I came across this wild stat today, from an energy modelling friend, Tom Brown, who had another modeller refine it to provide the 90/90/90 framing, that I want more people to know about. Basically, at at 2030 prices, solar and batteries can provide 90% of the world population with 90% of their electricity demand for below 90 €/MWh. I definitely want to come back to this, and I expect others will want to as well, so I’m collating a few links for future reference.
What does this 90 / 90/ 90 thing mean?
90/90/90 is a reference to being able to serve 90% of the electricity, all year round, for 90% of the world’s population based on where the world’s population is globally distributed, at a cost of 90 euros per megawatt hour.
Here’s a bit more from Tom Brown’s personal blog post, where he details the assumptions, and the open source model used to back up these claims, with a few high level takeaways:
- Solar and batteries have the potential to dominate electricity supply in most regions of the world, providing cheap and clean power.
- Where there is enough space, this supply can be provided close to demand (only minimal grid costs were included in these results).
- The more expensive regions are concentrated at high latitudes in the North, where seasonal variations make it attractive to include wind and other energy sources.
- We can get far without worrying about the last 5-10%. The solutions for the last 5-10% could be fossil fuels in the short-term, long-duration storage as it matures, or easily storeable e-biofuels.
The modelling he used to carry out this analysis is the open source https://model.energy, and the prices he uses are prices expected in 2030, not in what they are expected to be in 2026.
Most of the world is a net importer of fossil fuels, leading to massive concentrations of wealth and power in a few countries, and you only need to look at the news right now to see how control over fossil fuel is weaponised. Bill Mckibben is good on the liberatory implications of moving on from fossil fuels in recent post, Solar as Solidarity, which is why I find this so exciting.
Relating this to what people spend on tech infrastructure and digital services.
At work, we talk about a fossil free internet, and part of this involves looking at the economics of an energy transition. We know that datacentres are one of the big new sources of energy demand – can this be decarbonised? Is this 90/90/90 stat, with 90 EUR / MWh reasonable for seeing if decarbonisation of new demand like this is plausible?
The short version is that the last 10 percent of decarbonisation is likely to be more expensive than the first 90, but if we step back from looking at the cost of energy services, to the cost of providing AI services, it can give us an idea of how economically feasible this might be.
One new company, Neuralwatt, who I have written about previously have done something really interesting here, by selling their AI services by the kilowatt-hour, not the token, and being radically transparent about the energy consumed when using their services.
They launched, selling AI inference at 2 USD per kilowatt hour (KW/h), which works out to be 2000 USD per megawatt hour (MW/h). The founder of Neuralwatt worked at Microsoft leading on some of the internal carbon accounting for digital infrastructure which makes me think they likely have a good idea of the costs of operating a services (see my interview with him on a podcast last year), and I’d be surprised if they lost money at 2 USD per kilowatt hour of inference. So, I think we can use it as a guide for what larger firms sell inference at by comparison to.
Even at the current abysmal (for the US) exchange rates, that works out to be something like 1680 EUR MW/h – that still leaves headroom to cover the cost of development, hardware, wages and so on, as well as closing the gap for that last 10%.
Further reading and links:
Tom’s original post on his blog
Tom’s post on linkedin, which triggered the 90/90/90 response post from a fellow modeller, Emil van Druten
Model.energy, the ‘toy’ modelling tool, that provides a front end to a more sophisticated open source energy modelling tool, PyPSA.
PyPSA – the open source modelling tool used to come up with these numbers
My post on how much energy hyperscalers use, and how much it might cost to go fossil free 24/7
Neuralwatt’s AI pricing for selling AI inference by the kilowatt hour
A (Claude assisted) thought experiment of mine, in taking this selling of digital services by the kilowatt hour, and researching the implications of applying it to managed WordPress hosting, with rough assumptions on costs, breakeven and technical architecture.