AI Power Infrastructure Faces Its First Rotation Test

AI power infrastructure is the cleanest follow-on test after the semiconductor rally because the next bottleneck is no longer only computing chips. It is whether enough electricity can be delivered, transformed, stored and cooled at the pace AI data centers require. That does not make the sector automatically cheap or safe. It does make the evidence stronger than in more crowded rotation stories where the link between orders and earnings is harder to verify.

The market has already learned the first AI trade: GPUs, high-bandwidth memory and advanced packaging. The second trade is messier. Power grids are slower than chips, utility permits are political, transformers have long lead times, and cooling systems are still moving from old air-based designs toward liquid-heavy architectures. That mess is exactly why the theme deserves attention. A hard physical constraint can keep demand visible even after the original software narrative cools.

📊 The bottleneck has moved from racks to megawatts

The strongest top-down number in the current debate comes from Gartner. According to a CIO report summarizing Gartner’s July 2026 forecast, global data center electricity consumption is projected to rise from 447 TWh in 2025 to 565 TWh in 2026, a 26% increase. Gartner also estimated data center power demand at 105 GW in 2025, 133 GW in 2026 and 291 GW by 2030.

Those figures matter because a 26% one-year increase in electricity consumption is not a normal IT upgrade cycle. It reaches into substations, grid connections, emergency generation, power management and cooling. Gartner’s estimate that AI-optimized servers will account for 31% of data center power consumption in 2026 adds the key distinction: ordinary server demand is not the main pressure point. AI workloads are.

The U.S. piece is also important for Korean suppliers. Gartner estimated that U.S. data centers would consume about 204 TWh in 2026, or 36% of global data center electricity use. Of that, AI-dedicated data centers were expected to consume 68 TWh. If that forecast is broadly right, North America remains the demand center that sets the tone for transformer, cable and backup power orders.

Policy has made Korea a live test case

Korea adds a second layer to the story: the state is trying to build a domestic AI infrastructure base, not merely export equipment into U.S. demand. According to EBN’s summary of the government’s “three mega-projects” announcement, the plan includes AI data centers totaling 18.4 GW around SK, GS and Naver-linked projects, with a first phase of 8.4 GW and a broader blueprint described at roughly 550 trillion won.

The headline number is huge, but the more investable detail is the policy machinery underneath it. The same report said the government intends to expand transmission networks, stabilize the power system, consider dedicated electricity tariffs for AI data centers, disclose 345 kV substation information and speed up grid-impact reviews. Those are not glamorous items. They are the plumbing that determines whether a data center plan can move from a press event to an energized site.

This is where power infrastructure looks different from themes such as shipbuilding, defense or nuclear. Those sectors may have powerful long-term stories, but they are already crowded with geopolitical narratives and project-specific uncertainty. AI power infrastructure has a simpler chain of evidence: more AI racks require more electricity; more electricity requires higher-capacity grid equipment; constrained supply gives qualified manufacturers pricing and backlog visibility. The chain can still break, but it is easier to audit.

🏭 Export data shows demand is already visible

The export numbers are the bridge between theme and reality. Herald Economy, citing Korea Customs Service trade statistics, reported that Korean exports of large transformers above 10,000 kVA and cables above 1,000 volts reached $721.56 million, or 1.0639 trillion won, in the first quarter of 2026. It was the first quarter above 1 trillion won and exceeded the previous quarterly record of $674.72 million set in the fourth quarter of 2025.

The destination matters as much as the total. The same report said the U.S. accounted for $397.24 million, or 585.7 billion won, meaning more than half of those exports went to the market where data center power demand is most concentrated. That helps explain why the Korean power equipment story is not only a local-policy trade. It is tied to North American grid replacement, AI campus construction and a transformer market where qualified capacity is scarce.

There is an important caveat. The customs figures are reported through a media summary, not linked here directly to the official database. That makes them useful but not perfect. For a serious investor, the next step is to reconcile those export categories with each company’s official order backlog, segment revenue and geographic exposure. The export data shows the tide. Company filings show who is actually catching it.

Order targets are becoming the company-level proof point

One company-level signal is HD Hyundai Electric’s 2026 order target revision. According to a company-distributed release carried by Korea Economic Daily, HD Hyundai Electric raised its 2026 order target from $4.222 billion to $5.185 billion. The increase was $963 million, or 22.8% above the prior target.

