Fuse is training an energy foundation model to be a 'superhuman energy trader', and says it already cuts demand forecast error to 1%
What's the deal? UK energy company Fuse Energy has started pre-training its own energy foundation model on its in-house customer demand data, co-founder and CEO Alan Chang said on 3 November. Two weeks in, the transformer model's demand forecast error is 1.00%, against 3.46% for Fuse's existing classical machine learning stack and 7.84% for what Chang calls the industry standard. That is a 3.5x improvement on Fuse's own best model and about 8x better than the industry, on the company's own figures. Chang called the results early but "the best I've seen in my career", and said the goal is an AI that works as a superhuman energy trader.
Why it matters: For a supplier, every point of forecast error is cost. Misjudge demand and you buy too much or too little power in a volatile wholesale market, and the margin goes. Fuse already forecasts at household level using a digital twin of each property, and in October Chang said the next step is a transformer model that co-optimises its whole portfolio, from utility-scale solar and gas to home batteries, EV chargers and heat pumps, against consumption and market prices.
Context: Founded in 2022 by former Revolut executives Alan Chang and Charles Orr, Fuse runs generation, trading and retail supply itself. Last week it said it is moving its AI workloads in-house onto NVIDIA hardware and open-weight models, starting with a customer support agent that resolves four in five chats without a human. In June it said it had reached profitability at a $550M revenue run rate. Chang is now hiring AI researchers for the model.
The signal: Vertically integrated energy challengers are starting to treat their own consumption data as training data, and build foundation models on it rather than buy forecasting from vendors.
Read more: Alan Chang (X), IT Brief UK, Fuse, Tech Funding News
Image credit: Fuse Energy / Alan Chang on X
