Beyond Nvidia: How AI Data Centres Are Driving Demand for Power, Transformers and Cooling

AI infrastructure is becoming a power-and-thermal engineering story as much as a semiconductor story. More AI servers mean more electricity must be transformed, distributed, backed up and converted inside a facility—and more heat must be removed safely. That expands the addressable market for transformer makers, power-system specialists and liquid-cooling suppliers beyond Nvidia.

Why the infrastructure spend is accelerating

McKinsey projects global data-centre demand could grow from about 82 GW in 2025 to roughly 220 GW in 2030. AI-related demand is expected to rise 3.5 times, from about 44 GW to 155 GW, compared with 1.7-times growth for non-AI workloads. 22

The investment figures are correspondingly large: McKinsey estimates data centres may require $6.7 trillion globally by 2030, including $5.2 trillion for AI-ready capacity. These are estimates of required capital expenditure, not booked orders or committed spending. 28

Higher rack density is the mechanism that spreads spending across the supply chain. A Bank of America scenario cited in reporting projects AI-rack power demand could exceed 1.5 MW by 2030. If that occurs, each deployment requires materially more electrical infrastructure and more capable cooling than a conventional server environment. 20

The suppliers benefiting—and why

Transformers and grid equipment: HD Hyundai Electric and Hainan Jinpan

Transformers step down grid electricity to levels usable by servers, cooling systems and power-distribution equipment. They have become a practical bottleneck as utilities and data-centre developers place orders further ahead to secure scarce equipment. Reuters reported strong first-half 2026 AI-infrastructure demand, particularly from North America, for South Korea's HD Hyundai Electric and China's Hainan Jinpan Smart Technology. 142

HD Hyundai Electric provides a useful example of what constrained supply can mean operationally. It reported more than $1.4 billion in second-quarter new orders and an order backlog approaching $8.5 billion at the end of June 2026. 36 A sizeable backlog supports revenue visibility, but it is not the same as revenue: delivery schedules, component costs and project completion still determine conversion into earnings.

Power conversion and systems: Delta Electronics

Data centres need far more than a grid connection. They require power conversion, distribution, backup systems and controls within the facility. Delta Electronics has exposure across AI power, cooling and data-centre infrastructure, and told Reuters that these solutions remain a growth engine as it expands production across Thailand, the United States and other locations. 1

This breadth can be valuable because integrated systems address a harder customer problem than a single component. The trade-off is that system suppliers still face global competition and must demonstrate reliability at scale.

Thermal management: AVC, Auras and Envicool

Asia Vital Components (AVC), Auras Technology and Shenzhen Envicool are among the thermal-management suppliers exposed to the shift toward denser AI systems. As racks run hotter, the opportunity moves beyond conventional fans and heatsinks to cold plates, coolant-distribution units, pumps, manifolds, heat exchangers and facility-level thermal controls.

Liquid cooling is central to that shift. Microsoft says traditional air cooling cannot handle the density of modern AI hardware in its AI data centres, which use liquid circulated directly into servers to extract heat. 8 One Bank of America forecast expects liquid cooling to account for 70% of new AI-data-centre installations by 2030, up from roughly 30%. 2021

That forecast should be read as a scenario rather than a certainty. Adoption will vary by chip architecture, rack density, reliability requirements, operator preference and the availability of qualified cooling systems.

Why power access may matter more than compute capacity

The International Energy Agency expects global data-centre electricity consumption to more than double to around 945 TWh by 2030. It also estimates that about 20% of planned data-centre projects could face delays if grid risks are not addressed. 43

The problem is a mismatch of timelines. AI facilities can be planned and built in a few years, while transmission upgrades, substations and interconnections can take much longer. In many regions, grid connections can take four to 10 years. 53

For equipment makers, this creates two opposing effects:

• Near term: shortages of transformers, switchgear and related equipment can support orders, pricing and backlog visibility. 42
• Long term: the same shortages can delay customers' projects and postpone the equipment deliveries on which supplier revenue depends.

This is why the strongest participants are likely to be those with certified products, dependable manufacturing capacity, component supply and the ability to meet customer schedules—not merely broad exposure to the AI theme.

Efficiency claims need careful interpretation

Solid-state transformers and silicon-carbide power architectures could eventually support more efficient, DC-oriented data-centre designs. Wolfspeed cites up to a 5% end-to-end efficiency improvement and 25% to 40% lower conversion losses for certain SiC-enabled stages. Those are vendor architecture claims, not a sector-wide measure of realised savings. 45

Similarly, no provided evidence substantiates a universal 27%-plus energy-saving figure for liquid cooling. Cooling performance depends on the air-cooled baseline, chip density, coolant temperatures, facility design and whether waste heat is reused. The more defensible conclusion is that high-density AI deployments are increasing the need for liquid cooling; the precise energy benefit is site-specific. 824

The available evidence also does not support treating a 40% adoption forecast for solid-state transformers as an established industry outcome. Conventional transformers remain essential to grid and facility infrastructure, while solid-state designs must still clear cost, qualification, reliability and grid-compliance hurdles.

The margin opportunity—and the risks

Power-equipment suppliers can benefit when qualified capacity is scarce. High-voltage transformers and related equipment are specialized, long-lead products, and the installed base requires grid upgrades as well as new data-centre construction. HD Hyundai Electric's reported order growth and backlog illustrate that favorable dynamic. 36

Thermal suppliers have a different profile. Liquid cooling raises content per rack and rewards engineering, reliability and system integration. Yet the thermal ecosystem is competitive, and individual components can become commoditized more quickly than complete systems. The key risks include rapid architecture changes, customer concentration, pricing pressure and manufacturing expansion that eventually eases today's supply constraints.

A practical framework for assessing likely winners is to ask whether a supplier can:

1. manufacture qualified equipment at scale;
2. secure critical components and capacity ahead of demand;
3. deliver integrated power or thermal systems rather than isolated parts;
4. preserve pricing and reliability through fast product cycles; and
5. expand internationally without damaging quality, cash flow or delivery performance.

Are floating and underwater data centres a major answer?

Floating and subsea data centres may help in specific locations where land, freshwater, cooling or permitting is constrained. Underwater facilities can use surrounding seawater for cooling, and China has moved projects toward commercial operation, including an offshore wind-powered underwater data centre near Shanghai. 47

But these designs remain niche alternatives rather than replacements for mainstream land-based campuses. They introduce their own engineering, maintenance, connectivity and permitting challenges. Their near-term importance is less about displacing conventional construction and more about showing how operators may respond when power, water and land constraints become severe.

Bottom line

The AI boom is broadening the infrastructure opportunity beyond GPUs. Transformer makers such as HD Hyundai Electric and Hainan Jinpan are positioned where grid equipment is scarce; Delta has exposure to the power-and-cooling stack; and AVC, Auras and Envicool are exposed to the increasing thermal content of high-density racks. 1

The durable winners will be suppliers that turn demand into dependable deliveries and profitable systems. The main caveat is equally important: grid access, equipment lead times and project execution can slow the buildout even when AI demand remains strong. In this cycle, electricity delivery and heat removal are not secondary details—they are core constraints on how quickly AI capacity can be built. 4342