The AI Boom Is Becoming an Electricity Supply Chain Story

The bottleneck has moved outside the server rack

A modern AI campus is effectively two capital projects being built together. One is the computing system. The other is a miniature electrical and thermal infrastructure system that begins at the transmission network and continues through substations, large power transformers, medium-voltage switchgear, busways, uninterruptible power supplies, cooling equipment, backup generators, and finally the racks themselves.

The physical intensity of that second system is increasing. CBRE estimates that the most demanding U.S. data center developments now cost roughly $14 million to $16 million per megawatt to construct.[4] At the rack level, Nvidia's DGX GB200 NVL72 consumes approximately 120 kilowatts and uses liquid cooling, compared with the much lower rack densities historically common in enterprise data centers.[5] Higher rack density does not eliminate infrastructure. It concentrates the infrastructure requirement into less space and raises the performance requirements for electrical distribution and cooling equipment.

This distinction matters for valuation because the AI buildout can remain strong while the composition of spending changes. A dollar that might once have been modeled as incremental server demand can increasingly require associated spending on substations, electrical distribution, thermal systems, generation, storage, and transmission upgrades. The suppliers able to deliver those systems therefore participate in the AI investment cycle without needing to manufacture a semiconductor.

It also helps explain why the data center market can simultaneously report enormous construction pipelines and limited available capacity. CBRE found 7.48 gigawatts under construction across its primary North American markets in the first half of 2026, up 24.8% from a year earlier, yet vacancy still fell to 1.4%. More than 80% of that construction was already committed.[6] The scarce product is not an empty building. It is an energized building with a credible delivery date.

Grid connections and transformers are becoming the first gating items

The first major constraint sits upstream of the data center itself. A 100 megawatt or 500 megawatt campus cannot simply plug into an existing distribution circuit. Utilities must determine whether generation, transmission lines, substations, and local grid equipment can accommodate it, then determine who pays for whatever must be added.

The pressure has become large enough to change federal grid regulation. In June 2026, the Federal Energy Regulatory Commission ordered all six regional grid operators under its jurisdiction to justify or reform the rules governing the connection of data centers and other large electricity users. FERC's separate large-load proceeding generally focuses on loads above 20 megawatts and is examining issues including grid-upgrade costs, flexible service, project readiness, and whether generation and large loads should be studied together.[7] The existence of the proceeding is itself economically significant. Data center interconnection has moved from a utility engineering issue into a system-planning problem.

Even after a project receives a viable interconnection path, it must obtain the hardware. Reuters reported in July that lead times for some high-voltage transformers had reached about 160 weeks, compared with 143 weeks in 2024, as utilities and data center developers competed for transformers, circuit breakers, and switchgear.[8] Siemens Energy says approximately 80% of U.S. transformers are imported and that delivery times for some large transformers can extend as long as five years.[9]

Manufacturers are spending aggressively to expand capacity, but the timing is important. Hitachi Energy broke ground in June 2026 on a $457 million large power transformer factory in Virginia, part of more than $1 billion of planned U.S. grid-equipment investment. The new plant is intended to support transmission, generation, industrial customers, and data centers.[10] Capacity that starts production several years into an AI construction boom can eventually relieve shortages, but it does not retroactively accelerate projects that need equipment today.

This creates an increasingly important distinction between a data center development pipeline and a power-secured development pipeline. Land, permits, fiber access, and even a signed tenant are not equivalent to energized megawatts. For investors evaluating developers, utilities, and data center operators, the quality of a project pipeline should therefore depend partly on the maturity of its interconnection, substation, transformer, and equipment procurement rather than simply its announced megawatt count.

Gas turbines and backup generators are turning AI campuses into power projects

Grid delays are also changing the architecture of the data center itself. Developers that cannot obtain enough utility power quickly are considering onsite generation, microgrids, and temporary or bridge-to-grid systems. That shifts a portion of the AI investment cycle into natural gas turbines, reciprocating engines, fuel infrastructure, batteries, and the electrical controls needed to coordinate them.

GE Vernova provides unusually clear evidence of how quickly that market has tightened. At the end of the second quarter of 2026, its gas power equipment backlog and slot-reservation agreements totaled 116 gigawatts, up from 100 gigawatts three months earlier. The company expected to reach at least 125 gigawatts by year end. GE Vernova was producing gas turbines at an annualized rate of about 20 gigawatts in the third quarter of 2026 and plans to raise capacity to 30 gigawatts by 2030.[11]

The more revealing number may sit outside the turbine division. GE Vernova said data center orders in its Electrification segment exceeded $5 billion in the first half of 2026, more than double the total booked during all of 2025.[11] That means AI demand is appearing simultaneously in generation and in the equipment required to move and control the electricity after it has been generated.

Reciprocating engines are participating in the same shift. Cummins said robust orders for data center standby power contributed to record second-quarter 2026 results.[12] Caterpillar separately entered an agreement supporting 2 gigawatts of natural gas generator sets for a hyperscale power platform, with deliveries scheduled from September 2026 through August 2027.[13] Backup generation remains a separate function from continuous primary generation, but the boundaries are becoming more strategically important as campuses consider grid-connected, bridge-to-grid, and partially islanded designs.

