The Numbers: How Much Power AI Actually Consumes
Artificial intelligence has transformed data centers from steady-state infrastructure into the fastest-growing category of electricity consumption in the developed world.
In 2026, US data centers consume approximately 4% of total national electricity — up from roughly 2% in 2020. AI workloads are the primary driver of this growth. A single H100 GPU draws 700 watts under full compute load. An 8-GPU server node consumes 10–12 kW. A rack of these nodes draws 80–140 kW. A 10,000-GPU training cluster consumes 10–15 megawatts — enough to power a small town.
The Scale of Growth
| Year | US Data Center Consumption | % of US Electricity | Primary Growth Driver |
|---|---|---|---|
| 2020 | ~80 TWh | ~2.0% | Cloud migration |
| 2022 | ~90 TWh | ~2.3% | Cloud + early AI |
| 2024 | ~120 TWh | ~3.0% | AI training boom |
| 2026 | ~170 TWh | ~4.0% | AI training + inference |
| 2028 (projected) | ~260 TWh | ~6.0% | AI inference dominance |
| 2030 (projected) | ~350–500 TWh | ~8–12% | AI as utility |
The projection to 8–12% by 2030 is not speculative — it is the direct consequence of current data center construction commitments. Over 100 GW of new data center capacity is in various stages of planning and development in the US alone. Each gigawatt of capacity requires a dedicated power source.
From Air to Liquid: Why GPU Cooling Changed Everything
Traditional data centers cooled servers with air — precision air conditioning units pushed cold air through server racks. This worked for CPUs drawing 150–300W. It does not work for racks drawing 80–140 kW.
The Cooling Reality for GPU Operators
Air cooling is dead for AI. Any deployment involving 8× H100 or Blackwell nodes requires liquid cooling. This immediately disqualifies most existing colocation facilities, which were built for 10–20 kW per rack with air cooling infrastructure.
Direct-to-chip liquid cooling is the current standard for H100 deployments. Cold plates mounted directly on GPUs and CPUs carry heat away via liquid loops. This handles racks up to 100 kW and is deployable as a retrofit in some existing facilities.
Immersion cooling — submerging entire servers in dielectric fluid — supports the highest power densities (140+ kW per rack) and is required for NVIDIA’s Blackwell NVL72 architecture, which packs 72 GPUs into a single liquid-cooled rack.
The cost implication: Liquid cooling infrastructure adds $500K–$2M per megawatt of capacity in capital costs. For a 10 MW GPU cluster, cooling infrastructure alone costs $5–$20 million. This is a significant barrier to entry for smaller GPU operators. For detailed operating cost breakdowns including cooling, see our GPU cluster economics guide.
The Power Wall: Why Electricity Is the New Bottleneck
In 2024, the primary constraint on AI compute was chip supply — NVIDIA could not manufacture H100s fast enough. In 2026, the constraint has shifted to electricity. GPUs are available; the power to run them is not.
Data Center Occupancy and Power Availability
Data center occupancy rates in major US markets are expected to exceed 95% by late 2026. This does not mean servers are full — it means electrical capacity is fully committed. New customers cannot deploy GPU racks because the facility has no remaining power allocation.
Key bottleneck metrics:
- Northern Virginia (the world’s largest data center market): Utility connection wait times now exceed 3–5 years for new large-scale deployments
- Grid interconnection queues: Over 2,500 GW of generation and storage projects are waiting for grid connection in the US — many specifically to serve data centers
- Transmission constraints: Even where generation exists, the high-voltage transmission lines to deliver power to data center sites are at capacity in many regions
Why Power Is Harder to Scale Than Chips
NVIDIA can double GPU production by expanding manufacturing at TSMC. Power generation and transmission infrastructure takes 5–15 years to build. A new natural gas plant takes 3–5 years. A nuclear plant takes 10–15 years. Transmission lines require years of environmental review and permitting.
This mismatch — GPUs that can be deployed in months versus power infrastructure that takes years — is the fundamental constraint on AI scaling for the rest of the decade.
