Why Microsoft and Google Are Buying Nuclear Reactors for AI Datacenters

In September 2024, Microsoft signed a 20-year power purchase agreement to restart the Three Mile Island Unit 1 reactor — a deal worth an estimated $16 billion over its lifetime. Google followed with commitments to purchase power from Kairos Power's small modular reactors (SMRs). Amazon has invested in X-energy and acquired a nuclear-powered datacenter campus from Talen Energy. These are not speculative bets; they represent a fundamental strategic conclusion: AI's energy appetite cannot be satisfied by intermittent renewables alone.

This article examines the energy economics driving hyperscalers toward nuclear power, the technology landscape, and the implications for enterprise AI infrastructure strategy.

The Energy Problem Nuclear Solves

AI datacenters present a unique load profile that exposes the limitations of renewable energy:

24/7 baseload demand: GPU clusters run continuously at 85-95% utilization during training
Massive scale: Single AI campuses now require 500 MW-2 GW — equivalent to a city of 400,000-1,600,000 people
Grid stability: AI loads cannot tolerate voltage fluctuations or intermittent supply
Geographic constraints: Datacenters locate where fiber, land, and talent exist — not necessarily where renewables are abundant

The Renewables Gap for AI Workloads

Energy Source	Capacity Factor	Availability	Carbon Intensity	Cost (LCOE $/MWh)	AI Suitability
Solar PV	20-30%	Daytime only	20-50 gCO2/kWh	$30-50	Low (intermittent)
Onshore wind	25-45%	Variable	7-15 gCO2/kWh	$25-50	Low (intermittent)
Solar + storage (4h)	60-70%	Extended	30-60 gCO2/kWh	$60-90	Moderate
Natural gas (CCGT)	85-90%	On-demand	350-450 gCO2/kWh	$45-70	High (but carbon)
Nuclear (existing)	90-93%	24/7 baseload	5-12 gCO2/kWh	$30-40	Excellent
Nuclear (new SMR)	90-95%	24/7 baseload	5-12 gCO2/kWh	$70-120	Excellent

Sources: Lazard LCOE 2024; EIA Annual Energy Outlook; IAEA PRIS Database

Nuclear's 90%+ capacity factor with near-zero carbon emissions makes it uniquely suited to AI's 24/7 baseload demand. No combination of solar and wind, even with battery storage, matches this profile economically at GW scale without significant overbuilding.

The Hyperscaler Nuclear Strategy

Microsoft: Three Mile Island Restart

Microsoft's agreement with Constellation Energy to restart Three Mile Island Unit 1 (a pressurized water reactor, distinct from the damaged Unit 2) provides 835 MW of carbon-free baseload power. Key deal parameters:

Capacity: 835 MW (enough for ~6 large AI datacenter campuses)
Duration: 20 years (2028-2048)
Estimated cost: $16 billion total ($95-110/MWh estimated)
Timeline: Restart targeted for 2028
Carbon offset: Eliminates approximately 5.6 million tonnes of CO2 annually vs. gas baseline

Microsoft has also invested in Helion Energy (fusion) and explored SMR partnerships, indicating a portfolio approach to 24/7 carbon-free energy (24/7 CFE).

Google: Kairos Power SMRs

Google's commitment to purchase power from Kairos Power's fluoride-salt-cooled SMRs represents a bet on next-generation nuclear technology:

Technology: Kairos Hermes — fluoride-salt-cooled, high-temperature reactor
Initial deployment: 75 MWe demonstration by 2027
Scale target: 500 MW+ by 2030
Location: Pacific Northwest (leveraging existing transmission)
Differentiation: Higher operating temperature enables more efficient cooling of adjacent datacenters

Google's approach is more aggressive technologically but offers potentially lower long-term costs if the SMR technology matures as projected.

Amazon: Talen Energy Acquisition + X-energy

Amazon acquired a 960 MW nuclear-powered datacenter campus directly adjacent to Talen Energy's Susquehanna nuclear plant in Pennsylvania. This co-location model eliminates transmission losses and grid interconnection delays:

Capacity: 960 MW direct from adjacent nuclear plant
Investment: $650 million campus acquisition
X-energy investment: $500 million for SMR development
Strategy: Co-locate datacenters directly at nuclear generation sites

The Economics of Nuclear for AI

The economic case for nuclear AI power rests on several factors that differ from general electricity markets.