The IEA still expects renewables to supply a large portion of incremental data center electricity demand over time. Natural gas, however, is expected to play a meaningful dispatchable role, particularly in the United States, while nuclear capacity contributes more materially later in the decade.[14] The near-term investment question is therefore not simply which energy source wins. It is which combination of generation can be permitted, financed, interconnected, and equipped quickly enough to energize a campus on schedule.

That favors technologies with different attributes at different stages. Solar and batteries can be deployed relatively quickly but may not independently provide the continuous power profile required by a hyperscale campus. Large combined-cycle plants can provide substantial firm capacity but require turbines, pipelines, permits, and grid integration. Reciprocating engines can be deployed modularly and may bridge a project toward a later utility connection. Nuclear offers very different operating economics but typically works on a substantially longer development schedule. The AI power buildout is therefore unlikely to produce a single generation winner. It is producing a premium on speed, availability, and integration.

Switchgear, copper, and cooling create a second layer of constraints

Once high-voltage power reaches the site, it still has to be safely distributed. That is why switchgear has become one of the less visible beneficiaries of AI infrastructure spending. Eaton's Electrical Americas business reported that its rolling twelve-month orders were up 41% organically in the second quarter of 2026, while backlog was 33% above the prior year.[15] The company has also announced a new 370,000-square-foot Nebraska facility specifically intended to expand medium-voltage switchgear production amid AI data center demand, with production expected to begin in 2027.[16]

Switchgear does not attract the attention of an accelerator launch because its job is comparatively mundane: protect, isolate, and control electrical equipment. Yet a campus cannot operate without it. The more electrical power concentrated into a facility, the more consequential those protection and distribution systems become. The investment implication is that apparently ordinary industrial equipment can acquire strategic value when its lead time becomes longer than the customer's tolerance for delay.

Copper runs through much of the same chain. It is required in transmission and distribution equipment, transformer windings, busbars, cabling, motors, generators, and numerous cooling-system components. AI is only one source of incremental copper demand and should not be treated as the sole driver of the commodity. But AI infrastructure is arriving at the same time as grid expansion, renewable generation, storage, transport electrification, and industrial investment are consuming more electrical equipment. The IEA's 2026 critical minerals outlook still projects a copper supply shortfall of roughly 25% relative to expected requirements in 2035 based on the current project pipeline.[17]

That makes copper a useful example of a second-order constraint. A data center developer may never negotiate with a copper miner directly, yet the price and availability of copper can affect the transformers, cables, switchgear, busbars, and generators that determine project cost. Investors looking only at the direct customer-supplier relationship between a hyperscaler and its server vendors can miss these shared upstream dependencies.

Cooling is moving in the opposite direction, from a building-level utility toward a technology closely coupled to the compute architecture. Nvidia's GB200 NVL72 is explicitly designed as a liquid-cooled rack-scale system.[5] Vertiv has been expanding liquid-cooling manufacturing and acquiring thermal-management capabilities, while Eaton completed its $9.5 billion acquisition of Boyd Thermal in March 2026 after valuing the business at 22.5 times estimated 2026 adjusted EBITDA when the deal was announced.[18]

That purchase multiple is notable because it shows where established electrical companies believe part of the next value pool sits. Eaton was not buying another utility-scale generator. It was buying the equipment that removes heat from increasingly dense compute systems. The strategic logic is that power delivery and heat removal are becoming a single infrastructure problem.

Vertiv's September agreement to acquire UtilityInnovation Group makes the same point from the other direction. The transaction would add microgrid controls, onsite-generation orchestration, specialized switchgear, and behind-the-meter power architecture to a company already heavily exposed to data center power and cooling.[19] The boundary between a data center equipment supplier and an energy infrastructure integrator is becoming less distinct.

The investment signal is shifting from GPU count to deliverable megawatts

The clearest implication is that investors should treat megawatt availability as a fundamental operating variable for the AI infrastructure cycle. Semiconductor shipments still matter, but they are increasingly downstream of a separate physical capacity cycle in electricity.

For data center operators and developers, power-secured land should command a different valuation framework from speculative acreage. A campus with transformer allocations, switchgear procurement, transmission capacity, and a credible energization schedule has removed several risks that a similarly sized announced project may still face. As construction pipelines grow, the gap between announced capacity and deliverable capacity could become more important than the headline pipeline itself.

For electrical-equipment manufacturers, long backlogs provide revenue visibility and pricing support, but investors should distinguish durable structural demand from scarcity that may eventually be relieved by new factories. Eaton, Hitachi Energy, Siemens Energy, GE Vernova, and others are all adding capacity. If manufacturing catches up with orders later in the decade, some of today's lead-time premium can normalize. The more durable opportunity may therefore belong to suppliers that combine equipment with installed-base service revenue, engineering expertise, controls, and broader system integration.

