AI power consumption is rising rapidly as modern data centers scale to support increasingly complex AI systems. Training large models and running billions of real-time inference requests now requires massive amounts of electricity, pushing global energy infrastructure to its limits.
The plant that nearly melted down in 1979 is coming back online. This time to power Microsoft’s AI.
Three Mile Island, Pennsylvania. Unit 1 is being refurbished under a 20-year deal between Microsoft and Constellation Energy. The 837-mw facility is expected to restart by 2028, and Microsoft’s going to take 100% of its output. That’s a signal that the energy math behind AI has broken down in ways that renewables alone can’t fix.
The deeper you go into how AI infrastructure actually consumes power, the more obvious it becomes that nuclear is not a trend. It’s a structural response to a structural problem.
Here’s what that problem actually looks like.
Key Takeaways
- AI power consumption is projected to grow by over 160% by 2030 because of increasing AI workload.
- Renewables paired with storage cover roughly 80% of a data center’s power needs. The other 20% requires reliable baseload generation.
- Nuclear provides 24/7, near-zero direct carbon emissions, and stable long-term costs.
- Microsoft, Google, Amazon, Meta, and Oracle have all signed nuclear deals since 2024. Big tech has committed to over 10 GW of new US nuclear capacity in the past year.
- SMRs are the longer-term bet. Faster to build, right-sized for hyperscale campuses.
- Less than 10% of the nuclear capacity needed by 2030 will actually be available by then. Nuclear is necessary but not sufficient.
What Is AI Power Consumption and Why Is It Increasing So Fast?
AI power consumption covers all electricity used to train, run, and support AI systems, including servers, networking, and cooling infrastructure.
A decade ago, data center workloads mostly served web pages and ran databases. Modern AI workloads, training large language models and running real-time inference at scale, are a different category of compute entirely. The chips are different, the density is different, and the power draw per rack is different by an order of magnitude.
AI power consumption is rising rapidly because AI is now embedded in search, productivity tools, coding assistants, etc. Every inference request draws power. At the scale of billions of daily queries, the energy requirements add up faster.
How Much Electricity Do AI Systems Actually Use?
AI power consumption becomes more visible when you look at how much electricity modern AI systems actually use.
GPU power consumption explained
A traditional server rack draws around 5–10 kW. A high-density AI GPU cluster running NVIDIA H100S can push more than 40 kW per rack. An individual H100 has a thermal design power of 700W; a single DGX H100 system draws roughly 10.2 kW.
Multiply that by thousands of servers per facility, and you’ll get the picture of what GPU power consumption looks like.
Training vs. inference energy use
Training a large model is a one-time event but power-intensive. GPT-4-class training runs consume tens of gigawatt-hours, roughly what a small city uses in a few days.
Inference is continuous, and it scales with adoption. This is why the AI energy problem doesn’t plateau; it compounds.
Why Hyperscale Data Centers Are Energy Intensive
Hyperscale data centers, typically 100,000 sq. feet and above, are operated by companies like AWS, Microsoft Azure, Google Cloud, and Meta. They’re massive parallel compute, with tens of thousands of servers running in coordinated GPU clusters. That architecture is inherently power-intensive.
Hyperscale infrastructure is one of the main drivers behind rising AI power consumption, as higher compute density directly increases electricity demand.
Cooling alone accounts for 30–40% of total data center electricity use. High-density AI compute generates significantly more heat per sq. meter than traditional workloads. So the cost of cooling increases with every expansion. Liquid cooling and immersion cooling are being deployed.
On top of that, hyperscale data centers can’t power down. Downtime costs millions per hour. Data center power usage has to come from a source that delivers an uninterrupted supply year-round. That “always-on” requirement is exactly where the case for nuclear power becomes hard to argue with.
The Growing Gap Between AI Power Consumption and Power Supply
AI energy demand is rising faster than power infrastructure can expand, creating a gap between electricity supply and the needs of hyperscale data centers.
The grid was not built for this.