Total Cost of Energy Comparison for AI Datacenters

Metric	Grid Power	Dedicated Solar+Storage	Gas (PPA)	Nuclear (existing)	Nuclear (new SMR)
LCOE ($/MWh)	$60-100	$70-100	$45-70	$30-40	$80-120
Capacity factor	99.9% (grid)	65-75%	85-90%	90-93%	90-95%
Carbon intensity	300-600 gCO2	30-60 gCO2	350-450 gCO2	5-12 gCO2	5-12 gCO2
24/7 CFE match	40-60%	65-75%	0%	90-93%	90-95%
Land requirement (per GW)	N/A	5,000-10,000 acres	30 acres	500-1,000 acres	50-100 acres
Interconnection timeline	3-5 years	2-4 years	2-3 years	1-2 years (restart)	5-8 years (new)
Price volatility	High	Low (fixed PPA)	Moderate (gas prices)	Very low	Very low

Sources: Lazard LCOE 2024; EIA; MIT Future of Nuclear Energy Study; NEI (Nuclear Energy Institute)

The Hidden Economics: Interconnection and Reliability

For AI datacenter operators, the headline LCOE comparison understates nuclear's advantage because:

Interconnection queue elimination: The average US interconnection queue wait is now 5+ years. Co-locating with existing nuclear plants bypasses this entirely.
Grid congestion avoidance: In markets like PJM (Eastern US), congestion charges add $10-30/MWh to effective prices. Behind-the-meter nuclear eliminates these.
Reliability premium: AI training runs that fail due to power interruptions waste millions in compute. Nuclear's 90%+ availability vs. grid outage risk has direct cost implications.
Carbon credit value: Nuclear generates zero-emission energy eligible for carbon credits ($15-50/tCO2 in compliance markets), offsetting 15-30% of power costs.

Small Modular Reactor Technology Landscape

SMRs represent the next generation of nuclear technology, designed specifically for distributed deployment scenarios like AI datacenters.

Leading SMR Designs for AI Datacenter Applications

Company	Design	Output (MWe)	Coolant	Timeline	Key Investor
NuScale	VOYGR	77 per module	Light water	2029-2030	Fluor Corp
Kairos Power	Hermes	75 (demo)	Fluoride salt	2027 (demo)	Google
X-energy	Xe-100	80 per module	Helium (HTGR)	2029	Amazon, DOE
TerraPower	Natrium	345	Liquid sodium	2030	Bill Gates, DOE
Rolls-Royce SMR	UK SMR	470	Light water	2031	UK Government
Last Energy	PWR-20	20 per unit	Light water	2026	Private

Sources: Company filings; DOE ARDP program; NRC licensing database

The modular nature of SMRs aligns with datacenter scaling: operators can add reactor modules as GPU deployment grows, matching power supply to demand without massive upfront capital for capacity that may sit idle.

Grid Impact and Regulatory Challenges

The rush to secure nuclear power for AI datacenters has triggered regulatory and grid reliability concerns.

Grid Reliability Concerns

When hyperscalers contract directly with nuclear plants for behind-the-meter power, the surrounding grid loses that baseload capacity. In the Talen Energy/Amazon case, the Federal Energy Regulatory Commission (FERC) intervened, concerned that diverting 960 MW from the grid would increase prices and reduce reliability for other customers.

Key regulatory questions include:

Should existing nuclear capacity serve AI datacenters or the general grid?
Do behind-the-meter arrangements unfairly shift grid costs to other ratepayers?
Should new nuclear construction receive expedited permitting for AI applications?
How should states balance economic development (datacenter jobs) against grid reliability?

NRC Licensing Timeline

The Nuclear Regulatory Commission's licensing process remains the primary timeline risk for new nuclear construction. Current processing times:

License Type	Typical Duration	Status for AI-Related Projects
Existing plant restart	2-3 years	TMI Unit 1 application filed
Existing plant uprate	1-2 years	Multiple applications pending
New SMR design certification	3-5 years	NuScale certified; others in review
New SMR construction permit	2-3 years	No AI-specific applications yet
Combined license (new site)	4-6 years	TerraPower Natrium in progress

The 24/7 Carbon-Free Energy Framework

Microsoft, Google, and others have adopted the "24/7 Carbon-Free Energy" framework, which goes beyond traditional renewable energy certificates (RECs) to match actual hourly consumption with local carbon-free generation.

