The Power Wall: How Energy Grid Constraints Are Becoming the Biggest Bottleneck in the AI Data Center Boom

A single Nvidia H100 GPU consumes 700 watts. A rack of 8 H100s consumes 5.6 kilowatts. A data center row of 40 racks is 224 kilowatts. A 1,000-rack AI data center hall is 5.6 megawatts. A hyperscale AI training facility with 50,000 GPUs is 280 megawatts — equivalent to a mid-sized coal power plant.

The numbers scale brutally. And the world is building thousands of these facilities simultaneously.

The US Department of Energy's January 2024 report on data center electricity demand projected that data centers will consume between 6.7% and 12% of US total electricity by 2028, up from approximately 4% in 2023. The midpoint — 9% — represents the addition of roughly 150 gigawatts of new electricity load over five years. For context, the entire US nuclear power fleet generates approximately 100 gigawatts.

This is the power wall: the constraint that no amount of investment can instantly solve, because building new power generation capacity takes 5-15 years. And it is reshaping where AI infrastructure gets built, what it costs, and who can afford to build it.

Where the Load Is Coming From

The GPU demand for AI training is the headline number, but the electricity footprint of AI is broader:

AI training clusters: The most power-intensive per unit of compute. A single GPT-4-scale training run consumed an estimated 51.7 megawatt-hours of electricity. A Stargate-scale facility running continuous training workloads consumes hundreds of megawatts continuously.

AI inference at scale: Inference is less power-intensive per query than training, but the volume is orders of magnitude larger. OpenAI processes approximately 10 billion words per day via ChatGPT. Google Search now includes AI-generated summaries for almost every query. The aggregate inference load across all deployed AI is already significant and growing faster than training.

Traditional cloud compute: Even before AI, cloud data centers were growing at 10-15% annually driven by video streaming, e-commerce, financial services, and enterprise software. AI is layered on top of this baseline growth.

Cooling systems: In hot climates or dense deployments, cooling can consume 30-40% of total data center power. The Power Usage Effectiveness (PUE) ratio — total facility power divided by IT equipment power — for new AI data centers targets 1.2-1.3 (meaning 20-30% overhead for cooling and other infrastructure). Legacy facilities have PUE of 1.5-2.0.

The Grid Response

Power grid operators across the US and globally are responding to the load growth with alarm:

ERCOT (Texas): The Texas grid operator projected in its 2024 long-term study that data center load in Texas could reach 70 GW by 2030 — more than Texas's total current installed generation capacity of approximately 100 GW. ERCOT has implemented new data center interconnection studies and large load management requirements.

PJM (US Mid-Atlantic and Midwest): Northern Virginia — the world's largest data center concentration, known as "Data Center Alley" — is served by PJM. PJM reported in 2024 that it was struggling to maintain reserve margins due to accelerating data center load growth. Dominion Energy (the primary Virginia utility) has a backlog of 90+ GW in interconnection requests.

National Grid (UK): The UK's electricity system operator flagged in its 2025 Future Energy Scenarios report that AI data center load could require an additional 15-20 GW of UK generation capacity by 2035, equivalent to building 15 new nuclear plants.

Singapore: Singapore's data center moratorium (see Southeast Asia coverage) was partly a response to data centers consuming 7% of the city-state's electricity on approximately 0.7% of its land area.

The Nuclear Response

Faced with grid constraints, hyperscalers have made an extraordinary pivot: they are signing long-term agreements for nuclear power at a pace not seen since the 1970s.

Microsoft — Three Mile Island / Crane Clean Energy Center: Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1 (now renamed Crane Clean Energy Center) in Middletown, Pennsylvania. The plant came back online in September 2024. It provides 837 MW of carbon-free power, exclusively to Microsoft's data centers.

Google — Kairos Power SMR: Google signed an agreement with Kairos Power to purchase electricity from multiple small modular reactor (SMR) sites, with first power targeted 2030 and 500+ MW by 2035.