Regional Power Economics: Where GPU Clusters Make Sense
Electricity prices vary 4× or more across US regions, making location the single largest variable cost factor for GPU operations. A GPU cluster in the wrong location can lose money; the same cluster in the right location can thrive.
| Region | Avg. Industrial Power Cost | Climate Cooling Benefit | Data Center Activity |
|---|---|---|---|
| Pacific Northwest (WA, OR) | $0.04–$0.06/kWh | Moderate (cool climate) | Growing rapidly |
| Texas (ERCOT) | $0.05–$0.08/kWh | Low (hot, but cheap power) | Fastest growing |
| Midwest (IA, OH, NE) | $0.06–$0.09/kWh | Good (cold winters) | Emerging market |
| Southeast (GA, TN, NC) | $0.07–$0.10/kWh | Low (hot, humid) | Established market |
| Northern Virginia | $0.08–$0.12/kWh | Moderate | Saturated — no new capacity |
| California | $0.12–$0.16/kWh | Good (mild climate) | Expensive but in-demand |
| Nordics (Europe) | $0.03–$0.06/kWh | Excellent (cold climate) | Growing for AI workloads |
The Dollar Impact
For a 1,000-GPU cluster running at 80% utilization:
- At $0.05/kWh (Pacific Northwest): ~$350,000/year in power costs
- At $0.10/kWh (Virginia): ~$700,000/year in power costs
- At $0.15/kWh (California): ~$1,050,000/year in power costs
The difference between the cheapest and most expensive locations is $700,000/year — for the same hardware running the same workloads. Over a 3-year GPU lifecycle, location choice alone accounts for a $2.1 million cost difference per 1,000 GPUs.
For GPU operators evaluating infrastructure decisions, power cost is not a secondary consideration — it is a primary driver of profitability. See our rent vs. buy analysis for how infrastructure costs factor into the ownership decision.
Nuclear, Solar, and SMRs: The Race for AI-Scale Energy
The mismatch between AI power demand and available electricity has triggered an unprecedented wave of energy investment — much of it explicitly for data centers.
Nuclear Power Agreements
Microsoft signed a deal to restart the Three Mile Island Unit 1 nuclear reactor, securing 837 MW of carbon-free power for its data centers. This single agreement provides enough power for a data center campus running hundreds of thousands of GPUs.
Google signed a contract with Kairos Power for small modular reactors (SMRs) to power its data centers. SMRs are factory-built, smaller-scale nuclear reactors that can be deployed in 3–5 years versus 10–15 years for traditional nuclear.
Amazon has acquired a nuclear-powered data center campus and invested in multiple SMR companies. Amazon’s strategy focuses on co-locating data centers directly adjacent to nuclear generation, eliminating transmission losses and grid constraints.
Why Nuclear Makes Sense for AI
Nuclear power provides what AI data centers need: 24/7 baseload generation that does not depend on weather, time of day, or fuel supply chains. A nuclear plant produces power continuously at a marginal cost of $0.02–$0.03/kWh — the cheapest non-hydro electricity available.
Solar and wind are cheaper per kWh when generating, but their intermittency is a problem for GPU clusters that run 24/7. Battery storage to smooth solar variability adds $0.02–$0.04/kWh, closing much of the cost gap with nuclear.
The Timeline Problem
Even the fastest new energy projects take years:
| Energy Source | Time to Deploy | Capacity | Baseload? |
|---|---|---|---|
| Solar farm | 1–2 years | 50–500 MW | No (30% capacity factor) |
| Natural gas plant | 3–5 years | 200–1,000 MW | Yes |
| Small modular reactor (SMR) | 4–7 years | 50–300 MW | Yes |
| Traditional nuclear | 10–15 years | 1,000+ MW | Yes |
| Geothermal | 3–5 years | 10–100 MW | Yes |
The implication: power constraints on AI compute will persist through at least 2028–2030, even with aggressive investment. This creates a structural advantage for GPU operators who have already secured power capacity in favorable locations.
Impact on GPU Pricing and Availability
Energy constraints affect GPU markets in two direct ways:
1. Limited deployment capacity keeps GPU rental rates elevated. If facilities cannot add more racks due to power constraints, GPU supply on marketplaces and cloud platforms grows more slowly than demand. This supports pricing.
2. Operating costs set a floor on rental rates. A GPU that costs $0.05/kWh to run in Oregon must be priced higher than a GPU running at $0.03/kWh in Norway. Power cost differences between operators flow directly into competitive pricing dynamics.