Goldman Sachs Research estimates 85–90 GW of new nuclear capacity would be needed to meet all projected data center power demand growth by 2030. The global nuclear pipeline coming online by then covers less than 10% of that.
Grid interconnection queues are backed up for years. Transmission infrastructure takes a decade to permit and build, compared to renewable capacity, which expands rapidly. But the expansion is intermittent.
Why Renewable Energy Alone Isn’t Enough for AI Power Consumption
Renewables are already a significant part of hyperscaler energy supply. Goldman Sachs Research forecasts that 40% of new capacity built to meet data center demand will be renewables. The problem isn’t the source. Its availability.
Utility-scale solar runs roughly 6 hours per day on average. Wind averages about 9 hours. Battery storage helps, but at hyperscale, the economics get complicated quickly.
Wind and solar paired with storage can serve roughly 80% of a data center’s power demand. The other 20% needs something that runs regardless of the weather. That’s baseload power. Nuclear is the lowest-carbon option for it.
Why Big Tech Is Using Nuclear Power for AI Data Centers
Reliability and baseload power
Nuclear plants run at roughly 90–93% capacity factor. Solar runs around 25%, wind around 35%. For a facility that needs 24/7 power with no degradation, nuclear is the most reliable option available. The operational requirement for always-on power at hyperscale essentially excludes anything that depends on sunlight or wind.
Sustainability and emissions
Nuclear has near-zero CO2 emissions. Tech companies have aggressive net-zero targets while expanding their power footprint by hundreds of MW.
A scenario where 60% of data center demand growth is met by natural gas would add 215–220 million tons of global emissions, about 0.6% of global energy emissions. Nuclear avoids that cost.
Long-term cost stability
Goldman Sachs Research puts the cost of large-scale onsite nuclear at around $77/MWh with a $100/ton carbon price factored in, cheaper than a near-100% renewable solution ($87/MWh) or natural gas combined cycle ($91/MWh) under the same scenario.
Uranium prices are more stable than gas prices. For a company locking in a 20-year PPA, that predictability matters.
Nuclear vs Renewable vs Gas: Which Powers AI Best?
| Energy Source | Reliability (24/7 Power) | Carbon Emissions | Cost Stability | Scalability for AI | Overall Suitability |
| Nuclear Power | Very High (90%+ capacity factor) | Very Low | High | High (long-term) | Excellent |
| Solar & Wind | Low (intermittent) | Very Low | Medium | Medium (needs storage) | Partial |
| Natural Gas | High | High | Low (price volatility) | High (short-term) | Transitional |
Nuclear power is currently the most suitable energy source for AI data centers because it provides reliable 24/7 electricity with low carbon emissions, unlike renewables, which are intermittent, or natural gas, which increases emissions.
What Are Nuclear Power and Data Centers?
Nuclear power data centers get their electricity from nuclear generation through one of three models: a power purchase agreement with an existing plant, co-location next to one, or integrating SMRs directly into the campus.
Microsoft’s Three Mile Island deal is the PPA model: 100% of the plant’s 837 MW output for 20 years. Amazon took the co-location approach, acquiring a campus adjacent to Talen Energy’s Susquehanna plant for $650 million, with the logic of putting the compute next to the reactor. FERC rejected the interconnection permit twice before the deal was restructured.
The third model, SMRs purpose-built for data center campuses, is where the industry is heading.
The Rise of SMR Data Centers
What are SMRs?
Small modular reactors or SMRs are nuclear reactors with an output below 300 MWe (megawatts electric). They’re physically smaller than conventional reactors, designed for factory manufacturing and modular deployment, and intended to be faster and cheaper to build than traditional large-scale nuclear plants.
Several different designs are in development, using different reactor types, like light water, gas-cooled, and molten salt. The unifying idea is modularity. Build units, stack capacity, and deploy closer to the point of use.
Why do they fit the AI infrastructure?
Three things make SMRs a better fit for AI infrastructure than regular reactors:
- Right-sizing: A 100–300 MW reactor matches a large data center campus. Traditional plants produce 1,000+ MW, which creates the complexity of grid integration.
