Precise analysis of AI data center electricity supply and demand across the United States, China, and Korea over the next decade (2026~2035). Among the five candidate bottlenecks (power, HBM, GPU, SSD, capital), which one is the real constraint — and where do the asymmetric investment opportunities created by that bottleneck lie?
* This ranking is not external; it reflects this article's synthesis of the five-bottleneck analysis below. Not a buy recommendation. Time horizon, risk tolerance, and portfolio fit depend on your situation.
Big Tech AI infrastructure CapEx began its full-throttle climb in 2024. In 2025 alone, the four largest US cloud companies are deploying roughly $315B into data centers and AI infrastructure — a record. That figure is equivalent to about 18% of South Korea's annual GDP.
Estimates based on Goldman Sachs / McKinsey / Synergy Research consensus:
The US faces the deepest demand-vs-supply gap. Combining IEA, Pew Research, EIA, and ICF data, US data center electricity consumption rises from 183 TWh in 2024 to ~426 TWh in 2030 and ~730+ TWh by 2035. New generation additions, however, run at less than half the pace of demand growth.
| Year | DC demand | New supply | Cum. supply | Cum. demand | Gap | Risk |
|---|---|---|---|---|---|---|
| 2026 | 230 | 52 | 4,520 | 4,540 | -20 | Manageable |
| 2027 | 275 | 56 | 4,576 | 4,608 | -32 | Warning |
| 2028 | 330 | 58 | 4,634 | 4,706 | -72 | PJM crisis |
| 2029 | 380 | 62 | 4,696 | 4,818 | -122 | Rate spikes |
| 2030 | 426 | 68 | 4,764 | 4,944 | -180 | Structural shortfall |
| 2031 | 490 | 72 | 4,836 | 5,090 | -254 | SMR online begins |
| 2032 | 555 | 78 | 4,914 | 5,255 | -341 | Distributed accel. |
| 2033 | 620 | 85 | 4,999 | 5,440 | -441 | — |
| 2034 | 680 | 92 | 5,091 | 5,640 | -549 | — |
| 2035 | 730 | 98 | 5,189 | 5,855 | -666 | Full distributed |
* 2026 and 2030 figures use IEA / Pew / EIA base-case scenarios. Intermediate years (2027–2029, 2031–2034) are this article's interpolations. Forecasting institutions can vary by ±15%.
Loudoun County in Northern Virginia carries an estimated 70% of global internet traffic. As of 2024, data centers consume about 26% of state-wide electricity, and Dominion Energy is quoting 4–7 year wait times for new DC interconnections. An additional ~16 GW of demand is expected by 2030.
Cheap power and faster permitting have made Texas a hotspot for new data centers, although the 2021 cold snap left lingering concerns about grid resilience. xAI Colossus (near Memphis) and Stargate Phase 1 (Abilene) are headline cases. The energy mix is gas + wind + solar + batteries.
Phoenix anchors a Microsoft / Meta / Apple cluster. Reno (Nevada) hosts Google and Tesla. Iowa is Microsoft's AI supercomputing hub. The decision driver is consistent: cheap power + cool climate + state incentives.
Key shifts: Gas generation rises in absolute terms but loses share. Solar climbs explosively (8% → 30%). SMR comes online materially in the late 2030s, edging nuclear share higher. Coal nearly retires by 2035 (10% → ~2%).
China is playing a different game than the US. With state-controlled power allocation, grid bottlenecks bite less than in the US's market-based system. At the same time, NVIDIA H20 / B200 export controls impose strong domestic-substitution pressure on chips and memory. The "East-Data-West-Computing" (东数西算) initiative is the centerpiece.
