Before getting into today’s story, a word from Distilled’s only sponsor: me!
In addition to writing the Distilled newsletter, I run a couple businesses: Cleanview and Campfire Labs. Entrepreneurial ventures like this have always funded my far less profitable impact projects like Distilled. Over the years, they’ve funded code schools for refugees, grants to climate nonprofits, and climate media projects like this newsletter and Carbon Switch.
Campfire Labs is a content agency that has historically worked with software companies. But over the years, we’ve been working with more climate tech and clean energy companies and nonprofits. We’ve produced everything from data-driven reports on food waste to clean energy tax credit guides.
If you’re interested in working with our award-winning team of content strategists and writers to produce content for your climate tech or clean energy company or nonprofit, I’d love to chat. You can reply to this newsletter or email me at michael@campfirelabs.co
Alright, onto this week’s story!
How the DeepSeek panic rewrote everyone’s assumptions about future AI power demand
Last week, I published a story on why we should all be more skeptical of AI power demand forecasts. The thrust of my argument was that the world is always changing in unpredictable ways and we should build infrastructure that is resilient no matter what happens.
A few days after I published that story, the stock market had a minor meltdown in response to some big AI news. DeepSeek, a Chinese AI startup, announced that it built one of the world’s best-performing AI models with a fraction of the chips, capital, and computing power that rivals like OpenAI, Anthropic, Facebook, and Google have used.
The DeepSeek news was a big deal because it rewrote certain assumptions everyone has been making about the future of AI. Since the beginning of the generative AI era, companies have said the only way to improve models is to use more computing power, electricity, and data. These so-called “scaling laws” have been core inputs to every AI power forecast.
DeepSeek’s latest model put all of that into question. The company’s open-source model and research paper raised the possibility that the big AI labs might be able to train the next generation of models with a fraction of the resources.
There’s still much uncertainty about what all this means for the future of AI, electricity demand, and grid emissions. The big AI labs may all fail to replicate DeepSeek’s results. The Chinese startup could be lying about how much computing power they used. Low AI training costs could lead to a surge in demand offsetting any efficiency gains (see Jevon’s paradox). We can’t know any of that right now. But in just a week the mainstream narrative on how much electricity AI will gobble up has shifted dramatically.
One way to see this is to look at the stock performance of the great “AI power plays” of 2024. For more than a year, investors have been buying up the stock of the largest power plant owners in the country—companies like Constellation Energy and Vistra which own huge nuclear and natural gas power plants. This was a bet on rising power demand.
On Monday morning, those stocks fell by roughly 20%. It was as if the quants all updated their power forecasts and realized the “scaling laws” were actually more like “scaling hypotheses” subject to change.
History is full of technological shocks and surprises like the one we saw from DeepSeek this month. Here are just a few examples from stories that I’ve written about energy technologies:
In 2010 IEA predicted that solar panels would generate 630 TWh of electricity annually by 2035. By 2019, the world generated that much solar annually, 16 years ahead of schedule. Last year solar power generation was about three times that. (Read the full story)
In the 1990s, analysts said lithium-ion battery technology could never be used in EVs. After all, building a Tesla-sized battery with the technology would cost $500,000 at 1990s prices. Over the next two decades, costs fell by 97%. (Read the full story)
In 2010, LED lights were installed in less than 1% of sockets. They cost $40 per bulb. Then the cost fell to as little as $2 per bulb and they became one of the most quickly adopted new technologies in history. (Read the full story)
The DeepSeek panic is a reminder that the world can change in the blink of eye. It’s a reminder that “It is difficult to make predictions, especially about the future,” as Niels Bohr said. It’s a reminder that nothing is certain when it comes to technological development.
We should build infrastructure accordingly.