- Proximity: SMRs can be sited closer to demand, cutting transmission losses.
- Scalability: Extra modules can be added incrementally as the power requirement increases.
Google signed a deal with Kairos Power to deploy 500 MW of advanced nuclear capacity through a fleet of SMRs by the early 2030s, with the first reactor expected online by 2030. Oracle announced plans for a gigawatt-scale data center powered by three SMRs, with CEO Larry Ellison stating that building permits were already secured.
Deployment and safety
No commercial SMR has been deployed yet in the US. Kairos received a construction license for its 35 MWe Hermes demonstration reactor in Oak Ridge, Tennessee. It’s also the first new US nuclear construction license in decades. The progress is meaningful, but it’s still a demonstration project, and not a commercial rollout.
Advanced SMR designs rely on passive safety systems that improve on older reactor profiles. But real-world operational data is limited, and investors are watching the demonstration projects carefully before committing to full-scale deployment.
Big Tech Companies Investing in Nuclear Energy
Here is where the major players stand as of early 2026:
| Company | Deal | Capacity | Timeline |
| Microsoft | Three Mile Island restart (Constellation Energy) | 837 MW | 2028 |
| Amazon | X-Energy SMR + Susquehanna campus | 320 MW (phase 1), up to 960 MW | Early 2030s |
| Kairos Power SMR fleet | 500 MW | First reactor ~2030 | |
| Meta | Constellation Energy (Illinois nuclear plant) + open RFP | 1–4 GW target | Early 2030s |
| Oracle | Three-SMR campus | ~1 GW | Unspecified |
In the US alone, big tech has signed contracts for more than 10 GW of possible new nuclear capacity in the past year. Goldman Sachs Research sees potential for three new plants to come online by 2030. That’s not the 85–90 GW that would be needed to fully meet demand, but it’s a real infrastructure commitment, not just a roadmap.
Environmental Impact: Is Nuclear the Green Solution?
Nuclear vs. other fuels
The lifecycle CO2 emissions for nuclear power are roughly 12 gCO2eq/kWh. That’s way lower than natural gas (490 gCO2eq/kWh) or coal (820 gCO2eq/kWh). For data center operators with strict climate targets, that gap’s huge.
Using nuclear power also avoids the emissions from natural gas peaking plants that fill renewable intermittency gaps. A fully renewable solution for a hyperscale campus still costs more per MWh than a large-scale nuclear plant. Nuclear isn’t just cleaner than gas. It’s cost-competitive with renewables when the full system is costed honestly.
Waste and public concerns
Nuclear waste is the real concern. The US has no permanent repository for high-level nuclear waste, and all existing plants store spent fuel on-site in dry casks. SMRs have better passive safety profiles than 1970s light water reactors, but it isn’t risk-free.
The environmental case for nuclear is strong, but it’s more complicated regarding waste, water use, and land use. Anyone writing nuclear off is ignoring the emissions math. Anyone calling it fully clean is ignoring the waste problem.
Challenges of Using Nuclear Power for AI
The case for nuclear is real. So are the obstacles.
- Lead time: Building a new nuclear plant takes 10–20 years from planning to operation. The AI energy crunch is happening now. Restarting existing plants (Three Mile Island) and SMR demonstration projects are faster paths, but still have multi-year timelines.
- Cost overruns: Large nuclear construction projects have a poor track record on budget and schedule globally. Vogtle Units 3 and 4 in Georgia came in at roughly $35 billion, more than double the original estimate.
- Specialized labor: Nuclear construction and operation require a workforce that doesn’t exist at the scale needed. Scarcity of specialized labor is one of the primary constraints Goldman Sachs analysts identify in nuclear expansion.
- Uranium supply: Permitting and uranium sourcing are an active bottleneck for new nuclear development. Significant portions of enriched uranium are concentrated in Russia and Central Asia.
- Regulatory complexity: FERC has already rejected direct nuclear-to-data-center interconnection agreements twice. Regulatory frameworks weren’t designed for these arrangements.