Amazon — Susquehanna Nuclear Campus: Amazon purchased the Cumulus Data data center campus adjacent to PPL Corporation's Susquehanna nuclear power plant in Pennsylvania — a direct nuclear-to-data center power connection.

Oracle / Stargate — SMR planning: The Stargate Project is actively planning on-site SMR deployment for its Texas campuses. Larry Ellison has publicly discussed SMR-powered AI campuses.

The nuclear pivot is driven by two factors: carbon commitments (all major hyperscalers have net-zero pledges) and reliability. Nuclear plants operate at 90%+ capacity factor — they produce power reliably regardless of weather. Solar and wind produce power intermittently, which creates grid management complexity for always-on AI workloads.

Geothermal and the Iceland Factor

Iceland has an interesting position in the global AI infrastructure map: abundant geothermal power at low cost, arctic cooling that reduces data center cooling costs dramatically, and political stability. Iceland's data center industry has grown 400% since 2018.

Iceland's limitations are connectivity (submarine cables to Europe only) and total power capacity (Iceland's entire electricity system is approximately 3 GW — too small for hyperscale concentration). But it is an increasingly popular location for AI training workloads that can tolerate higher latency (batch training, not real-time inference).

Greenland and Norway (large hydroelectric capacity) are also being evaluated by hyperscalers for training facility siting.

Impact on Cloud Costs

The energy constraint is beginning to flow through to cloud pricing. For several years, competition among hyperscalers drove GPU instance prices down despite rising electricity costs — AWS, Azure, and Google absorbed margin compression to gain market share.

In 2025-2026, this dynamic is reversing. GPU instance price reductions have slowed. Long-term reserved instance pricing for GPU compute has increased. And the most energy-constrained geographies — Northern Virginia, Singapore, parts of Europe — are seeing higher on-demand GPU spot prices.

Three dynamics are reshaping cost:

Power Purchase Agreement amortisation: Hyperscalers that sign 20-year nuclear or renewable PPAs at fixed prices are insulated from electricity market volatility. Those PPAs cost capital upfront but reduce long-term energy cost uncertainty. The capital cost is being amortised into GPU compute pricing.

Location-based pricing divergence: AI compute from data centers with cheap, abundant power (Texas wind, Pacific Northwest hydro, Nordic geothermal) is becoming cheaper than equivalent compute in power-constrained markets. Some sophisticated enterprises are already routing AI workloads to specific cloud regions based on real-time energy cost data.

Liquid cooling premium: AI GPUs generate heat that air cooling handles poorly at density. Liquid cooling (direct-to-chip or immersion cooling) is more efficient but more expensive to build and operate. New AI-optimised data centers are all-liquid-cooled; retrofitting existing facilities adds cost. This capital expenditure increase is being priced into next-generation GPU instance offerings.

Developer Implications

Region selection now includes energy considerations. When choosing cloud regions for AI workloads, energy availability and cost are relevant factors alongside latency and data sovereignty. AWS us-east-1 (Northern Virginia) is increasingly power-constrained; AWS us-west-2 (Oregon, hydro power) is not.

Spot instance availability for AI training will vary. In power-constrained regions, GPU spot instances may be unavailable during peak grid demand periods. Plan for training run interruptions and implement checkpointing.

On-premise AI infrastructure economics have shifted. The hyperscaler power access advantage — they can sign nuclear PPAs that individual enterprises cannot — means cloud GPU cost is relatively more competitive than it was in 2022-2023 for organisations that would otherwise build private GPU clusters.

Watch for power-aware AI architectures. The energy constraint is driving research into more efficient model architectures (smaller MoE models, quantisation, distillation) that deliver comparable capability with less compute — and thus less power. DeepSeek's architectural efficiency innovations are partly a response to China's limited access to GPU compute; the same efficiency approach reduces energy consumption globally.

The power wall is real, it is growing, and it will constrain AI infrastructure expansion for the next decade. The question is not whether AI development is energy-intensive — it is. The question is where the power comes from, who pays for it, and whether the efficiency improvements from better model architectures keep pace with the raw demand growth.