AI Data Center Energy Consumption: Why Big Tech Is Racing Toward Nuclear Power

The Pulse

Global data centre electricity consumption is projected to rise from about 485 terawatt-hours in 2025 to roughly 950 TWh by 2030, according to the International Energy Agency. The IEA expects electricity use from AI-focused data centres to triple over the same period, even as the energy required per AI task falls. [IEA]

That growth is pushing Microsoft, Google, Meta, Amazon, and other technology companies beyond ordinary electricity procurement. They are signing decades-long nuclear power agreements, financing small modular reactors, supporting plant restarts, and investing directly in energy infrastructure. The race toward nuclear power is not replacing renewables, which the IEA expects to meet nearly half of additional data centre demand through 2030. It is a response to a different requirement: large AI facilities need reliable electricity every hour, including when wind and solar output falls. [IEA]

Core Significance

Why it matters

• Electricity is becoming an AI deployment constraint: Data centre demand grew 17% in 2025, while electricity use at AI-focused facilities increased even faster. The IEA also estimates that around 20% of planned data centre projects could face delays if grid and supply-chain bottlenecks are not addressed. [IEA]
• Big Tech is moving closer to the power source: Alphabet agreed to acquire Intersect for $4.75 billion plus assumed debt to expand data centre and generation capacity, while Amazon, Microsoft, Meta, and Google have signed long-term nuclear agreements. [Alphabet]
• Nuclear solves a reliability problem, not a speed problem: Existing reactors can provide continuous low-carbon power, but new reactors and small modular reactors take years to license and build. The IEA expects renewables and natural gas to supply most near-term growth, with nuclear becoming more important toward the end of the decade and after 2030. [IEA]

Deep Context

Data centre energy demand did not begin with generative AI. For years, cloud computing, streaming, enterprise software, and digital services expanded while improvements in chips, cooling, and server utilisation limited the rise in total electricity use.

Generative AI altered that balance by accelerating demand for specialised servers and high-density computing. The IEA projects that electricity use from accelerated servers, which is mainly driven by AI adoption, will grow by about 30% annually through 2030 in its base case. Accelerated servers account for nearly half of the projected increase in global data centre electricity use. [IEA]

The physical scale is also changing. The IEA says conventional data centres commonly use between 10 and 25 megawatts, while hyperscale AI centres can exceed 100 MW. The largest announced developments can require electricity comparable to millions of households. [IEA]

This creates a second infrastructure race alongside the competition for advanced chips. The Nvidia AI chip competition determines who controls compute economics. The energy race determines where that compute can actually be installed and operated.

Data Insights

By the numbers

• 485 TWh: Estimated global data centre electricity consumption in 2025. [IEA]
• 950 TWh: The IEA’s updated central projection for global data centre electricity demand in 2030, equal to around 3% of worldwide electricity consumption. [IEA]
• 1,200 TWh: The IEA base-case projection for global data centre electricity consumption by 2035. [IEA]
• 15% annually: The projected growth rate for global data centre electricity consumption from 2024 to 2030, more than four times faster than electricity demand from other sectors. [IEA]
• Up to 500 MW: The total capacity covered by Google’s agreement with Kairos Power for multiple small modular reactors, with the first project targeted for 2030 and additional deployments through 2035. [Google]
• 1,920 MW: The full contract quantity in Talen Energy’s expanded nuclear power agreement with Amazon, expected to ramp to full volume by 2032 and continue through 2042. [Talen Energy]
• 2,000 MW: The electricity allocation announced by Pakistan for bitcoin mining and AI data centres in the first phase of a national initiative. [Reuters]

Table 1: Big Tech nuclear power agreements supporting AI and data centre growth

Company	Agreement	Capacity	Timing	Source
Microsoft	20-year agreement supporting the restart of the Crane Clean Energy Center	835 MW	Constellation has targeted a return to service before the end of the decade	Constellation
Meta	20-year agreement for the Clinton Clean Energy Center	1,121 MW, including a planned 30 MW uprate	Begins June 2027	Constellation
Google	Multiple Kairos Power small modular reactors	Up to 500 MW	First project targeted for 2030; further deployments through 2035	Google
Google	First Kairos deployment with TVA at Oak Ridge	50 MW	Expected to supply grids serving Google data centres from 2030	Google
Amazon	Expanded PPA with Talen’s Susquehanna nuclear plant	Up to 1,920 MW	Full quantity no later than 2032; contract through 2042	Talen Energy
Amazon	X-energy and Energy Northwest SMR development	320 MW initially, expandable to 960 MW; broader 5 GW target by 2039	Early 2030s onward	Amazon

These agreements are not identical. Microsoft’s deal supports the restart of an existing reactor, Meta’s agreement helps preserve and expand an operating plant, Google is backing advanced reactors that have not yet reached commercial operation, and Amazon is combining existing nuclear supply with investment in future SMRs.

Table 2: AI data centre electricity demand compared with major power benchmarks

Benchmark	Electricity or capacity	Context	Source
Conventional data centre	10–25 MW	Typical range cited by the IEA	IEA
Hyperscale AI centre	More than 100 MW	Comparable to the annual electricity use of about 100,000 households	IEA
Largest under-construction data centre used in IEA comparison	About 2,000 MW	Equivalent to several million households in the IEA comparison	IEA
Global data centres in 2025	485 TWh	Starting point for the IEA’s updated 2030 outlook	IEA
Global data centres in 2030	About 950 TWh	Roughly 3% of global electricity demand	IEA
Global data centres in 2035	About 1,200 TWh	IEA base case after 2030 uncertainty widens	IEA

Business Case

Why nuclear appeals to hyperscalers

Nuclear power offers two features that are difficult to combine: continuous output and low operational carbon emissions. AI data centres run around the clock, and their electricity demand cannot always be shifted to periods of strong wind or solar generation.

