I’m in a chat at work, and recently this question came up:
Are folks expecting a reduction in energy demands if DeepSeek-style models become dominant vs the ones you see from Open AI?
I’ve paraphrased it slightly, but it’s an interesting question, so rather than obnoxiously share a long answer into a that chat, I figured linking to an longer answer might be slightly less obnoxious, and allow me to use the answer more times in similar conversations.
My answer
TBH, I think the question is more about what you compare it to.
I don’t think the existence of smaller, more open models reduced the total energy demand from AI training when people realised it was possible to build them.
On the contrary, it seems to have accellerated it, as you now have companies building dedicated new chips outside of the big tech firms, that are specifically designed to work with new, less resource-intensive, more open models like the Llama series from Meta. Groq is one example.
Not good news for big-ass centralised “god-project” datacentres
However, it definitely calls into question the need for the OpenAI-style massive training datacentres and the silly stories about nuclear fusion.
Compared to that, yes, you might see a reduction in energy usage, but it’s not clear they were realistic anyway.
More efficient steam engines
Making language models cheaper, assuming you see them as an enabling technology, makes me think that a better mental model to use when exploring this question might actually be steam engines.
Making steam engines more efficient led to people using steam engines in new places, and more often, because they were accessible to more people, which resulted in more usage of coal in absolute terms.
So far, we’re just talking Jevons Paradox – something that gets invoked all the time when people talk about efficiency, and often in really narrow and disempowering ways. Why even bother pursuing efficiency if the gains are only ever going to be eaten up anyway?
Not every rebound is the same
There’s some nuance that’s worth taking into account though.
Things like the rebound effect absolutely matter when you’re looking at a resource you can consume, and where the limiting factor is the cost of using it.
In the steam engines example, you can think of the limiting factor not being the number of clothes you could sell to a willing audience, but the amount of coal you could afford to power your steam engine with. If you can sell twice as many clothes as you can make with the same amount of coal, and your goal is maximising shareholder value ahead of “provide the world with a sufficient quantity of clothing”, then you will, because the upside, in terms of making a bunch of cash, is twice as big.
So, when we view AI through this lens, then yes, models like DeepSeek would increase the number of people using them, see them used in new places, and see them used more. This would result in greater total energy usage.
Even if you don’t assume the ultimate goal of a your organisation is just making rich guys richer, this argument might work.
In geopolitical terms, AI is largely seen as a key to staying competitive by various governments in various countries, so making LLMs more accessible will likely increase the amount of countries relying on them in their industries to stay competitive. You can easily see this leading to more absolute use if you think that twice as much adoption means you being twice as competitive – there’s no real upper limit to how ‘competitive’ you might be.
There’s of course an open question in terms of how you define ‘competitive’ here. For the most part, countries tend to look at something like output in the form of GDP, (without looking really looking to hard at who the gains are accruing to, but that’s a separate post).
More efficient vacuum cleaners
However, not everything gets used more as it gets more efficient. My vacuum cleaner today is more than twice as efficient as my old one, largely because of various regulations mandating power usage.
Making it twice as efficient hasn’t doubled how much I vacuum my flat, because there’s an upper limit to how much I am prepared to vacuum.
Once my flat has reached an acceptable levelof cleanliness, I ‘bank’ the savings, because I can’t really perceive the extra cleanliness I might get from using the same amount of energy to vacuum twice as much.
So if I see AI as a thing that delivers some benefit, but where the benefits trail off once I reach a certain point, then it’s more likely that making LLMs more efficient does reduce total energy use.
Which is it?
Right now, I’m honestly not sure, because I don’t think I understand the technology well enough yet.
I get some benefit from using AI (and specifically GenAI) in services I use daily sure. I reckon there are a bunch of places I can see it being useful in my professional life. There’s not an infinite number of things though.
I’m more inclined to see this leading a to reduction in total energy use in the longer term (i.e. 5+ years), because there’s just so much inefficiency and waste everywhere, but in short term (as in 3-5 years), an increase.
Who should I read to understand this better?
Clearest person work I’ve seen exploring this rebound question is probably the work from Vlad Coraoma. If you want to spend 30 mins diving in, then this video, Digitalisation and the Rebound Effect is an absolute must.
An outcome I’m more confident about: more diffuse usage, and better suited to cleaner, more distributed energy
One thing I am more certain about though – I think smaller models likely lead to more diffuse energy demands globally, rather than massively concentrated ones, like the kind you would see in a massive gigawatt-scale datacentre.
This is important, because it’s the density of energy demand that often necessitates the use of stored chemical energy in the form of fossil fuels, and is used an excuse to power datacentres with gas, or coal, or even nuclear.
If you have more diffuse energy demand, it becomes easier to accomodate into the energy systems we already have. In many scenarios it’s also easier to meet with cleaner energy like renewables, which work by harvesting existing energy around us, at a lower density, rather than releasing stored energy through combustion.
There’s also another reason, that isn’t about the efficiency of the model directly, but more about the hardware models like DeepSeek required. If you can run AI on cheaper, older, less expensive hardware, then it can mean that the cost of the hardware as a share of the cost of offering service will likely decrease.
If you have lower fixed costs for hardware (i.e. lower capital expenditure, and capex), then cost of energy plays a larger role in the economics of making a service work – you don’t need to be running hardware continually to spread the cost of the massive capex over all the usage.
Illustrating this with some toy numbers
To make this more concrete, when you’re using extremely expensive, high capex GPUs, you want to run the chips as many as possible to maximise the number of inference hours. A 30k USD GPU card, running for 100 hours a year costs 300 USD per hour, and the same GPU card running 1000 hours a year costs 30 USD per hour. Let’s say you think you can break even selling an AI service at 5 USD per hour on average over the year. You’d need to sell around 6000 hours each year of AI compute.
If the initial capex isn’t so high for your hardware, you can likely get away with running it an in more variable and intermittent way. If the kit costs, say… 10k USD you don’t need as close to the maximum possible 8760 hours as year possible to bring down the cost per hour to something manageable.
If you only have spread the cost of across 100 hours of clean energy over the entire year, then you’re paying 100 USD per hour. Over 1000hrs and it’s 10 USD per hour. If you want to hit the same 5 USD figure, you only need to sell 2000hrs over the year. It’s easier to find 200 hours of clean electricity generation than 6000 hours of clean electricity generation over a year.
I live in Berlin, which gets around 1600 hours of sunlight per year, so if I only looked at solar, it’s pretty close already. Let’s say I lived in LA, California. That city gets twice that – around 3200 hrs a year. Much easier.
And this is before we think about how the cost of energy, and in particular clean energy changes as we move to a grid that with a larger share of renewables.
Those are my two cents anyway – hope this is of interest and use to someone!
Leave a Reply