On 22 October, Beebop Labs hosted the first edition of The Open Problem at Wintercircus Ghent. The format is simple: a small, closed room with energy trading desks and industry and academic experts working on related problems. Three short statements, then open discussions. Around the table: trading desks and researchers from leading European energy companies and universities.

We wanted the conversations to explore what is happening, what is coming, and what is possible.

Compute just got started

Jan Fostier (imec) gave an overview of the hardware behind AI. His data showed the expressive power of frontier models rising roughly linearly with time: there is no evidence that progress is stagnating, nor are we seeing a self-reinforcing feedback loop (yet).

The compute being thrown at this is truly mind-boggling. The large frontier labs operate GPU clusters far larger than what European players such as Mistral have access to, and they just keep scaling up. Sovereignty concerns are real. Hardware is developing quickly on many fronts.

Powering the data center

How do we energise all this, in terms of generation and grid capacity? Where does it even fit? A natural segue to Ana Radovanovic (Google), one of the world’s leading experts on this topic, who presented data centres as grid-aware, flexible load.

There can be (some) flexibility in data centres, but tapping into it is not easy. Understanding and extending data centre flexibility is one of the topics Beebop Labs focuses on.

Speed matters

As all this energy needs to be bought and sold somewhere, David Schaurecker (ETH Zürich) presented recent results on high-frequency trading for grid-scale batteries on the continuous intraday market.

The core idea: every delivery period has its own order book, and every order-book update is a new decision. A battery that re-optimises its schedule after each update captures more value than one that decides every minute or every hour. Execution latency leads to missed orders and lost profit; acting on stale price signals leads to slippage.

The baseline in this line of work is rolling intrinsic, which reacts only to what is executable in the book now. David outlined two routes beyond it: adding forecast-based price opportunities to the current book, or estimating the continuation value of future trading through stochastic optimisation or value-based reinforcement learning, which we are exploring within Beebop and the AI4 Energy lab. This connects directly to the trade-off between mathematical programming and reinforcement learning we discussed in Q, f or a tree?.

With this many traders at the table, this one sparked the most discussion. One thing was clear: speed matters when trading.

Thanks to Jan, Ana and David, to the trading desks around the table, and to imec, Ghent University, and ID Lab. Beyond the formal programme, the informal conversations with everyone involved were excellent.

For research collaboration: labs@beebop.ai

Beebop Labs is where we foster the foundations that keep our orchestration platform state-of-the-art, for the power system of tomorrow rather than the one we operate in today (more here). Two of the fundamentals Labs spends its time on, trading integration and understanding data centre flexibility, were at the core of this first edition.