Alphabet, Amazon, Meta and Microsoft alone are expected to spend over US$650 billion in 2026, yet close to half of this year’s planned US data centre builds have been delayed or cancelled. Each month of delay on a 60 MW facility costs roughly US$14.2 million. The blockers sit in the physical world: electrical gear with multi-year lead times, a skilled-trades shortfall, and organised supply-chain theft. This article answers four questions: how long operators wait on gear, why labour is the second bottleneck after power, what drove theft losses up 60% to US$725 million in 2025, and whether modular is faster and cheaper in 2026. By the end you will be able to read a delivery date the way operators do, and know what modular construction can and cannot buy you in 2026. It is the mechanism behind why so much planned capacity is delayed or cancelled, and part of the collateral damage of the AI data centre boom.
Why is there a transformer and electrical equipment shortage, and how long are operators waiting on gear now?
Waits now run in years, not months, because two decades of flat grid investment are colliding with AI demand. Wood Mackenzie puts waits on key components at 18 to 36 months, large power transformers at 80 to 128 weeks, big custom units at three to five years. Operators wait years, and energisation dates slip even when the building itself is finished.
The scarcity covers the electrical room’s whole bill of materials: power and substation transformers, medium-voltage switchgear, breakers, high-voltage cable. Transformer demand is up 119% since 2019, grain-oriented electrical steel is tight, and there are fewer than six major producers worldwide. Wood Mackenzie’s grid-equipment data shows the US market growing from roughly US$20 billion toward US$65 billion, and vendors returning with year-old purchase orders to impose 20% price increases just to hold delivery slots. Hyperscalers pay, because a missed launch costs more.
As waits stretch past two years, Washington now treats the queue as a national-security problem: in April 2026 the President invoked Section 303 of the Defense Production Act, classifying grid infrastructure as essential to national defence and authorising emergency financing for domestic supply (full text). GE Vernova’s equipment backlog was up 69% year on year inside a US$176 billion total, and Mitsubishi Electric will not finish doubling transformer output until 2030.
NVIDIA is pushing racks toward 1 MW each, and solid-state transformers using silicon-carbide power electronics convert medium-voltage AC straight to 800V DC. In 2026 these are pilot-stage products from vendors like Heron Power and DG Matrix, a medium-term answer at best.
Why is skilled labour the data centre industry’s second bottleneck after power?
Power decides whether a site is viable; trades decide whether it is delivered on time. Electricians, high-voltage specialists, welders and commissioning crews are being bid away by fabs, renewables and grid work. US construction was 439,000 workers short in late 2025, and CBRE data shows a 32% pay premium for data centre work. Even when power and equipment line up, schedules still stretch.
Interconnection queues of five to seven years gate viability first, so trades capacity only governs delivery once power is secured. Mercer‘s William Self calls labour the single biggest constraint of this buildout and puts the shortfall at 75,000 to 140,000 skilled workers. Peak crews grew from about 750 in the cloud era to 4,000 to 5,000, while roughly 300,000 new electricians are needed this decade on top of 200,000 retirements. Turner & Townsend‘s cost index puts power access as the top scheduling constraint, ahead of labour.
Data centre electricians in Northern Virginia and Texas report US$140,000 to US$280,000, and 60% of providers struggle to fill roles. Abilene, where the Stargate campus has already launched 1.2 GW, is reorganising around a “functioning mini-economy” of housing, hospitality and healthcare (DataCenterKnowledge). The industry’s answer is to build its own pipelines: apprenticeships, community college partnerships, veteran pipelines and in-house academies, plus Meta’s training programme with CBRE (Tom’s Hardware).
Australia draws on the same global trades pool with far thinner data. Specialist labour shortages are expected to intensify, and AEMO’s planning data shows data centre electricity demand climbing steeply. Australia’s labour-market evidence lags the US, so local schedule risk is harder to quantify. This is how these bottlenecks feed the wider crunch.
Why did data centre supply chain crime losses surge 60% to US$725 million in 2025?
The same queues now attract a different kind of pressure: organised crime. Cargo theft losses in the US and Canada hit nearly US$725 million in 2025, up 60% on 2024, with average stolen loads up 36% to US$273,990, per an FBI alert and Verisk CargoNet. The driver is organised rings using ghost carriers, fraudulent pickups and cyber-enabled freight fraud. Electronics made up 22% of incidents, per Overhaul.
Copper is the recurring magnet: transformer coils, cabling and copper conductor, alongside electrical equipment, servers and GPUs in transit. The pattern runs through logistics corridors, truck yards and construction sites, which thieves now target directly; Cook County investigators recovered US$1.3 million in stolen copper wire and data centre equipment from two trailers.
Overhaul’s David Warrick describes syndicates that run like big corporations, penetrated at driver and warehouse level, moving goods toward China, Russia and Iran, where black-market RTX 6000 Pro prices more than doubled. Every stolen shipment becomes a replacement order that rejoins the same queues, and disruption can cost several times the goods’ value.
Is modular data centre construction actually faster and cheaper in 2026?
With schedules stretched by gear, trades and theft, factory-built construction is the industry’s main answer, and the question is what it actually buys. For standardisable scope, yes on speed: factory-built power skids, modular substations and prefabricated data halls cut on-site schedules, with highly modularised projects reporting 30 to 50% compression. On cost the verdict is mixed, and a power skid still needs the same scarce transformers and switchgear.
More of the build moves into factories, assembled and tested off site, cutting on-site labour and weather exposure. Inflect’s cost comparison, benchmarked against CBRE data, puts traditional builds near US$10 to 12 million per MW against an industry-cited US$7 to 9 million for modular, though transport, cranes and staging eat into it. Meta’s US$10 billion Louisiana campus shows the hyperscaler version, with electrical rooms prefabricated years ahead; materials suppliers such as Saint-Gobain, which feeds hundreds of projects, sit behind that chain (Tom’s Hardware).
