Powering the AI Factory: The Electrical-Equipment Bottleneck of 2026

In 2026 the binding constraint on AI-factory deployment is no longer silicon. It is power, and specifically the electrical equipment that converts and delivers it. Transformers sit at the epicenter of the squeeze.

Average lead times have stretched to 128 weeks for power transformers and 144 weeks for generator step-up (GSU) units, per Wood Mackenzie's Q2 2025 survey,¹ with prices up 45–95% since 2019 and a 30% supply shortfall for power transformers. Roughly 60% of global transformer production capacity sits in China,¹ while North American manufacturers have committed ~$1.8 billion in new capacity since 2023; first output from the largest plants arrives only in 2027–2028.¹ Omdia projects 30–50% of planned 2026 U.S. datacenter capacity will slip to 2028, with interconnection queues exceeding 2,100 GW.⁵

The architectural response is a shift to 800 VDC distribution, anchored by NVIDIA's Kyber rack and Vertiv's H2-2026 portfolio, which cuts copper and conversion losses.^8,9 Solid-state transformers (SSTs) are emerging as a longer-horizon mitigation: DG Matrix raised a $60M Series A with ABB as a strategic investor, and Delta demonstrated 800 VDC power at COMPUTEX 2026.^10,11,12 But SSTs remain at demonstrator / early-pilot stage in 2026: a credible 2027+ bet, not a 2026 supply fix. The constraint has moved downstream from silicon to electrons.

The squeeze is demand-driven. Global datacenter power demand reached ~415 TWh in 2024 and is projected to roughly double to ~945 TWh by 2030.⁴ U.S. hyperscale demand is on a path from ~50 GW to over 134 GW by 2030, with datacenters growing ~15% per year, more than four times the rest of the economy, and rising from ~4.4% of U.S. electricity in 2023 to a projected 6.7–12% by 2028.⁴

Capital is not the limit: hyperscalers have committed roughly $660 billion in capex but have, in Omdia's words, "hit a wall" on electricity availability.⁵ Of approximately 12 GW of AI datacenter capacity announced for 2026, only ~5 GW is under active construction; nearly half of planned 2026 U.S. datacenters have been delayed or canceled, with transformers, switchgear and batteries the recurring cause.^2,5 Grid-connection timelines now run 3–7 years.⁵

The centerpiece of the 2026 supply crunch. Transformers are long-lead, materials-intensive, geographically concentrated, and demanded simultaneously by datacenters, utilities, electrification and renewables, all pulling from the same constrained base.

Lead times have detached from historical norms

Demand has surged faster than any plausible supply response

U.S. demand growth from 2019–2025, by equipment class:¹

Bars scaled to the +274% peak. *Manufacturing construction spending growth measured over three years, not 2019–25. All figures per Powermag / Wood Mackenzie.¹

Pricing has reset structurally higher

Structural causes: materials, concentration, and inflexible capacity

• A 30% supply shortfall for power transformers and 10% for distribution units in 2025 (Wood Mackenzie); Western supply gaps run up to 30%.^1,4
• ~60% of global transformer production capacity sits in China, leaving Western buyers facing 2–4 year waits versus shorter Chinese delivery cycles, creating geopolitical and tariff exposure.¹
• Grain-oriented electrical steel is a single-point input: Cleveland-Cliffs is the only domestic U.S. GOES producer, and prices have roughly doubled since 2020.¹
• Aging base compounds demand: more than half of U.S. distribution transformers, roughly 40 million units, are already beyond their expected service life.¹
• Tariffs raise the floor: copper faces duties "up to 50%," and Section 232 steel/aluminum duties have been expanded.¹
• Capacity is slow to add: large power transformers require specialized steel, copper, engineering, testing and factory capacity; "new capacity takes years, not months."³ Transformers are 20–30% of datacenter electrical CAPEX and now drive site selection, schedule underwriting, pre-entitlement deposits and lease terms.^3,4

The North American capacity response (~$1.8B since 2023)

The signal is unambiguous, but the timing is the problem. The largest plants do not produce until 2027–2028, so the 2026 shortage is effectively baked in regardless of how much capital is now flowing into expansion.

