As Artificial Intelligence slowly takes over the world – and more than its fair share of our energy consumption – there’s growing pressure on AI companies to ‘go green’ and power their massive data centres with renewable energy. But do the numbers – and not just the ones associated with profit – actually add up? Geoff Russell wades into the mire.
A recent Australian Financial Review article reported praise for our renewable resources as giving us some kind of edge in the data centre boom. I’d suggest that whenever you see the phrase “renewable energy superpower”, you should ask to see the numbers. It’s a great phrase for hooking naive investors, but is it anything more?
It’s very difficult to judge the size and power requirements of a data centre from a picture. The image above shows a small data centre pair in Holland. You could power this pair with a trio of ice-breaker-sized nuclear power plants. In contrast, the big data centres being built in China, the US and even Australia are 5 to 10 times bigger, or more. So how should we power them?
I wrote back in June on the feasibility of running something like NSW’s Tomago aluminium smelter using all of the renewable resources of the South Australian grid. It took me maybe 8-10 hours of work, spread over a week, thinking about it and running some tests using a data analysis tool I’ve spent many, many hours writing over the past couple of years. I used real data on the output of wind and solar farms in South Australia over the course of a week. Not any particular week, just the week before I started writing.
I calculated the impact of adding more batteries, and my computer code carefully measured the shortfall under a large range of additional “big” batteries. The bottom line here is that on a still night, the entire South Australian grid is virtually useless. On such nights, we could be seeing only about 3% of the output of our 2.7 GW of wind farms. Getting through even one night would need billions of dollars worth of batteries… about 12 GWh of them.
I didn’t bother thinking too hard or doing calculations on how to get through a “wind lull” lasting many days.
At around the time I did the Tomago story, I wrote a piece about how bad Google Gemini was. Even when I told it exactly where to find the answer to a question, it got the answer wrong.
Gemini 3.0 Pro; doing more than regurgitating opinions
Fast forward a few months, and Gemini 3.0 is out, and some superficial testing showed me it was worth signing up to the Pro version and giving it a spin.
I don’t know if it is intelligent, but it can do a whole bunch of stuff that I can’t do. I’m not talking about what it knows, but what it can do. It can work stuff out; thus providing new information and insights. It still makes mistakes. Just like me. I am constantly thinking about how to check my estimations, and it’s the same with Gemini’s.
One of my recent tests was to ask Gemini something similar to my Tomago question. Namely, could you run a 1 gigawatt (GW) AI data centre with wind, solar and batteries?
First, why would you want to? Wouldn’t you just connect to the grid? Not if you care about the cost of your electricity. Having a dedicated power supply has obvious advantages for big centres. When you buy from the grid, about 50-70 per cent of the cost is for transmission and distribution. If you can build a power supply next to your data centre, the electricity will be much cheaper. This is exactly why big smelters and the like are typically located very close to a power plant and get a great deal on their power.
Why did I pick one GW? That is massive. The world’s biggest data centre is currently 650 megawatts (MW), but the first 1 GW centre isn’t far off. Such a centre could cost close to $6 billion to build, given that a recently announced 504 MW centre is estimated to cost $3.1 billion.
So how did this latest version of Gemini do? It generalised the problem, understood the key features of a configuration that would drive costs and efficiency and devised a model to find good configurations. It realised, for example, that you would need to test for various mixes of wind and solar. Is it best to build 50/50 solar/wind, or is 80/20 solar/wind a better choice, what about 20/80 solar/wind? That choice will determine the amount of battery storage required; not to mention the amount of land. So Gemini chose a set of possible mixes and then tested the amount of storage required for each mix. But it also considered the impact of decisions about the level of “overbuild”.