Where it fits is narrower than the marketing suggests, and that is the part to weigh if you are comparing build methods. Secondary sites and standardised halls gain most; the highest-density liquid-cooled AI halls, with bespoke power chains, remain custom builds. Sequencing, transport logistics and standards have become a management discipline of their own.
Taken together, the three problems snap into a loop: scarce gear sets the pace, scarce trades stretch whatever the gear allows, and every theft adds a replacement order to the same queue. Nothing in the loop eases until manufacturing and workforce pipelines mature later this decade. Higher component and hardware costs are the working assumption for your planning; see planning for higher component and hardware costs. See the bigger picture of the buildout’s fallout.
Frequently Asked Questions
When will transformer lead times for data centres come down?
Not before manufacturing capacity matures later this decade. Backlogs are still growing: GE Vernova’s equipment backlog rose 69% year on year inside a US$176 billion total, and Mitsubishi Electric will not finish doubling transformer output until 2030. For planning, assume waits of 44 to 65 weeks for medium-voltage switchgear, 80 to 128 weeks for transformers, and 3 to 5 years for large custom units.
Can operators pay a premium to jump the transformer and switchgear queue?
Premiums buy a place in line, not faster production. Hyperscalers already pay above list to hold delivery slots, and Wood Mackenzie reports vendors repricing year-old orders roughly 20% upward, yet waits still run years because manufacturing lines, not order books, set the ceiling. Earlier commitment is still the strongest lever, which is why power-first siting and slot locking now shape delivery dates.
What is grain-oriented electrical steel, and why does it constrain transformer supply?
Grain-oriented electrical steel, or GOES, is the specialist steel used in transformer cores, where its magnetic properties reduce energy losses. Its supply is constrained just as transformer demand has risen 119% since 2019, and the manufacturing base is thin, with fewer than six major large-transformer producers worldwide. That combination is why expanding output takes years and the bottleneck starts upstream of the assembly line.
What is the difference between solid-state and conventional transformers?
A conventional transformer uses iron and copper to step voltage up or down, and takes years to build. A solid-state transformer uses silicon-carbide power electronics to convert medium-voltage AC directly to 800V DC, which is the direction AI rack power is heading as racks approach 1 MW each. In 2026 these units remain at pilot-deployment stage, so they are a medium-term answer, not a fix.
Why did the US invoke the Defense Production Act for large-scale grid infrastructure?
The US now treats the equipment crunch as a national-security problem, not just a market one. The April 2026 Defense Production Act Section 303 determination classified large-scale grid infrastructure as essential to national defence, authorising emergency federal financing to expand domestic supply. It signals that grid gear carries federal priority, though it expands supply over years rather than clearing existing transformer and switchgear queues overnight.
Where can I read the full text of the April 2026 presidential determination?
The full text sits on whitehouse.gov and is the primary source for the April 2026 determination. It sets out the Section 303 classification of large-scale grid infrastructure as essential to national defence, along with the emergency financing authorised to expand domestic supply. Reading it directly shows the scope of the authorisation, which matters when judging how far federal support can reach into the equipment bottleneck.
Where can I find transformer lead time and grid equipment market data?
Wood Mackenzie publishes the lead-time and market data behind these constraints: critical-component waits of 18 to 36 months, transformers averaging 80 to 128 weeks, and a US data centre electrical equipment market growing from roughly US$20 billion toward US$65 billion. Its grid equipment research is the benchmark to cite when checking whether a promised delivery date is realistic.
Is higher pay enough to fix the data centre electrician shortage?
No. Pay is how operators compete for a fixed pool of trades, not how the pool grows. CBRE records a 32% premium for data centre work, with electricians in Northern Virginia and Texas earning US$140,000 to US$280,000, yet 60% of providers still report difficulty filling roles. The data centre workforce averages 53 years old, and roughly 200,000 electricians are expected to retire this decade against about 300,000 needed.
Is data centre cargo theft mostly opportunistic crime?
No, it is overwhelmingly organised. Rings use ghost carriers, fraudulent pickups and cyber-enabled freight fraud to target copper, electrical gear and GPU shipments in transit, a pattern serious enough that the FBI issued a public service announcement. US supply chain crime losses against data centre cargo hit US$725 million in 2025 (Verisk CargoNet), up roughly 60%, and electronics accounted for 22% of incidents (Overhaul).
How is prefabrication being used on large AI campuses today?
Adoption is real but selective. Meta’s US$10 billion Louisiana campus is the clearest hyperscaler example, and suppliers such as Saint-Gobain are building out prefabricated systems, including work with Microsoft. Factory-built power skids, modular substations and prefabricated halls cut on-site schedules where scope is standardisable. The highest-density liquid-cooled AI halls, with bespoke power chains, remain largely custom builds for now.
Does Australia face the same equipment and labour crunch?
For equipment, largely yes: Australia draws from the same constrained global transformer and switchgear market, so extended waits and premiums carry across. On labour, specialist shortages are expected to intensify, and AEMO expects data centre electricity demand to treble within four years. The caveat is the evidence gap: there is far less granular data in Australia than in the US on these pressures, so local schedule risk is harder to quantify.
What can operators do to protect their data centre build schedules?
The emerging playbook has three parts: lock long-lead gear early, because early commitment secures manufacturing slots; build trades supply through partnerships, apprenticeships and in-house academies, since experienced crews take years to produce; and standardise designs so prefabrication can absorb on-site work where scope allows. None of these removes the constraints. Together they buy schedule protection while manufacturing capacity matures.
