Submit a tool
Nuclear Energy for AI Hyperscalers: Three Mile Island, SMRs & Clean Grids
Nuclear EnergyAI DatacentersHyperscalersClean EnergySMRConstellation

The Gigawatt Compute Wall and Grid Interconnection Bottlenecks

The rapid scaling of frontier foundation model training clusters has collided directly with the physics of global electrical grids. In 2023 and 2024, state-of-the-art supercomputing facilities like the xAI Colossus Cluster scaled to 100,000 NVIDIA H100 GPUs, drawing upwards of 150 megawatts of continuous electrical load. Next-generation clusters operating 300,000 to 1,000,000 accelerators are projected to require between 1 and 2 gigawatts (GW)β€”a power draw equivalent to that of a major metropolitan area.

For cloud hyperscalers (Microsoft Azure, Amazon Web Services, Google Cloud), securing this power through traditional public utility interconnections has become impossible: 1. PJM and Regional Transmission Queue Delays: Regional Transmission Organizations (RTOs) such as PJM Interconnection, ERCOT, and MISO face interconnection queues extending 5 to 7 years for new utility-scale solar and wind projects. 2. Intermittent Capacity Factors: Solar (20–25% capacity factor) and wind (30–45% capacity factor) cannot deliver the 99.999% baseload reliability required by distributed gradient-descent training runs, where a momentary voltage sag can crash an entire multi-node checkpoint. 3. Hyperscaler Net-Zero Pledges: Hyperscalers are bound by corporate commitments to match 100% of their operational energy with 24/7 carbon-free energy (CFE), ruling out open-cycle natural gas peaker plants as permanent primary power.

Faced with these structural limits, hyperscalers in late 2024 executed an unprecedented series of multi-billion-dollar commercial agreements to secure dedicated nuclear power.

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”‚ Hyperscaler Nuclear Energy Strategies β”‚ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β”‚ β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”Όβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β–Ό β–Ό β–Ό β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”‚ Microsoft & β”‚ β”‚ Amazon AWS & β”‚ β”‚ Google & β”‚ β”‚ Constellation β”‚ β”‚ Talen Energy β”‚ β”‚ Kairos Power β”‚ β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€ β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€ β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€ β”‚ - TMI Unit 1 β”‚ β”‚ - Susquehanna β”‚ β”‚ - 6-7 SMRs β”‚ β”‚ Restart β”‚ β”‚ Nuclear Campus β”‚ β”‚ (500 MW) β”‚ β”‚ - Crane Clean β”‚ β”‚ - 960 MW direct β”‚ β”‚ - Molten salt β”‚ β”‚ Energy Center β”‚ β”‚ behind-the- β”‚ β”‚ cooled (FHR) β”‚ β”‚ - 835 MW baseloadβ”‚ β”‚ meter access β”‚ β”‚ - Low-pressure β”‚ β”‚ - 20-yr PPA β”‚ β”‚ - FERC scrutiny β”‚ β”‚ coolant loop β”‚ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

---

Microsoft and Constellation Energy: The Crane Clean Energy Center

In September 2024, Microsoft signed a historic 20-year Power Purchase Agreement (PPA) with Constellation Energy to resurrect Unit 1 of the Three Mile Island nuclear generating station in Londonderry Township, Pennsylvania (*Constellation Energy Press Release: Crane Clean Energy Center*, Sept 2024).

Unit 1β€”an 837 MW pressurized water reactor (PWR) physically adjacent to the damaged Unit 2 reactor that suffered a partial meltdown in 1979β€”operated safely and reliably for 45 years until it was prematurely shut down in 2019 due to uneconomic competition from cheap shale gas.

Under the agreement, Constellation will invest approximately $1.6 billion to restore Unit 1, rebranding the facility as the Crane Clean Energy Center in honor of former CEO Chris Crane: - Capital Restoration: Complete overhaul of the reactor cooling loops, steam turbines, main electrical generator, and digital instrumentation and control (I&C) systems. - Regulatory Pathway: Re-licensing through the US Nuclear Regulatory Commission (NRC) under 10 CFR Part 50, aiming for commercial re-synchronization to the grid in 2028. - Operational Life: Re-licensing will extend operations to at least 2054, supplying 835 MW of 100% carbon-free baseload electricity directly contracted to offset Microsoft Azure datacenter workloads across PJM.

---

Amazon AWS and Talen Energy: Susquehanna Behind-the-Meter Colocation

While Microsoft pursued a virtual grid PPA, Amazon Web Services pursued physical, "behind-the-meter" colocation. In March 2024, AWS acquired the Cumulus Data Center campus from Talen Energy for $650 million.

The 1,200-acre Cumulus campus is located immediately adjacent to the Susquehanna Steam Electric Station in Luzerne County, Pennsylvania, a 2.5 GW dual-unit boiling water reactor (BWR): - Direct Interconnection: AWS contracted to receive up to 960 MW of continuous baseload nuclear electricity directly off the generator bus, bypassing PJM's regional transmission grid entirely. - Transmission Cost Elimination: Bypassing public transmission lines eliminates network transmission charges, line congestion risk, and interconnection queue timelines. - Regulatory Scrutiny by FERC: In late 2024, regional utilities (American Electric Power and Exelon) filed formal complaints with the Federal Energy Regulatory Commission (FERC), arguing that diverting 960 MW of existing nuclear power behind-the-meter shifts grid transmission upgrade costs onto residential and industrial ratepayers. FERC's ongoing rulings will set the legal precedent for all future behind-the-meter datacenter campuses.

