URLs were retrieved across three rounds, of which the citations above represent the subset that yielded usable, attributable figures. A material number of high-priority sources, Bloomberg, WSJ, and several regulatory and IR pages, including Cyxtera bankruptcy filings, AEP Ohio's PUCO tariff order, EirGrid's Dublin connection policy, Hillsboro's moratorium ordinance, Georgia Power's large-load tariff, and QTS/CyrusOne/DigitalBridge financing disclosures, returned no extractable content and are therefore marked "n.a." rather than paraphrased from memory. Several Part 1 categories (structured cabling MPO/MTP, piping and valves, raised floor, ATS/STS vendor-level allocation, busway vendor-level allocation) lack primary-source lead-time or allocation data and are explicitly marked as such. Where two fetched sources give differing figures for the same category, as with busway and MV switchgear lead times, both ranges are shown with their respective sources rather than averaged.