Exowatt用SageMaker提升AI模型精度

Exowatt builds modular, dispatchable solar systems for energy-intensive applications like AI data centers. Its Exowatt P3 captures solar energy through lenses, stores it as heat in a thermal battery, and dispatches electricity on demand via a heat engine.

Unlike grid-tied systems, P3 powers data centers directly, bypassing multi-year grid interconnection queues. Exowatt deploys AI agents to monitor and triage live alarms and support operational decisions, especially at the edge with limited connectivity.

As its fleet scales from test benches to manufacturing and field units, the intelligence layer must scale accordingly.

Exowatt used Amazon SageMaker AI serverless model customization to build a repeatable AI pipeline, producing task-specific fine-tuned models for edge and cloud deployments, achieving higher accuracy than frontier models at up to 49× lower cost.

These models are initial outputs, with the more durable result being the pipeline itself: a repeatable path from raw operational data to a certified training set, fine-tuned model, and evaluated deployment artifact, now runnable in days instead of months.

Alarm triage and energy-market analysis are examples of narrow, high-value tasks where this approach applies. The challenge was deploying AI on physical energy infrastructure.

Exowatt’s energy management system runs across growing operational sites, each producing continuous alarms. As the fleet scales to thousands of field units, human operators can’t monitor every site in real time.

Exowatt needed an AI agent to autonomously read alarm states and recommend actions. Frontier models couldn’t solve this due to the task’s narrowness, proprietary requirements, edge deployment needs without GPUs or internet, and unscaled costs.

Exowatt turned to model customization. Initial fine-tuning workflows were slow and required manual GPU selection, blocking experimentation. Exowatt needed a compact, task-specialized model deployable at the edge and cloud.

Solution overview: Exowatt adopted a cloud-native, serverless model customization approach on Amazon SageMaker AI, creating a pipeline that produces fine-tuned models for edge and cloud deployments, achieving higher accuracy than frontier models at up to 49× lower cost.