Under this framework:

Hourly matching: Energy consumption must be matched with carbon-free generation in the same hour and same grid region
Additionality: New clean energy must be additional to what would have been built anyway
Deliverability: The clean energy must be physically deliverable to the datacenter location

Nuclear power is the only technology that consistently scores 90%+ on hourly matching metrics, making it the backbone of any credible 24/7 CFE commitment for AI workloads.

Strategic Implications for Enterprise CTOs

For enterprise technology leaders evaluating AI infrastructure options, the nuclear-AI convergence creates several strategic considerations. The same leaders focused on Kubernetes cost optimization today will need to consider power source economics in their total cost models:

1. Cloud Provider Selection

Cloud regions powered by nuclear energy offer the lowest carbon intensity for AI workloads. AWS's us-east-1 (Virginia, near Dominion nuclear plants) and GCP's us-central1 (Iowa, near Duane Arnold nuclear + wind) provide structurally lower emissions than regions dependent on fossil generation.

2. Long-Term Cost Trajectory

Nuclear-powered AI compute will likely carry a premium of 10-20% versus fossil-grid alternatives in the near term, but offers:

Price stability (no fuel price volatility)
Immunity from carbon pricing (which will increase over time)
Regulatory risk reduction (sustainability mandates)

3. On-Premise Nuclear Considerations

For organizations with sustained 50+ MW AI power needs, direct nuclear power procurement (via PPAs or co-location) may become viable by 2030 as SMR deployments scale. Early engagement with nuclear developers positions enterprises for priority allocation.

FAQ

Why are tech companies choosing nuclear over solar and wind for AI?

AI datacenters require 24/7 baseload power at 85-95% utilization. Solar (20-30% capacity factor) and wind (25-45%) cannot provide consistent power without massive overbuilding and expensive battery storage. Nuclear delivers 90%+ availability with near-zero carbon emissions, making it the only technology matching AI's load profile at scale.

How much does nuclear power cost for AI datacenters?

Existing nuclear plant power costs $30-40/MWh via long-term PPAs. New SMR construction is projected at $80-120/MWh initially, declining to $60-80/MWh at scale. These compare to grid power at $60-100/MWh (with volatility and carbon risk) and solar+storage at $70-100/MWh (with intermittency).

When will SMRs actually power AI datacenters?

The earliest SMR-powered AI datacenters are expected in 2028-2030. Microsoft's Three Mile Island restart targets 2028. Kairos Power's demonstration reactor targets 2027. Full commercial deployment of SMR fleets for AI is projected for 2030-2035.

Is nuclear power safe for datacenter co-location?

Modern reactor designs, including SMRs, incorporate passive safety systems that require no human intervention or external power for shutdown. The safety radius for SMRs is typically limited to the plant boundary, making co-location with datacenters architecturally feasible. The NRC regulates siting requirements.

How does nuclear AI power affect corporate carbon reporting?

Nuclear power produces 5-12 gCO2/kWh lifecycle emissions (including construction and fuel cycle), compared to 350-450 gCO2/kWh for natural gas. Organizations using nuclear-powered AI compute can report near-zero Scope 2 emissions for those workloads under GHG Protocol guidelines.

Conclusion

The convergence of AI energy demand and nuclear power represents one of the most significant infrastructure investments of the decade. Microsoft, Google, and Amazon are collectively committing $50+ billion to nuclear energy procurement and development — a signal that the industry's largest and most sophisticated buyers have concluded nuclear is essential to AI's future.

For enterprise CTOs, this shift has immediate practical implications: cloud region selection, sustainability reporting, long-term cost modeling, and infrastructure resilience all benefit from understanding the nuclear-AI nexus. The organizations that secure access to reliable, carbon-free power — whether through cloud provider selection or direct procurement — will hold a structural advantage in the AI era. For practical guidance on managing LLMs in production within these power constraints, see my separate guide.

Data sources: Constellation Energy/Microsoft PPA announcement (September 2024); Google/Kairos Power announcement (October 2024); Amazon/Talen Energy filing; FERC Docket ER24-2726; Lazard Levelized Cost of Energy v17.0; IEA Nuclear Power and Secure Energy Transitions (2024); NRC ADAMS document repository; Nuclear Energy Institute (NEI).

Why Microsoft and Google Are Buying Nuclear Reactors for AI Datacenters