| Hub | Region | Use | Power edge |
|---|---|---|---|
| Yangtze Delta | Shanghai · Jiangsu · Zhejiang | Real-time AI inference | Gas + nuclear |
| Greater Bay Area | Guangdong · Guangxi | Finance · real-time | Nuclear + hydro |
| Jing-Jin-Ji | Beijing · Tianjin · Hebei | Government / finance AI | Nuclear + wind |
| Chengdu-Chongqing | Sichuan · Chongqing | Training + inference | Abundant hydro |
| Inner Mongolia | Hohhot · Ulanqab | Large-scale training | Wind + solar |
| Guizhou | Guiyang | Large-scale training | Hydro + solar |
| Gansu | Qingyang · Yinchuan | Large-scale training | Solar + wind |
| Ningxia | Yinchuan | Large-scale training | Solar |
| Year | DC demand | YoY | Renewables share | Key issue |
|---|---|---|---|---|
| 2026 | 240 | +22% | 32% | H20 export controls bite |
| 2027 | 285 | +19% | 35% | Huawei Ascend volume ramp |
| 2028 | 335 | +18% | 38% | Domestic HBM (CXMT etc.) starts |
| 2029 | 390 | +16% | 42% | — |
| 2030 | 440 | +13% | 45% | Closing on 50% renewables target |
| 2031 | 510 | +16% | 48% | — |
| 2032 | 580 | +14% | 52% | — |
| 2033 | 650 | +12% | 55% | SMRs come online |
| 2034 | 720 | +11% | 58% | — |
| 2035 | 790 | +10% | 60% | Post-carbon-peak effects |
* China DC demand combines unofficial government statistics with private estimates (IDC China · Frost & Sullivan · Synergy). Source data is thinner than for the US, and ranges are wider.
In February 2025, Alibaba committed ¥380B (~$53B) to AI and cloud over three years. Footprint in all eight East-Data-West-Computing hubs. Self-developed Qwen LLM family. Targeting 10%+ global cloud share by 2030.
Tencent ¥100B+, ByteDance (TikTok parent) ¥80B in AI infrastructure. Baidu runs Apollo + Ernie. All three are building proprietary DCs across the East-West clusters. China's own AI CapEx supercycle is underway.
In 2025, DeepSeek-V3 and R1 claimed comparable performance to leading US frontier models at around 1/10 the cost (DeepSeek's own claim, still under academic and industry verification). If DeepSeek-style efficiency gains generalize, the demand projections above could move down meaningfully. The counter-scenario, though, is "Jevons Paradox" — efficiency gains expanding total usage rather than shrinking it.
Korea supplies 80%+ of the world's HBM but its own data center build-out is constrained by a triple bind: RE100 commitments, transmission shortages, and renewable lag. The 11th Basic Plan for Long-term Electricity Supply and Demand (2024–2038) defines the next decade.
| Year | DC capacity (MW) | DC power (TWh) | Renewables share | RE100 gap |
|---|---|---|---|---|
| 2026 | 2,400 | 17.5 | 11% | Large |
| 2027 | 3,100 | 22.6 | 13% | Large |
| 2028 | 4,500 | 32.9 | 16% | Very large |
| 2029 | 5,400 | 39.4 | 19% | Very large |
| 2030 | 6,320 | 46.1 | 21.6% | Improving |
| 2031 | 7,300 | 53.3 | 24% | Easing |
| 2032 | 8,400 | 61.3 | 27% | Easing |
| 2033 | 9,600 | 70.1 | 29% | Easing |
| 2034 | 10,900 | 79.6 | 31% | Stable |
| 2035 | 12,300 | 89.8 | 33% | Stable |
* 2025 and 2030 figures derive from the 11th Basic Electricity Plan. 2031~2035 extrapolates the same trajectory; renewables share interpolates toward the plan's 32.9% (2038) target.
Samsung Electronics Pyeongtaek campus (P1–P5) plus SK Hynix Yongin Semiconductor Cluster. Roughly 10 GW of additional electricity load required by 2030. KEPCO's transmission-line construction from East Coast nuclear and LNG plants is in progress, but resident opposition is causing 7+ year delays.
NAVER Gak (Chuncheon · Sejong), KT IDC, LG U+, Kakao, Kakao Enterprise. 76 new dedicated AI DCs are planned by 2028, averaging 50–100 MW each. Roughly half are expected to slip on schedule due to power infrastructure shortages.
AWS Seoul Region adding a 4th AZ; Microsoft Korea Central; Google Cloud Seoul. Korea's location offers optimal latency between Japan and China. All three carry RE100 commitments, making Korea's renewable shortage a direct business risk.
