Exowatt makes modular solar power systems, called the P3, designed to supply reliable electricity directly to energy-hungry sites such as AI data centers, bypassing long grid-connection queues. Running these systems in the field means constantly handling alarms: a sensor out of range, a dropped connection, a component operating outside its normal limits. As Exowatt's fleet grows toward thousands of deployed units, no human team can watch every alarm in real time. The company's answer, described in an AWS engineering blog post, was to build AI agents that read alarm data and either act or recommend an action automatically. The interesting part is not just that it built an AI system, but how it built one that is both accurate and cheap enough to run at scale.

What is new

Exowatt reports that small, custom fine-tuned language models it trained outperformed much larger general-purpose ('frontier') models such as Claude Sonnet and Claude Opus on two specific jobs: triaging alarms at solar sites, and interpreting live electricity-market data to decide when to store or release power. The fine-tuned models were not only more accurate on these narrow tasks, they were also far cheaper to run, in one case by roughly 49 times. Exowatt built this using Amazon SageMaker AI's serverless model customization service, a managed way to fine-tune open-weight models without provisioning GPU servers or writing custom training code.

How it works

Fine-tuning means taking an existing, general-purpose AI model and further training it on a smaller, task-specific dataset so it specializes in one kind of job rather than trying to do everything. A common lightweight technique for this is called LoRA (Low-Rank Adaptation), which adjusts only a small part of a model's internal parameters instead of retraining the whole thing, making the process faster and cheaper. Exowatt combined LoRA with supervised fine-tuning, where the model is shown example inputs paired with correct answers, on SageMaker's serverless infrastructure, meaning the company never had to request or wait for specific GPU hardware; the cloud allocated capacity automatically and billed only for actual training time.

Exowatt built two models for two different jobs. The first, based on an open-weight model called Qwen3-1.7B, handles alarm triage directly on-site, on ordinary CPU hardware with no GPU and no reliable internet connection, because many solar installations have limited connectivity. The model was trained on a small, carefully checked dataset: 684 verified decisions about alarm scenarios, expanded through paraphrasing (having a larger model, Qwen3-235B, rewrite the wording while keeping the underlying facts and numbers fixed) into 2,522 examples. Each paraphrase was automatically checked to confirm it had not altered the decision, labels, or any figures; 1,838 of 2,218 candidate rewrites passed that check. Training took about 32 minutes and cost roughly 20 to 25 US dollars. The finished model was compressed ('quantized') to around 1.71 gigabytes so it could run inside a small software container on the site's existing server.

The second model, based on the larger Qwen3-14B, analyses real-time electricity market data from the Texas grid operator ERCOT to help decide when to store or dispatch energy. It was trained on data drawn from about 163 million historical records, automatically labeled into 25,000 examples and then expanded to 86,281 through paraphrasing by Anthropic's Claude Sonnet 4.6. Training ran as a single pass (one 'epoch') over the data and took about 8.5 hours of managed computing time; the company ran 23 training jobs over six days before settling on its best-performing version. The finished model is hosted through Amazon Bedrock's Custom Model Import feature, which only charges while the model is actively answering requests rather than sitting idle.

The results

Accuracy on these tasks is measured using a statistic called macro-F1, a standard way to score how well a model classifies things into categories, balancing correctness across all categories rather than just the most common ones; a score of 1.0 would mean perfect classification.

  • Edge alarm triage: fine-tuned Qwen3-1.7B scored 0.911 macro-F1 on 320 test cases, versus 0.698 for Claude Sonnet 4.6, 0.680 for Claude Opus 4.7 and 0.519 for Claude Haiku 4.5. The untuned base model could not produce valid answers at all.
  • Energy-market analysis: fine-tuned Qwen3-14B scored 0.975 macro-F1 on 2,000 test cases (0.968 on a later, separate test set), versus 0.625 for Claude Opus 4.7, 0.538 for Claude Sonnet 4.6 and 0.415 for the untuned base model.
  • Cost per 1,000 answers for the market-analysis model was about $0.60, versus $29.26 for Claude Opus 4.7, roughly a 49-fold difference.
  • The edge alarm model runs at effectively zero per-decision cost because it runs on the site's existing CPU hardware, with no metered API or GPU bill.
  • Both models were built to output structured answers with citations back to specific input data; unsupported claims were automatically discarded before reaching an operator.

Why it matters

The result challenges a common assumption that bigger, general-purpose models are always the safer or more capable choice. For narrow, repeatable, well-defined tasks, a much smaller model trained specifically on that task can be more accurate, and it avoids the practical problems of relying on a cloud-hosted frontier model: dependency on an internet connection, per-call costs that add up across thousands of sites running continuously, and a model that was never specialized for the company's own hardware and data in the first place. Exowatt's own account attributes most of the accuracy gain not to model size but to the quality and variety of training data. On the alarm-triage task, expanding 684 verified examples into 2,522 through varied paraphrasing lifted the score from 0.584 to 0.911, while simply repeating training over the same 684 examples for more passes only reached 0.879. The company also frames the real outcome of the project as the reusable pipeline itself, which it says can now take a new narrow task from raw data to a deployed, evaluated model in days.

Limits and open questions

The source describes results for two specific tasks on Exowatt's own infrastructure and data; it does not say how the approach would perform on less narrowly defined or noisier real-world tasks, nor how the models behave outside the shadow-mode testing mentioned for the alarm system. The comparisons are against particular versions of Claude models accessed through Amazon Bedrock at list prices, not against every possible frontier model or configuration. The post does not detail how the system handles cases genuinely outside its training distribution beyond noting that the edge model is designed to abstain rather than guess when evidence is insufficient.

How to try it

The underlying tools are generally available: Amazon SageMaker AI's serverless model customization service, which provides ready-made fine-tuning recipes for open-weight models from a public hub, and Amazon Bedrock, including its Custom Model Import feature for hosting fine-tuned models without maintaining dedicated servers. Exowatt's pipeline used seven AWS services in total, covering data storage, fine-tuning, evaluation, model registry and deployment, none of which required the company to manage its own compute instances or clusters directly.

Sources