Microsoft’s agreement with Constellation supports the return of 835 MW of nuclear generation at the Crane Clean Energy Center. Meta’s 20-year agreement supports the continued operation of the 1,121 MW Clinton plant and enables a 30 MW increase through uprates. [Constellation on Microsoft] [Constellation on Meta]

Google and Amazon are making a different bet by supporting small modular reactor development. Google’s Kairos agreement covers up to 500 MW, while Amazon says its X-energy investment is intended to help bring more than 5 GW of new nuclear capacity to the US grid by 2039. [Google] [Amazon]

Why nuclear cannot solve the near-term gap alone

The IEA expects renewables to meet nearly half of additional data centre electricity demand through 2030. Natural gas and coal are expected to supply more than 40% of the increase because they can often be added or dispatched faster than new nuclear capacity. Small modular reactors begin contributing meaningfully after 2030 in the IEA base case. [IEA]

Big Tech is therefore pursuing a portfolio strategy rather than a nuclear-only strategy: existing nuclear plants for continuous supply, renewables and storage for scale, natural gas for near-term reliability, and advanced reactors for the next decade.

Why energy ownership is becoming strategic

Alphabet’s agreement to acquire Intersect for $4.75 billion plus assumed debt illustrates how close the technology and energy sectors are becoming. Alphabet said the acquisition would help bring data centre and generation capacity online faster and included multiple gigawatts of projects under development or construction. [Alphabet]

This is more than sustainability procurement. A company that can coordinate compute, land, grid connections, generation, and storage can open capacity faster than a competitor waiting for a utility interconnection.

Expert Nuance

The phrase “AI data centre energy consumption” can be misleading when it is treated as a precisely measured global category. Data centres also support cloud computing, streaming, storage, enterprise applications, and conventional digital services. The IEA separates accelerated servers and AI-focused facilities where possible, but uncertainty remains because companies do not disclose complete workload-level electricity data.

Training is also not the only issue. Inference runs whenever users query models or agents complete tasks, turning electricity use into a recurring operating cost. The difference between AI training and inference power demand matters because inference can scale with billions of daily interactions even when individual tasks become more efficient.

Efficiency gains are substantial. The IEA says power use per AI task is falling rapidly, and Alphabet reported that it reduced Gemini serving unit costs by 78% during 2025 through model optimisation, efficiency, and utilisation improvements. Lower cost per task, however, can increase total use by making AI cheaper and more widely deployed. [IEA] [Alphabet]

Nuclear agreements also do not mean that a specific data centre receives electrons directly from one reactor at every moment. Many deals operate through regional grids and contractual clean-energy attributes. Their strategic value lies in financing continued or additional generation and securing long-term supply economics.

Strategic Outlook

1. Grid connections will remain the immediate bottleneck: The IEA estimates that about 20% of planned data centre projects could face delays unless grid, equipment, planning, and permitting constraints are addressed. [IEA]
2. Existing nuclear plants will be monetised first: Plant restarts, licence extensions, uprates, and long-term PPAs can add or preserve capacity sooner than entirely new reactor fleets.
3. SMRs face a commercial proof point around 2030: Google’s first Kairos-linked project targets 2030, while Amazon’s SMR plans also point to the early 2030s. Delays would increase dependence on gas and existing grids. [Google] [Amazon]
4. Energy infrastructure acquisitions will continue: Alphabet’s Intersect agreement shows that technology companies may increasingly buy or finance the organisations that develop generation and co-located data centre sites. [Alphabet]
5. Pakistan will need to distinguish surplus capacity from reliable capacity: Reuters reported a 2,000 MW allocation for bitcoin mining and AI data centres, but dependable data centre operations still require suitable locations, transmission, cooling, tariffs, and continuous power quality. [Reuters]

Key Question

Why is Big Tech turning to nuclear power for AI data centres?

Big Tech is turning to nuclear power because AI data centres require large amounts of reliable electricity around the clock. Nuclear plants can provide continuous low-carbon generation, making them useful alongside variable renewable sources such as wind and solar.

Microsoft is supporting the restart of an 835 MW reactor, Meta has signed a 20-year agreement supporting the 1,121 MW Clinton plant, Google has contracted for up to 500 MW from Kairos Power reactors, and Amazon has agreements covering existing nuclear supply and future SMR development. [Constellation] [Constellation] [Google] [Amazon]

Nuclear will not meet the full increase by itself. The IEA expects renewables to supply nearly half of additional data centre demand through 2030, while natural gas and coal cover a substantial share of the near-term gap. Nuclear becomes more important later because new projects require longer development timelines. [IEA]

The Takeaway

The AI infrastructure race is no longer only about chips, models, and data centres. It is increasingly about electricity, grid access, and the ability to bring reliable generation online at the same speed as new compute capacity.

Nuclear power appeals to hyperscalers because it offers continuous low-carbon output and long-term supply certainty. But the current nuclear deals cover several different strategies: preserving existing plants, restarting retired reactors, expanding output through uprates, and financing advanced reactors that may not operate until the next decade.

The near-term solution will remain mixed. Renewables, storage, natural gas, existing nuclear plants, efficiency gains, and better grid planning will all be required. The companies that coordinate these layers successfully will be able to deploy AI infrastructure faster. Those that secure chips but not power will discover that compute capacity is only useful when the electricity arrives.