Transformers are the epicenter, but the constraint is system-wide. Switchgear runs 52–78 weeks; U.S. datacenter equipment lead times averaged 42 weeks in 2026, 83% above the 2019 baseline (33 weeks globally, 50% above pre-2020).^3,6

Backup & on-site generation is booked out

• Rolls-Royce is booking 2027–2028 slots for datacenter power generation; datacenters now exceed 80% of its Power Systems generation revenue.⁶
• Caterpillar holds a long-term agreement with Hunt Energy for datacenter power systems.⁶
• Wärtsilä announced a 34SG 412 MW-nameplate project in April 2026 (contracted, not yet operational).⁶
• Bloom Energy & Equinix are deploying 100 MW of fuel cells (75 MW operational, 30 MW under construction) across 19 sites in six states.⁶
• Diesel remains the default emergency backup, but gas reciprocating engines, solid-oxide fuel cells and behind-the-meter gas+battery hybrids are moving from backup toward primary power.⁶

The silicon picture, for context: the bottleneck that moved

In 2025 the gating items were advanced packaging and HBM, not logic dies; the four largest AI chip designers (NVIDIA, Google, AMD, Amazon) consumed ~90% of global CoWoS capacity and HBM supply while using only 12% of advanced logic die production.⁷ That bottleneck is now easing as packaging scales, even as power tightens:

HBM market grew from $18B (2024) to $35B (2025), with prices up 20–30% YoY; NVIDIA H100/H200 lead times ran 36–52 weeks.⁷ Advanced logic (3–5nm) supply was ~3.2–3.5M wafers in 2025, abundant relative to packaging. The lesson: as one constraint is funded away, the system finds the next one. In 2026 that is electrical.

Rising rack density is forcing a rethink of the power chain itself. The Hopper-to-Blackwell transition brought a 3.4× increase in rack power density, pushing racks from tens of kW to well over 100 kW, with "a megawatt per rack on the horizon."⁸ At those levels, traditional low-voltage delivery (54 VDC) is physically and economically impractical because the current required produces excessive resistive losses and an unsustainable volume of copper.⁸

The industry's answer is 800 VDC distribution. At 800 VDC the same wire gauge carries 157% more power than at 415 VAC, and a native-DC chain eliminates multiple AC-DC conversion steps where end-to-end efficiency can fall below 90%.⁸

• NVIDIA Kyber is the first rack architecture built on 800 VDC; it houses 576 Rubin Ultra GPUs, which begin shipping in 2027.^8,9
• Vertiv plans to release its 800 VDC power portfolio in H2 2026, timed to Kyber and Rubin Ultra.⁹
• The ecosystem is coordinating through the Open Compute Project (OCP), spanning silicon, power components and datacenter-systems vendors (ABB, Eaton, GE Vernova, Heron Power, Hitachi Energy, Mitsubishi Electric, Schneider, Siemens, Vertiv).⁸

This matters for supply chain because 800 VDC is precisely the architecture solid-state transformers are designed to feed, making the SST a strategic, not merely an incremental, technology.

The headline question: how mature is the SST segment in 2026? The evidence points to a technology inflecting out of the lab into first pilots: real commercialization signals, but no documented volume deployment, and meaningful supply-chain gaps still open.

What is demonstrated vs. what is funded

The enabling layer and remaining gaps

Market sizing: directional only

1. Make procurement the first move, not the last. Order transformers and switchgear at project inception, before entitlement, and structure deposits and lease terms around 2–3-year lead times. Site selection is now power-equipment-led.³
2. Treat "bring your own power" as a primary strategy. With interconnection queues above 2,100 GW and 3–7-year timelines, behind-the-meter generation (gas reciprocating, solid-oxide fuel cells, gas+battery hybrids) shifts from backup to primary, but OEM slots are booked into 2027–2028, so commit early.^5,6
3. Engage OEMs and SST developers directly. Secure allocation with transformer OEMs (Hitachi, Siemens, Eaton, Prolec GE) and co-engineer 800 VDC-ready equipment with SST players (DG Matrix, Delta) to hedge conventional lead times.^1,11,12
4. Design to 800 VDC now. Align builds to the 800 VDC bus, 1200V SiC, and liquid-cooled racks tracking Kyber/Rubin and Vertiv's H2-2026 portfolio; retrofitting later is costly.^8,9
5. Watch the SST inflection as a leading indicator. DG Matrix's pipeline, ABB's investment and Delta's demonstrations point to a 2027–2028 pilot-to-volume transition. Track commoditization of HV-SiC gate drivers and medium-frequency magnetics as the gating signals.^11,13