Our recent commentary on data center power demand was that AI can only improve through brute-force application of ever more processing power and ever larger datasets, with exponential increases in energy consumption for a linear improvement in performance.
In particular we cited the cost to train a model, which had grown to between a hundred million dollars to as much as a billion dollars apiece, which can be graphed with a clear upward trajectory even on a logarithmic scale. The natural conclusion was that the billion-dollar training mark would be passed, then $10 billion or even $100 billion dollars. Running these models, which is where long-term power demand enters the conversation, would likewise become ever more energy-intensive.
President Trump’s $100 billion, $500 billion Stargate AI data center plan seemed like a logical extension of those figures – now that figure merely seems like more of his typical golden-tinged bombast.
What changed is that DeepSeek brought the training cost crashing back down to just a few million dollars, according to its claims. On quality, it doesn’t beat the models trained for hundreds of millions or billions of dollars – but with a $5.6 million dollar training cost, enabled by a ‘Mixture of Experts’ method, DeepSeek V3 has outperformed models which cost tens of millions of dollars. And it is free, open-source and can be run locally.
DeepSeek states that its model also uses between 10 and 40 times less energy than its US equivalents.
So DeepSeek has dodged one, maybe one-and-a-half orders of magnitude of energy intensity – but now that Mixture of Experts is factored in, the same logic as before returns. All else being equal, more processing steps and more data are still fundamental routes to improve AI quality.
That may explain why Nvidia’s stock price is only down by 10%. It also means that data center power demand (which was growing disproportionately and rapidly even before LLMs entered the scene a few years ago) is still the central force driving electricity demand forecasts in the US and elsewhere. However, now we should regard this as being because Western demand growth is relatively weak otherwise (US electricity demand shrank 1.6% in 2023 and increased 3% in 2024), rather than because data center growth will be completely out of control in an objective sense.
Given the difficulty of building transmission, pumped hydro, nuclear power, and (under Trump) even renewables in the US, it’s still an easy call to state that data center and AI power demand will grow to match available power supply on the grid.
What will change is the price insensitivity of these data centers – with AI companies now undercut by the Chinese, they cannot engage in expensive investments such as SMR nuclear power plants as easily. Nvidia’s 10% stock fall is mild compared to independent power producers such as Vistra Corp (down 18%) and Constellation Energy (down 12%), which have considerable gas or nuclear assets.
NextEra Energy announced a partnership with GE Vernova last week, just days before the DeepSeek news, to develop a new fleet of gas turbine projects for long-term offtaker contracts. According to NextEra’s CEO and President John Ketchum, those offtakers would include data centers. Ketchum also commented that SMRs are a ‘next-decade’ asset type, and that new gas could also take until 2030 to come online – but that hyperscalers were not satisfied with the pace of the renewable energy build out alone. It’ll be worth keeping track of how this gas initiative pans out in the new context which DeepSeek has created.