---

Google and Kairos Power: Backing Small Modular Reactors (SMRs)

Recognizing that existing nuclear plants are a finite resource, Google signed a groundbreaking corporate master agreement with Kairos Power in October 2024 to finance the commercial deployment of a fleet of Small Modular Reactors (SMRs).

The agreement targets 500 MW of new carbon-free capacity across six to seven modular reactors, with the first unit scheduled to come online by 2030 and full fleet deployment by 2035: - Fluoride Salt-Cooled High-Temperature Reactor (FHR): Kairos utilizes low-pressure liquid FLiBe molten salt (a lithium fluoride and beryllium fluoride mixture) as a coolant rather than high-pressure water. - TRISO Fuel Pebbles: Fuel consists of Tri-structural Isotropic (TRISO) pebbles containing high-assay low-enriched uranium (HALEU) encased in silicon carbide and pyrolytic carbon, capable of withstanding temperatures exceeding 1,600Β°C without melting. - Passive Safety Physics: Because the molten salt operates at near-atmospheric pressure, the reactor cannot suffer high-pressure steam explosions or loss-of-coolant accidents. Decay heat is removed purely through natural circulation and radiation.

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”‚ TRISO Fuel Pebble Cross-Section (Kairos Power SMR) β”‚ β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€ β”‚ [ Uranium Oxycarbide Kernel (15.5% HALEU) ] β”‚ β”‚ └─ Porous Carbon Buffer Layer β”‚ β”‚ └─ Inner Pyrolytic Carbon (IPyC) Layer β”‚ β”‚ └─ Silicon Carbide (SiC) Pressure Vessel β”‚ β”‚ └─ Outer Pyrolytic Carbon (OPyC) Layer β”‚ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

---

Comparison of Hyperscaler Nuclear Approaches

| Parameter | Constellation / Microsoft | Talen Energy / Amazon | Kairos Power / Google | | :--- | :--- | :--- | :--- | | Project Model | Decommissioned Plant Restart | Operating Plant Behind-the-Meter | Greenfield Small Modular Fleet | | Capacity | 835 MW (Single Unit) | Up to 960 MW | 500 MW (Fleet of 6–7 Units) | | Reactor Technology | Pressurized Water Reactor (PWR) | Boiling Water Reactor (BWR) | Molten Salt High-Temp (FHR) | | Grid Connection | Virtual Grid PPA (PJM) | Direct Behind-the-Meter Bus | Localized / Microgrid / Grid-Fed | | Target Operational Date | 2028 | Immediate (scaling through 2027) | 2030–2035 | | Primary Regulatory Risk | NRC re-commissioning safety audit | FERC transmission cost hearings | NRC commercial SMR licensing & HALEU |

---

Technical & Economic Obstacles Facing Nuclear AI

While nuclear energy solves the baseload power constraint, engineering teams face significant headwinds: 1. HALEU Supply Chain Vulnerabilities: Advanced SMRs require High-Assay Low-Enriched Uranium enriched between 5% and 20% U-235. Historically, commercial HALEU was sourced primarily from Russian state supplier Tenex. The US Department of Energy (DOE) and domestic enrichers like Centrus Energy are investing billions to scale domestic enrichment cascades, but supplies remain constrained. 2. Thermal Dissipation Requirements: Large nuclear facilities discharge substantial waste heat into cooling towers or water bodies. When coupled with datacenter liquid cooling loops (examined in detail in our analysis of Thermal Physics of AI Datacenters), site water availability becomes a primary gating factor. 3. Capital Intensity & LCOE: While amortized nuclear operational costs are relatively low ($30–$40/MWh for existing plants), greenfield SMR construction costs are projected between $80 and $120/MWh. Hyperscalers are subsidizing these early FOAK (First-Of-A-Kind) premiums to ensure supply certainty for their cloud infrastructure on RunPod GPU Cloud and Together AI Inference.

Nuclear power has shifted from an energy afterthought to a core strategic pillar of frontier AI computing, anchoring the next decade of gigawatt datacenter architecture.

Enjoyed this article?
Share it with fellow builders and AI practitioners.
Share:𝕏 PostLinkedIn

More from the AI Journal

Hardware & Infrastructure
NVIDIA Blackwell Ultra & GB200 NVL72: Liquid-Cooled AI Megaclusters and Data Center Transformation

Analyzing NVIDIA's Blackwell architecture, the NVLink 5 interconnect, and why liquid cooling has become a non-negotiable requirement for next-generation frontier training....

Oct 1, 2026
Hardware & Infrastructure
ARM Neoverse V2 & NVIDIA Grace: Cache-Coherent AI Architecture

How cache-coherent CPU-GPU architectures like NVIDIA Grace (ARM Neoverse V2) eliminate PCIe bus bottlenecks during RAG vector retrieval, data ingestion, and multimodal token preprocessing....

Dec 22, 2024
Hardware & Infrastructure
Intel Gaudi 3: Integrated RoCEv2 Ethernet & Matrix Economics

Examining Intel's Gaudi 3 accelerator, analyzing its 24 integrated 200 GbE RoCEv2 ports, matrix processing engines, software ecosystem, and enterprise TCO economics....

Dec 20, 2024