We score the five candidate bottlenecks of AI data center build-out on a common scale. Higher score = harder bottleneck (more difficult to remove).
| Bottleneck | 2026 | 2027 | 2028 | 2030 | 2035 | Structural |
|---|---|---|---|---|---|---|
| ⚡ Power (grid) | 95 | 95 | 92 | 85 | 70 | Very high |
| 💾 HBM / CoWoS | 88 | 82 | 72 | 60 | 40 | Medium |
| 🎮 GPU silicon | 68 | 55 | 40 | 30 | 25 | Low |
| 💿 SSD / NAND | 35 | 38 | 42 | 45 | 42 | Very low |
| 💵 Capital | 15 | 18 | 22 | 28 | 35 | Very low |
* These scores are not from an external source — they are this article's composite judgment, weighted across (1) physical constraint, (2) time constraint, (3) capital-solvability. Higher = harder to clear.
Why this is the hardest: (1) New transmission lines 5–10 years (resident opposition, environmental review). (2) Large transformer global lead times 3–4 years. (3) New generation permitting 2–5 years + construction 2–5 years. Capital alone cannot solve it. Workarounds: behind-the-meter gas/SMR, direct PPAs, distributed microgrids.
Diagnosis: NVIDIA H100/H200/B200 and AMD MI300X all consume HBM3E. The true gatekeeper is CoWoS packaging at TSMC: capacity climbs from 30K wafers/month (2024) to ~75K by end-2025. SK Hynix ~50%, Samsung ~42%, Micron ~8%. Capacity normalization expected 2027–2028.
Diagnosis: H100 lead times of 12+ months in 2023–24 have normalized to 3–6 months in 2025. NVIDIA Blackwell B200/B300 ramp continues. AMD MI300X / MI350, Intel Gaudi 3, and hyperscaler ASICs (Trainium · TPU · MTIA) diversify the supply.
AI training datasets and checkpoints lift NAND demand ~45% YoY, but SK Hynix (incl. Solidigm), Samsung, Micron, and SanDisk (WDC) all supply. Temporary tightness in 2026 followed by normalization in 2027. Much weaker than HBM.
Microsoft + Google + Amazon + Meta combined CapEx ~$315B in 2025 alone — record. Combined cash positions ~$400B. The scenario "AI infrastructure runs out of money" is unrealistic. The risk is an AI revenue disappointment around 2027–2028 slowing CapEx, not capital scarcity per se.
The power bottleneck is fourfold: generation, transmission, conversion (transformers), and last-mile distribution. Any one of these breaking blocks the entire data center. The lead times below show the real depth of the constraint.
September 2024: agreement to restart Three Mile Island Unit 1. Microsoft buys all 835 MW for 20 years. Historic — the first time Big Tech has revived a closed reactor. Restart targeted for 2028.
AWS data center co-located with Talen's Susquehanna nuclear plant in Pennsylvania. ~$650M site purchase, up to 960 MW direct PPA. Behind-the-meter design. Note that FERC rejected PJM's revised ISA in Nov 2024; partial approval and refiling have followed. The first regulatory test of behind-the-meter at scale.
First SMR online targeted for the early 2030s. Amazon also invests in X-energy SMR; Microsoft signs Helion (fusion). Big Tech is becoming the new "anchor tenant" of future power.
NVIDIA H100 / H200 / B200, AMD MI300X / MI350, Google TPU v6, Amazon Trainium 2 — every modern AI accelerator depends on HBM. And every HBM stack must pass through TSMC's CoWoS packaging to be assembled with its GPU. These two stages are the real gate.
Korea = ~92%. That share is the real moat of Korean memory. (*Average of 2024 market-research-firm estimates. Quarterly variation possible.)
HBM bottleneck eases only as CoWoS capacity unlocks in 2027–2028. Until then, NVIDIA / AMD GPU shipments are effectively rate-limited by CoWoS.
HBM's critical "TC (Thermocompression) Bonding" tool ships almost exclusively from Hanmi to SK Hynix. Micron has also adopted Hanmi tools. Samsung is developing its own. The narrowest alley of the HBM supercycle = the sharpest pick.
Sharp re-rating in 2024–2025, then consolidation. HBM3E → HBM4 transition could re-accelerate revenue. Single-stock volatility is high — average in.
Translating the bottleneck analysis into a three-tier framework: Tier 1 = direct beneficiaries of the bottleneck, Tier 2 = picks-and-shovels infrastructure, Tier 3 = Korea-specific.