What is “overbuild”? If the data centre demand is 1 GW, then do you build 1 GW of generation capacity? Definitely not. A 1 GW data centre needs 1 GW at all times, not just when it is sunny or windy. Over the course of a year, it will need 8760 GWh of energy. That’s just 24h x 365 x 1 GW. But 1 GW of solar power will only deliver about 1/4th of what you need; and 1 GW of wind will only deliver about 1/3rd of what you need. So if you built 3 GW of wind, you’d get, on average about 8760 GWh over the course of a year, which is an overbuild factor of 1 (i.e. no overbuild). But you will need batteries to store the electricity when it is windy, so you can use it when it isn’t. If you built 6 GW of wind, that would be an overbuild factor of about 2. Would you need more storage or less? The answer isn’t obvious. Best to work it out.
The exact amounts of generation required for each overbuild level would depend on exactly where you were building. We know that sunniness and windiness can vary widely on the planet, but the factors of 1/4 and 1/3 are reasonable for rough modelling. If you are in a very sunny place, where solar can average more than 1/4 of the max, then you can just re-run the model with the appropriate figure.
So Gemini considered three ratios of wind to solar power and 5 levels of overbuild. Here are the first 9 cases.
- 8760 GWh at a mix of 20/80 solar/wind or 50/50 solar/wind, or 80/20 solar/wind.
- 2 x 8760 GWh at a mix of 20/80 solar/wind, or 50/50 solar/wind, or 80/20 solar/wind.
- 3 x 8760 GWh at a mix of 20/80 solar/wind or 50/50 solar/wind or 80/20 solar/wind.
How do you estimate storage requirements for each of these configurations? You could do this using real-world solar and wind output, but that would vary depending on where you are and which year’s data you selected.
Gemini chose to generate data that simulated sun and wind variability, and it then wrote the computer code for this, ran it and then examined the table of output and reasoned fairly much like I would to choose reasonable configurations.
Some things are obvious. For the first 3 cases, with an overbuild of 1 (i.e. no overbuild), you’ll need a massive amount of storage. With an overbuild of 2, you’ll need less. It’s not so obvious what impact the choice of mix will have on the amount of storage. That’s the power of modelling.
Somewhere in this set of cases will be the one with the smallest amount of storage. That case will probably be the least-cost case, because storage is very expensive. But it may require land that isn’t available. I didn’t give Gemini enough information to model everything.
Simulating the weather
When I looked at the code Gemini produced to simulate the solar farm output, it looked familiar. I still don’t recall exactly where I’d seen something similar, but I think in a book about the mathematics of time series analysis.
If you look at a solar panel’s output during a sunny day, you’ll see it rise and fall a bit like a mathematical sine function. If I’ve lost you, don’t worry about it. Gemini basically mocked up a function to simulate this shape and then added some random variation to simulate seasons.
For wind, the code was rather different, because wind is more variable. But again, the methods will be familiar to people with the right maths background in time series, autocorrelation and the like.
The bottom line is that Gemini crafted a reasonable simulation model for generating generic wind and sunshine data to test the various cases. Using real data would be required if you were building an actual data centre and needed the best possible match of investment to location, keeping in mind that weather systems change and you need to plan for this variation.
Here are Gemini’s results. Gemini didn’t produce the plot, just the numbers. I’m sure it could produce the plot, but I did that the old-fashioned way, because it’s pretty simple and using Gemini wouldn’t save much time or effort. Each little panel is one case. The top row gives the overbuild level from 1 to 5, with a couple of extra levels between 1 and 2.
Each panel represents one configuration. A choice of the ratio of wind to solar and overbuild. The red dot shows the battery size estimated for firming. It’s nice to see the pattern of the dots, but I’ve also printed numbers in the panels so you don’t have to estimate them from the graph axis. (See Appendix 3 below for the cost breakdown into battery and power components.)
The top number in each panel is the amount of storage required to firm the chosen levels of wind and solar. The next number is the total GW of power generation being built. That’s the sum of the solar and wind power. So, for example, the top left panel has 3.22 GW and it’s in the S20/W80 row, meaning we would build 0.6 GW of panels and 2.4 GW of wind turbines.