3. Location becomes a competitive moat. GPU operators with long-term power contracts at favorable rates — particularly those with access to nuclear, hydro, or low-cost natural gas — have a structural cost advantage that cannot be replicated by competitors in power-constrained markets.
Key insight: The AI industry’s growth rate is now fundamentally constrained by electricity, not silicon. Companies like NVIDIA can manufacture more GPUs, but there is nowhere to plug them in. This makes power access — securing megawatts at favorable rates in locations with adequate cooling — the key infrastructure asset of the AI era.
What This Means for GPU Providers and Renters
For GPU Providers and Operators
Secure power first, buy GPUs second. The most valuable asset in GPU operations is not the GPU — it is the power allocation. A 10 MW power contract at $0.05/kWh is worth more than the GPUs it runs, because GPUs can be upgraded but power contracts cannot be easily replicated.
Invest in cooling infrastructure. Facilities with liquid cooling can deploy 3–5× more GPUs per square foot than air-cooled facilities. The capital cost of cooling upgrades pays back quickly through higher GPU density and revenue per rack.
Prioritize efficiency. Power Usage Effectiveness (PUE) — the ratio of total facility power to IT equipment power — directly impacts profitability. A PUE of 1.1 (best-in-class liquid cooling) versus 1.5 (average air cooling) saves 27% on power costs. At scale, this is millions of dollars per year.
For operators considering building or expanding GPU infrastructure, see our cluster economics guide for complete cost modeling.
For GPU Renters and AI Companies
Provider location matters. Ask where your GPU instances are physically located. A provider in Texas at $0.06/kWh can offer lower rates than one in California at $0.14/kWh — for the same hardware. For current provider pricing comparisons, see our cloud GPU pricing guide.
Expect power-related pricing tiers. Some providers already charge different rates based on power availability. “Interruptible” instances (which can be paused during peak grid demand) cost less than guaranteed instances. This model will expand.
Plan for geographic constraints. If you need GPUs in specific regions (for data residency or latency), power availability in those regions directly affects your options. For the intersection of geographic constraints and sovereign AI requirements, see our sovereign AI analysis.
For an overview of the supply constraints that compound with energy limitations, see our GPU shortage analysis.
Frequently Asked Questions
How much electricity does training a large AI model use?
Training a frontier model like GPT-4 consumed an estimated 50–60 GWh of electricity — equivalent to powering roughly 5,000 US homes for a year. Next-generation models training on 50,000+ GPUs for months could consume 100+ GWh per training run. The electricity cost alone for a single frontier training run is $3–$8 million at typical industrial rates.
Why is liquid cooling required for AI GPUs?
An H100 GPU draws 700W under full load. Eight of them in a server node draw 10–12 kW. A rack of these nodes draws 80–140 kW. Air cooling physically cannot remove heat fast enough above approximately 30 kW per rack. Direct liquid cooling or immersion cooling is required to maintain safe operating temperatures at GPU-dense power levels.
Which regions have the cheapest power for GPU operations?
The Pacific Northwest (Washington, Oregon) offers the lowest US rates at $0.04–$0.06/kWh, primarily from hydroelectric power. The Nordic countries (Norway, Sweden, Finland) offer similar rates with excellent natural cooling. Texas and the Midwest offer moderate rates ($0.05–$0.09) with high availability. For the most comprehensive comparison, see our GPU guide which factors power efficiency into GPU selection.
Are nuclear power plants really being built for AI?
Yes. Microsoft’s deal to restart Three Mile Island Unit 1 (837 MW) is specifically for AI data center power. Google and Amazon have invested in small modular reactor (SMR) companies with explicit data center applications. These are not speculative — they are signed contracts with delivery timelines. The first SMR deployments for data centers are expected in 2028–2030.
Will energy constraints slow down AI progress?
In the short term (2026–2028), yes — energy is the binding constraint on AI compute scaling. However, the massive investment flowing into data center power ($50B+ in announced energy projects for data centers) will gradually relieve this constraint. By 2030, enough new generation capacity should be online to support the next wave of AI scaling, though demand will likely keep growing to fill any new supply.
Energy constraints keep GPU compute scarce and valuable. You can benefit from this dynamic by purchasing GPU packages on GPUnex starting at $59 to earn daily revenue from the sustained demand for compute capacity.