TerraPower Targets AI Data Centers with Natrium SMR and Molten Salt Thermal Storage

Nuclear technology developer TerraPower announced plans to finalize its first dedicated data center power project this year, positioning its Natrium sodium-cooled fast reactor architecture to meet the volatile power demands of artificial intelligence infrastructure. The project, slated to break ground in 2027, marks the company's second commercial deployment following its initial facility currently under construction in Kemmerer, Wyoming. In January, Meta signed an agreement with TerraPower to

2 min
TerraPower Targets AI Data Centers with Natrium SMR and Molten Salt Thermal Storage

Nuclear technology developer TerraPower announced plans to finalize its first dedicated data center power project this year, positioning its Natrium sodium-cooled fast reactor architecture to meet the volatile power demands of artificial intelligence infrastructure.

The project, slated to break ground in 2027, marks the company's second commercial deployment following its initial facility currently under construction in Kemmerer, Wyoming. In January, Meta signed an agreement with TerraPower to purchase power from eight Natrium reactor units.

TerraPower Molten Salt Thermal Storage and AI Load Balancing Architecture

The Load-Ramping Challenge in AI Compute

Data center power architectures face substantial operational strain due to the variable electrical loads imposed by large language model training clusters and real-time inference serving. Unlike conventional steady-state enterprise workloads, GPU clusters experience rapid power swings as compute cycles surge during training checkpointing or bursty inference queries.

Traditional baseload power generation struggles to accommodate these swings:

  • Conventional Nuclear Generation: Standard light-water reactors maintain high capacity factors (averaging 92.5 percent in the United States) but adjust output slowly, ramping at approximately 5 percent of rated capacity per minute according to the National Laboratory of the Rockies.
  • Gas Turbines: Rapid thermal cycling under sudden load fluctuations accelerates mechanical wear on turbine components.
  • Battery Storage Costs: Compensating for grid ramps using utility-scale battery banks introduces significant capital expenditures and maintenance overhead.

Molten Salt Thermal Buffering Architecture

TerraPower's 345-megawatt Natrium design decouples nuclear heat generation from electricity production by integrating a molten salt thermal energy storage system.

Under this architecture, the sodium-cooled reactor core operates continuously at full thermal output, maximizing capital efficiency and fuel utilization. During periods of lower data center demand, surplus thermal energy is routed directly into an insulated molten sodium and salt storage reservoir.

When AI workloads spike, the facility draws thermal energy from the storage reservoir to generate additional steam, boosting turbine output up to 500 megawatts for several hours without altering reactor core operating parameters. This thermal buffer absorbs short-term electrical transients, providing clean baseload capacity capable of responding to computational load shifts.

Commercial Deployment Timeline

As hyperscale operators seek dedicated power sources to bypass strained regional electric grids, small modular reactors (SMRs) with integrated storage offer a mechanism to deliver high uptime without fossil-fuel peaker plants. TerraPower expects to advance site selection and licensing for the data center facility ahead of the 2027 construction target.

Sources

Written by

More to read

  • Grokking in Large Language Models: How Weight Decay and Circuit Efficiency Drive Delayed Generalization

    Grokking in Large Language Models: How Weight Decay and Circuit Efficiency Drive Delayed Generalization In standard machine learning paradigms, model generalization closely tracks training loss: as an optimizer minimizes loss on training data, performance on held-out validation data improves in tandem until the model begins to overfit. In 2022, researchers at OpenAI observed a phenomenon that inverted this assumption: small neural networks trained on algorithmic tasks achieved near-zero trainin

    1 min
  • Serverless GPU Inference in Production: Cold Starts, GPU Memory Snapshotting, and Weight Paging Architectures

    Serverless GPU Inference in Production: Cold Starts, GPU Memory Snapshotting, and Weight Paging Architectures Deploying large language models on dedicated cloud GPUs creates an uncomfortable financial trade-off: keeping enterprise accelerators such as NVIDIA H100s or A100s warm 24/7 costs thousands of dollars per instance each month, yet scaling instances to zero introduces severe latency penalties. When traffic arrives at a dormant node, a standard inference server cold start can take anywhere

    1 min
  • Transformer Feed-Forward Networks as Key-Value Memories: How First-Layer Keys and Second-Layer Values Store Knowledge

    Transformer Feed-Forward Networks as Key-Value Memories: How First-Layer Keys and Second-Layer Values Store Knowledge In modern autoregressive Transformers, the division of labor between attention heads and multi-layer perceptron (MLP) blocks is often summarized through a clean functional split: attention routes information across sequence positions, while feed-forward networks (FFNs) process information per position. Yet for years, the exact mechanism by which FFNs process that information rem

    1 min