Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
One of the hottest debates in the energy sector over the past year has been exactly how much power artificial intelligence was going to soak up — and what the impact of this was going to be. On suppliers: good. On the green transition: mostly bad.
DeepSeek has turned such conversations on their head. Following the Chinese group’s release of a massively more efficient large language model, the question now is how much energy demand forecasts will need to be revised down — and where that leaves the utilities in this space.
It’s mainly a US issue. While the need for rapid answers means that data centres required to use AI products are best located close to their clients, those needed to train the models can be built where power is cheap. Europe, with its expensive electricity, was always going to be a bit player here — and has hence been relatively insulated both from the excitement that gripped US utilities and from this week’s ructions.
Intuitively, it seems clear that US power forecasts are due for a cut. Training DeepSeek’s model required less than a tenth of the computing power required to train Meta’s Llama. The impact on power demand may not be entirely commensurate given data centres still need to be cooled, but this is game-changing efficiency.
Once up and running, DeepSeek may also be more efficient to use than, say, OpenAI’s ChatGPT, because it can apparently switch off the parts of its brain it’s not using. True, lower costs may result in higher usage — as chipmakers the world over hope. But the risks are clearly on the downside.
Such insights are hard to translate into new long-term forecasts — not least because US electricity projections looked a bit dubious anyway. There are a lot of them. And they say very different things. That makes it difficult to understand what assumptions were baked in in the first place.
More pertinently, even if one believes that cheaper and more efficient AI leads to more AI overall, that is not going to do much to reassure investors in the US power sector. After all, assuming DeepSeek really has changed the AI calculus, greater uncertainty over the eventual shape and location of demand ought to make so-called hyperscalers more wary of committing to long-term energy contracts.
Of course, while a slowdown would be bad news for energy stocks, it would be better news for the energy transition. A rapid increase in power demand would have been partly met by building new gas-fired plants. A more gradual approach gives renewables, batteries and nuclear power a better chance of filling the gap. One lesson from the DeepSeek drama is that AI can benefit the world even as it scorches some investors’ portfolios.
Read the full article here