For generation-equipment companies, the key question is whether onsite and dedicated power remain bridge solutions or become a persistent feature of hyperscale development. Every year that utility interconnection takes longer than data center construction increases the economic value of an alternative that can produce usable power sooner. That does not mean every proposed onsite generation project will be permitted or economical. It means the option to bypass part of the grid delay has acquired measurable strategic value.

Utilities face a more complicated outcome. Large new loads can justify substantial generation and transmission investment and expand the future asset base, but the speed and concentration of data center demand create forecasting, reliability, financing, and cost-allocation problems. FERC's 2026 actions illustrate that regulators are increasingly focused not only on connecting large loads faster, but also on ensuring that speculative projects do not distort planning and that infrastructure costs are allocated appropriately.[7] A utility with exceptional access to power may benefit from the AI buildout while still facing political and regulatory resistance if residential customers believe they are subsidizing it.

The same framework creates second-order risks for the semiconductor industry. A delayed transformer does not reduce the theoretical demand for AI accelerators, but it can change when a customer can install them. If enough campuses are delayed simultaneously, electrical infrastructure can become a timing constraint on server deployment even when semiconductor supply is available. In that environment, power infrastructure functions as part of the effective supply chain of Nvidia, AMD, memory manufacturers, server assemblers, and cloud operators even when those companies do not purchase the equipment directly.

Conclusion

The AI infrastructure boom is no longer adequately described as a semiconductor capex cycle. It is becoming a coordinated buildout of computing capacity and electricity infrastructure, and the second system is increasingly determining how quickly the first one can grow.

The evidence is visible across the chain. Data center construction and electricity forecasts continue to rise, but interconnection procedures are being rewritten. Transformer factories and switchgear plants are expanding. Gas turbine slots are being reserved years in advance. Backup-generator manufacturers are reporting strong data center demand. Copper is shared with several other electrification cycles. Liquid cooling is becoming integrated into the architecture of high-density AI racks. Electrical and thermal suppliers are buying adjacent businesses to control more of the route from generation to chip.

The central investment conclusion is therefore specific: the marginal value in the AI buildout is shifting toward companies and assets that can shorten time to power. A server can be manufactured in months. The electrical system needed to operate a campus can take years. Until those timelines converge, the pace of AI infrastructure deployment will be set increasingly by transformers, turbines, switchgear, cooling systems, generators, transmission capacity, and grid approvals.

In practical terms, the most important question for the next AI data center may no longer be how many GPUs it plans to install. It may be when the first megawatt can actually be energized.

Sources

1. Americas Data Center Update, H1 2026, Cushman & Wakefield, September 14, 2026.

2. United States Data Center Energy Usage Report: 2025 Update, Lawrence Berkeley National Laboratory, June 2026.

3. Key Questions on Energy and AI, International Energy Agency, April 16, 2026.

4. U.S. Real Estate Market Outlook Midyear Review 2026: Data Centers, CBRE, 2026.

5. DGX GB Rack Scale Systems User Guide: Hardware, Nvidia.

6. North America Data Center Trends H1 2026, CBRE, August 27, 2026.

7. FERC Launches Targeted Action to Speed Large Load Integration, Federal Energy Regulatory Commission, June 18, 2026, and Interconnection of Large Loads to the Interstate Transmission System, Docket RM26-4-000.

8. US power companies scramble to secure equipment as surging data center demand strains supplies, Reuters, July 9, 2026.

9. Modernizing and expanding the power grid for tomorrow's energy, Siemens Energy.

10. Hitachi Energy breaks ground on the nation's largest facility for the production of large power transformers in South Boston, Virginia, Hitachi Energy, June 29, 2026.

11. GE Vernova reports second quarter 2026 financial results and raises 2026 financial guidance, GE Vernova, July 22, 2026.

12. Cummins Reports Strong Second Quarter 2026 Results; Raises Full-Year Outlook, Cummins, 2026.

13. American Intelligence & Power Forms Strategic Alliance with Caterpillar and Boyd CAT to Deploy 2 Gigawatts of Dedicated Power for Hyperscale AI Infrastructure, Caterpillar, January 28, 2026.

14. Energy and AI: Executive Summary, International Energy Agency, April 10, 2025.

15. Eaton Reports Record Second Quarter 2026 Results, Eaton, 2026.

16. Eaton expands operations in Nebraska with new manufacturing facility to meet increasing switchgear demand driven by AI data center boom, Eaton, April 8, 2026.

17. Global Critical Minerals Outlook 2026, International Energy Agency, July 16, 2026.

18. Eaton signs agreement to acquire Boyd Thermal, expanding solutions for data center customers to include critical liquid cooling technology, Eaton, November 3, 2025; acquisition completed March 12, 2026.

19. Vertiv Announces Agreement to Acquire UtilityInnovation Group to Accelerate Time to Power for AI Data Centers, Vertiv, September 2, 2026.

How to Cite This

For research inquiries or data access: press@altsets.com

Go Deeper