KOSUNI CRYPTO
Korea's Crypto Pulse, in English
$BTC $ETH $XRP $SOL $DOGE
Macro
Neutral

AI's Next Bottleneck Isn't GPUs—It's the Power Grid

Published September 27, 2026 10:59 PM · 13 views
AI's Next Bottleneck Isn't GPUs—It's the Power Grid

Artificial intelligence (AI) is now devouring the world’s power grids after consuming semiconductor supply chains. For years, the question in AI infrastructure competition was simple: "Who can secure more GPUs?" But having tens of thousands of chips isn’t enough—power must be supplied to run them, electricity delivered from power plants to data centers, and transformers and switchgear needed to convert high-voltage power for server use. This is where a new bottleneck emerges.

According to Goldman Sachs’ August 28 report on South Korea’s power equipment industry, U.S. electricity demand is projected to grow at an annual average of 3.5% between 2025 and 2030, with AI data centers accounting for about 70% of the increase. Goldman anticipates a shortage in power equipment supply until at least 2029, expecting steady new orders for South Korean power equipment firms through 2028.

AI has disrupted decades of stagnant U.S. electricity demand, which had remained nearly flat for 20 years due to energy-efficient appliances, LED adoption, and manufacturing shifts overseas. Goldman notes that AI will accelerate this trend: data center electricity consumption in the U.S. is expected to jump from 31GW in 2025 to 41GW in 2026, then to 66GW by 2027—a more than doubling within two years. The share of total U.S. electricity demand consumed by data centers will rise from 4.1% in 2025 to 8.5% by 2027.

Long-term, the scale is even larger: global power demand from AI chips could reach approximately 315GW by 2033, an increase of over 11 times compared to 2025 levels, with about 64% coming from the U.S. This shift marks a transition from computing competition to energy competition in the AI race.

The problem isn’t just power generation—it’s transmission. Building new power plants alone won’t solve the issue; without sufficient transmission infrastructure, electricity can’t reach data centers. Key components like large transformers and switchgear are already in short supply. The U.S. Department of Energy reports that general distribution transformer lead times have stretched from 3–6 months in 2019 to over one or two years now, while large power transformers require waits of up to three to four years, with some ultra-high-voltage models taking as long as five years.

Unlike GPUs, which can be acquired faster by paying more, transformers require custom specifications, complex manufacturing processes, quality certifications, and installation timelines spanning years. Building new factories to increase supply also takes several years. The pace of AI data center construction does not match the pace of power grid development.

This bottleneck isn’t exclusive to AI. The U.S. must simultaneously replace decades-old power grids, expand renewable energy facilities, connect new power plants, and accommodate manufacturing reshoring and electric vehicle growth. The DOE states that shortages in transformers, switchgear, substations, and power electronics are delaying not only data center connections but the entire expansion of the U.S. power grid.

South Korea has positioned itself advantageously. Unlike Western nations that scaled back power equipment production over the past two decades due to stagnant demand, South Korea maintained steady investment in its power grid infrastructure amid industrialization and growth in power-intensive sectors like semiconductors and displays. This has left Korean firms with advanced manufacturing capabilities, skilled labor, and supply chains for high-voltage transformers, switchgear, and transmission equipment.

Hyosung Heavy Industries, HD Hyundai Eelectric, and LS Electric have already benefited from this structural shift. In the first quarter of 2024, these three companies secured new orders exceeding KRW 7 trillion, with combined order backlogs surpassing KRW 30 trillion. This growth stems from both AI data centers and U.S. grid modernization efforts.

Hyosung recently won a KRW 78.7 billion contract for 765kV ultra-high-voltage transformers and reactors from a major U.S. transmission operator—the largest single project ever secured by a Korean power equipment firm in the U.S. HD Hyundai Electric also signed a long-term supply agreement with a global tech giant worth up to KRW 112.12 billion for data center distribution and power equipment.

While stock prices have risen significantly, concerns about a market peak persist. However, Goldman Sachs projects that new order momentum will continue through 2028, with the industry’s overall EPS growing at an annual average of 28% from 2026 to 2030. Hyosung is expected to see around 34% annual EPS growth during this period, while HD Hyundai Electric and LS Electric are projected at 19% and 31%, respectively.

It’s important to note that power equipment revenue recognition lags behind new orders. High-voltage transformers ordered today won’t appear on financial statements immediately—they require years for manufacturing, testing, shipping, and installation. Thus, even if order growth slows in the future, existing backlogs will gradually convert into revenue and profit over time.

Goldman specifically highlights transmission infrastructure (not distribution) as key. Data center projects face delays due to power availability, permits, and construction issues, but nationwide grid upgrades are less affected by individual project setbacks. Transmission equipment benefits from broader infrastructure investment trends, making it relatively insulated from AI project risks.

This is why Goldman recommends a "Buy" rating for Hyosung Heavy Industries while maintaining a "Neutral" stance on HD Hyundai Electric and LS Electric—these are Goldman Sachs’ investment recommendations, not financial advice from this publication.

AI’s competitive unit has shifted from GPUs to megawatts (MW). Having GPUs isn’t enough without power access. Even with sufficient electricity, data centers can’t operate without grid connection. Future AI infrastructure scarcity will likely include power contracts, substations, transmission lines, interconnection rights, and land connected to power plants.

The cryptocurrency mining industry has long understood this challenge. Bitcoin miners prioritize low-cost electricity, stable supply, and secured grid connections over raw hardware performance. Many North American crypto miners are now pivoting to AI/HPC data centers using existing infrastructure—this isn’t just diversification but leveraging their most valuable assets: hundreds of megawatts of power and interconnection rights.

GPU can be purchased anew, but grid connection is not something that can be bought immediately with money. This explains why AI and Bitcoin mining, despite seeming unrelated, converge at the infrastructure level.

The first "pick-and-shovel" opportunity in the AI era was GPUs; the second was data centers and cooling systems. Now the third wave has arrived: power equipment. Transformers, switchgear, HVDC lines, cables, and energy storage systems may not be as flashy as AI models, but without them, even billion-dollar GPU clusters remain idle.

As AI expands digital economies, physical infrastructure demands more capital investment. The future of data centers and grid infrastructure could become a new asset class for Real-World Asset (RWA) tokenization, though this market is still nascent.

For South Korea, this shift presents two opportunities: export growth as power equipment firms rise in the AI infrastructure cycle, and domestic AI development. However, South Korean AI expansion faces similar challenges—power grid constraints, interconnection bottlenecks, and regional electricity concentration—that could delay local industry growth even as Korean companies profit from U.S. demand.

AI’s first battle was for GPUs; the second was for data centers. Now a third war is beginning: who controls power access and transmission capacity. The next bottleneck isn’t chips—it’s the grid that powers them.

Korean Source

This article is an English localization of a Korean-language crypto news report. Original headline: [에너지분석] AI의 다음 병목은 GPU가 아니다… ‘전력망’이 막힌다