The explanation is more useful than the upgrade itself. The release cited rising demand for 765 kV power transformers, expanded North American power infrastructure investment, early orders ahead of the company’s second North American plant expansion, stronger demand for distribution transformers tied to North American data center expansion, and greater demand for land-based generators due to gas turbine shortages. That is a full physical AI stack: transmission voltage, distribution equipment and backup generation.

This is the right kind of evidence for a rotation candidate. A broad theme becomes more credible when management teams change numerical targets and explain the product categories behind the revision. It does not prove valuation upside, and it says nothing by itself about margins, execution or future competition. But it does show that AI data center demand is appearing in order planning, not just in market commentary.

Transformers, cables, ESS and cooling are not the same trade

The phrase “AI power infrastructure” can become too broad if used lazily. Transformers and high-voltage cables are capacity-constrained industrial products. ESS is partly a reliability and load-management solution. Cooling is a thermal-engineering response to denser server racks. They share the same AI driver, but their economics are different.

Transformers are currently the most direct bottleneck. Large power transformers are specialized, slow to qualify and difficult to scale quickly. A data center cannot simply buy a transformer from any supplier at the last minute, especially at ultra-high voltage. That is why Korean manufacturers with established product capability in high-voltage equipment have attracted attention.

Cables are linked to the same buildout but have a somewhat different profile. High-voltage cable demand rises when transmission and interconnection work expands, but project timing can be uneven because routes, permits and utility coordination matter. For investors, that means cable orders may validate the cycle while still producing lumpier revenue recognition than a simple demand chart suggests.

ESS sits between power availability and power quality. AI data centers need stable power, and storage can help manage peaks, backup transitions and renewable integration. Still, ESS exposure should not be treated as identical to transformer exposure. Battery pricing, chemistry, safety standards and project economics introduce a separate set of risks.

Cooling is the most technically dynamic part of the stack. AI servers generate far more heat than standard racks, which is why liquid cooling, immersion cooling and heat-exchange systems are gaining attention. The business opportunity is real, but the winning designs are less settled than in transformers. That makes cooling potentially attractive, but also more vulnerable to technology shifts and customer-specific designs.

⚠️ The counter-scenario starts with timing

The bearish case is not that AI uses no electricity. It clearly does. The real risk is timing. Data center projects can be delayed by permitting, grid queues, local opposition, power purchase constraints and construction bottlenecks. If orders are pulled forward faster than installations can proceed, suppliers may enjoy a backlog surge before investors discover that revenue conversion is slower than expected.

There is also the risk of over-extrapolation. A 22.8% order-target upgrade at one company is evidence, not a sector-wide valuation guarantee. Export records show demand strength, but they do not tell us margins, contract terms, warranty risk or how much capacity will be added by competitors. When a theme becomes fashionable, the market often capitalizes several years of demand before the earnings are fully visible.

Policy risk cuts both ways. Korea’s AI data center plan supports the domestic demand narrative, but policies can change with budgets, grid stress, local acceptance and electricity pricing. Dedicated AI tariffs, faster grid reviews and substation disclosure may improve project economics, but they can also become points of political debate if households or traditional industries see AI campuses as competing for scarce power.

Finally, there is a substitution risk inside the theme. More efficient AI chips, better workload scheduling, edge deployment, improved cooling and software optimization could reduce the intensity of power growth per unit of AI output. That would not eliminate the infrastructure need, but it could soften the most aggressive demand forecasts.

What would confirm the rotation

The next confirmation should come from orders, not slogans. Investors should look for official backlog growth in high-voltage transformers, distribution transformers, switchgear, cable systems, ESS contracts and cooling equipment. The most useful disclosures will separate geography, product type and delivery schedule. A larger backlog is more valuable when it comes with visible pricing discipline and realistic capacity plans.

Export data also deserves close tracking. If Korea’s large-transformer and high-voltage-cable exports continue to set records, and if the U.S. remains the dominant destination, the AI power infrastructure case becomes harder to dismiss as a local market fad. But export growth should be compared with company-level filings. The sector thesis is stronger when national trade data and individual order books point in the same direction.

The grid side is equally important. Government announcements about gigawatts are only the first step. The practical milestones are substation access, grid-impact approvals, transmission upgrades, signed power agreements and construction starts. In power infrastructure, a delayed interconnection can matter more than an exciting capex headline.

The disciplined view is this: after semiconductors, AI power infrastructure is a serious rotation candidate because it is tied to measurable electricity demand and visible equipment constraints. The case is not that every related stock should rise, or that the theme is early everywhere. The case is that the next stage of AI investment has to pass through steel, copper, batteries, cooling loops and substations. That makes the physical stack worth watching with filings open and forecasts on a short leash.