Alternatives to Nuclear for AI Power Needs
Nuclear is not the only option being pursued. The AI power consumption problem probably requires all of the following to run simultaneously:
- Natural gas with carbon capture: The near-term default. High emissions without CCS; expensive with it. Some companies are running gas now with SMR conversion built into their contracts.
- Offshore wind and long-duration storage: Promising but unproven at hyperscale, with high interconnection costs.
- Geothermal: Reliable baseload and low emissions, but geographically constrained.
- Fusion: Microsoft has contracted for future fusion power. It’s real, but not before 2035 in any scenario.
- Demand-side efficiency: Still the fastest lever. More on that below.
How Companies Are Building Energy-Efficient Data Centers
Energy-efficient data centers are not just about switching to greener sources. Managing AI power consumption on the demand side reduces how much generation capacity is needed in the first place.
Cooling innovations
Liquid cooling, immersion cooling, and direct-to-chip cooling are all seeing active deployment in high-density AI facilities. Liquid cooling can cut cooling-related electricity use by 30–50% compared to traditional air cooling. Companies like Meta, Google, and Microsoft are retrofitting existing facilities and designing new ones with liquid cooling as the default.
Efficient hardware
The gap between GPU generations in performance-per-watt is significant. NVIDIA’s H100 delivers roughly 3–4x better energy efficiency than the A100 for typical AI workloads. The B200 and future generations push that further. Hardware that refreshes at scale has a material impact on data center electricity trends. However, the efficiency gains are being partially offset by expanding deployment volumes.
Workload optimization
Scheduling AI inference and training jobs during off-peak grid hours, routing workloads to facilities with lower-carbon grids. And could also use smaller fine-tuned models instead of large general-purpose ones. Google’s Carbon-Intelligent Computing Platform has been shifting data center compute loads to match renewable availability since 2020. It’s not a complete solution, but it’s not trivial either.
Future Outlook: Will Nuclear Power AI Grow?
Partially, and later than most headlines imply.
The Three Mile Island deal delivers power in 2028. Most SMR projects are targeting the early 2030s. The AI energy crunch is happening now, in 2025, and 2026. The gap is being filled with natural gas, renewables, and grid purchases, not nuclear.
Goldman Sachs Research projects that nuclear and AI efficiency gains together could start reducing the carbon footprint of AI data centers in the 2030s. What’s not in doubt is that the AI energy demand problem is structural, and the companies building the most capable AI systems have concluded the math doesn’t work without nuclear in the mix.
Final Thoughts
When Microsoft agreed to take 100% of Three Mile Island’s output for 20 years, the striking part wasn’t that they needed more power. It was how far up the risk curve they were willing to go. A plant with a complicated history, a multi-billion dollar refurbishment, and a commitment stretching to 2048.
That’s not a hedge. That’s a company that looked at the AI power consumption math and concluded nuclear power data centers aren’t optional.
Hyperscale data centers will keep expanding. Data center electricity consumption will keep climbing. Renewables will keep growing, but won’t cover the baseload gap alone. Nuclear, with all its complications, is the only zero-carbon baseload option at the required scale.
The trajectory of AI power consumption makes it clear that energy infrastructure decisions made today will shape the future of computing.
SMR data centers are a 2030s story. The commitments being made right now determine whether that story arrives on schedule.
FAQs
AI workloads require massive parallel GPU compute for training and continuous inference. Usage is now embedded across search, enterprise software, and consumer apps simultaneously. Efficiency improvements haven’t kept pace.
Renewables with storage cover about 80% of demand, but solar averages 6 hours of generation per day and wind about 9. Data centers need 24/7 power. Nuclear is the only low-carbon source that provides it reliably.
On carbon, nuclear lifecycle emissions run about 12 gCO2eq/kWh, comparable to wind. The harder issue is waste. No permanent US repository exists, and spent fuel stays on-site. Significantly cleaner than natural gas. Not without environmental cost.
Training large AI models can consume tens of gigawatt-hours, while ongoing inference adds continuous, long-term energy demand.
Yes. AI-driven data center growth is expected to significantly raise global electricity demand over the next decade.