Thesis: ~50%+ HBM share. Core supplier of NVIDIA H100 / H200 / B200. HBM3E 12-Hi in volume, HBM4 ramping in 2026. "The Intel of the AI era" position. Operating margins at multi-decade highs.
Risk: Samsung catch-up + Micron entry could trim share. Memory cyclicality. Valuation: 2026 P/E ~8–10× — discount to global memory peers.
Thesis: US nuclear leader (~94 TWh/yr). Microsoft 20-year PPA signed. AI DCs craving 24/7 clean power = explosive pricing power. New SMR construction takes 5–7 years, but existing reactors are immediately available.
Risk: Tail-risk of nuclear incident (low but asymmetric). Policy changes. Valuation: Up ~90% in 2024; P/E ~30× — reasonable given growth.
Thesis: The narrowest, sharpest Tier 1 pick. Effective monopoly on the key HBM packaging tool. SK Hynix and Micron both run Hanmi tools. HBM build pace → Hanmi revenue, directly.
Risk: Customer concentration. Samsung in-house tool risk. Significant volatility. Size positions accordingly.
Thesis: Gas + nuclear + battery portfolio. Top-2 US IPP. Up ~260% in 2024 then consolidating. Surge in DC PPA demand + ERCOT pricing power. Higher gas weighting offers nearer-term visibility than CEG.
Thesis: The real gatekeeper of every AI chip. NVIDIA, AMD, Apple, Tesla AI silicon all packaged here. CoWoS expanding from 30K → 75K → 150K wafers/month. The most stable AI exposure for the decade.
Global leader in transformers, UPS, switchgear. AI DC electrical content runs 3–5× a traditional DC. Backlog at record highs. 5+ year revenue visibility.
NVIDIA reference-design partner. Surging demand for immersion and direct-to-chip cooling as AI chip TDP explodes. Up ~140% in 2024.
US transmission and distribution EPC leader. Direct beneficiary of Goldman Sachs' projected $720B grid investment by 2030. Backlog $35B+.
Direct beneficiary of behind-the-meter gas demand. Global gas turbine #1. Wind portfolio adds optionality. Up ~130% in 2024.
NuScale-based SMR + gas turbines + nuclear EPC. Core beneficiary as SMRs scale in the late 2030s. Volatile until SMR commercialization. Long-horizon position only.
Major Korean transformer exporter to the US. US grid transformer lead times 3–4 years = pricing power for HD Hyundai. Korean transmission supercycle adds a second tailwind. Up 200%+ in 2024 yet backlog remains strong.
Transformers + distribution + automation. US and India export ramp. Direct beneficiary of Korean grid build-out + DC growth. More attractive valuation than HD Hyundai Electric.
HBM3E 12-Hi qualified at NVIDIA (2025), HBM4 in volume ramp. Foundry chasing TSMC. Underowned + with momentum. Useful as a hedge alongside SK Hynix.
Logic: Tier 1 (60%) + Tier 2 (28%) + Tier 3 (12%). HBM axis (SK Hynix + Hanmi + Samsung ≈ 30%+) + Power axis (CEG + VST + ETN + VRT + PWR + GEV ≈ 50%) + CoWoS axis (TSMC = 12%) + Korean infrastructure (HD Hyundai etc. ≈ 8%). Time horizon 5–10 years.
If AI revenue disappoints in 2027–2028, Big Tech CapEx could pause. The 2000 dotcom analog is worth remembering. Mitigation: diversify across Tiers 1/2/3, average in, hold for 5+ years.
If DeepSeek-style efficiency (10× cheaper for similar performance) becomes the industry standard, GPU and power demand could trim 30–50%. But Jevons Paradox (efficiency drives more total usage) could pull the other way. Watch closely.
IRA tax credit cuts would weaken solar, EV, and SMR incentives. But the AI hegemony race tends to absorb policy shifts. Power demand itself is structural and policy-agnostic.
If China accelerates self-reliance via Huawei Ascend / CXMT / SMIC, NVIDIA and SK Hynix's China revenue compresses. But ~80% of the global market is the US and allies, limiting impact.
"The real gate of the AI era is not GPUs but power and HBM. A 5–10 year diversified position concentrated on those two axes delivers the most asymmetric returns. SK Hynix · CEG · Hanmi · TSMC · VST form the core. Don't enter without a long horizon."