In July 2026 CGTN, China’s international broadcaster, sent a camera crew to our office in Vienna to film a segment on artificial intelligence in electricity networks. Our CEO Clemens Wasner did the interview. The piece aired on 21 July.
Why they called
The reason was four years old. In 2022 our reinforcement learning team entered Learning to Run a Power Network, the international competition run by RTE, the French transmission system operator, against a field that included Baidu. We came first.
That result is still how a great many people find us, and it is a fair summary of what the company is for. The competition demonstrated that an agent can search the space of grid switching options at a scale no human dispatcher can work through under time pressure. Everything we have built since has been an attempt to make that useful on a real network rather than a competition one.
What we explained
The question the crew kept coming back to was a simple one: why do the lights not flicker every time the wind drops?
The short answer is that somebody, somewhere, is constantly rebalancing the system — and that this job is getting harder. Wind and solar are excellent and entirely uncooperative. The more of them a grid carries, the more often it finds itself in a state nobody planned for, and the fewer of those states can be handled by an operator manually reaching for a known-good configuration.
What FLEX does is build a digital twin of the network and put a reinforcement learning agent on top of it, so that when congestion appears there is a ranked set of options ready — which switches to open, which transformer taps to move, which flexibility to call on, and what each choice costs. This is not an Austrian problem or a European one. Every grid on a decarbonisation path runs into it.
What we were careful to say
Our competition result is a benchmark result. On the competition’s network models the AlphaZero-based agent achieved up to a 60 percent reduction in required redispatching — on realistic simulated grids, not on a live distribution network.
We keep saying this because the distinction is the whole argument. Numbers from a real grid come from real-grid projects, and we publish them once they have been measured. Our current proof of concept runs on an Austrian distribution grid with real network and smart-meter data, and that is where the measured figures will come from.
Nor is any of this an autopilot. No European grid operator runs reinforcement learning in live control today. Our validated results are in day-ahead planning, on historical operator data, with a person making the final call.
What is actually at stake
Congestion is expensive. In Germany alone, grid congestion management cost €3.07 billion in 2025 — renewable curtailment compensation, conventional redispatch, reserve plant provision and deployment, and countertrading, according to the Bundesnetzagentur’s annual figures. That is money spent every year to work around capacity that in many cases physically exists but cannot be reached with the tools currently in the control room.
A loop closing
One footnote that made the visit feel less accidental than it looked: Clemens spent ten years living in Asia, including Beijing, before moving back to Austria in 2016 to start the company.
The segment was broadcast on CGTN and is not currently available online.