The bottom figure is the capital cost of the panels, turbines and batteries. I used a CSIRO GenCost 2024-5 draft figures for 2026. The battery figure assumes 24-hour batteries, which GenCost estimates at $278 million per GWh. This is far lower than the prices anybody is actually paying at present because nobody is installing 24-hour batteries. You might like to think about why nobody would ever build a 24-hour battery except for weird exceptions like this.
Consider, for contrast, the recently built Waratah Super Battery. This is a 1.5-hour battery with a price of $588 million per GWh. Keep in mind that CSIRO cost figures have no legal force; they are not proscriptive. People can and do charge whatever they can.
I didn’t give Gemini any cost data, but it chose the 1.5 overbuild as a reasonable option. That’s over 5 GW of stuff to power a 1 GW data centre. This implies that the total capital cost of the electricity generation kit to power a $6 billion centre could be $16 to $22 billion. The cost of the electricity will depend on the interest rates you can wangle when you borrow the money.
For the 1.5 overbuild and S80/W20 panel, you’d have 4.5 GW of panels on about 11,000 hectares of land and some 228 x 5 MW wind turbines. Plus the batteries. As a guide, the solar panels alone would fill an area equivalent to about 5,500 MCGs.
As I pointed out in the previous story, if you’re thinking about a smelter, then an aluminium smelter like Tomago only produces about $1.5 billion worth of aluminium per year. That’s not profit, that’s raw revenue. I don’t know how much value adding is done to that product, but it should be blindingly obvious that running such a smelter on a dedicated wind, solar and battery kit would be crazy. Hence the desperate pleas for bailouts by our smelters.
But a data centre is different. The value of its product could be extremely high. If your AI can develop skills that people will pay big money for, then the cost of electricity may not matter in the short term. It’s no secret that AI moguls are prepared to pay top dollar for electricity at present. But I’m guessing that that won’t last.
What about the leftovers?
There is one thing I still want to think about in more detail for a later article. The excess electricity. Clearly, with an overbuild of 2, you generate twice the electricity that you require. But how much of that can be exported?
The data centre never uses more than 1 GW. But if you look at the overbuild 2 column, you have 6-7 GW of installed generation (wind and solar farms). So you need considerable bandwidth just to charge your batteries, even if you only need 1 GW of bandwidth between the batteries and the data centre.
Getting 7 GW out of that system to sell into the grid is both costly and non-trivial. Most transmission lines in Australia transmit well under a GW, and they are very expensive. Getting electricity out of such a system and into the grid would add billions more to the cost.
To explore this further, I need to speak to some actual engineers. A quick query to a WhatsApp list I’m on informs me that the charging bandwidth from the solar farm to the batteries shouldn’t be an issue. But I think I’ve done enough to warn you that when people speak in glowing terms about how we could be a paradise for data centres, thanks to all our renewable resources, you need to ask them to show you the numbers.
Lastly, what if we built a couple of nuclear reactors to power such a data centre? The GenCost 2024/5 draft puts the cost of a couple of large reactors at $17-18 billion. That’s using the price per GW, but you’d realistically get a couple of 1.2 or 1.4 GW reactors for this kind of money; perhaps a pair of Korean APR1400s.
Reactors need refuelling, so you’d want a pair rather than just one, so you’d have close to a 3x overbuild. But unlike a triple overbuild using wind and solar, you’d only need relatively normal transmission gear to shift the excess output. Moving 1.4 GW of excess power is considerably easier than moving 9 or 10 GW. Other intermittency problems not discussed above also vanish. You wouldn’t need synchronous condensers, for example. The entire grid would benefit from increased stability.
Appendix 1: What about a 400 MW data centre?
I used a 1 GW data centre as an example. That’s the size that is coming. But what about today’s smaller data centres? In particular, what about the Air Trunk 400 MW centre in Western Sydney? Here are the results of the simulation.
The lowest cost alternative is in the bottom left panel: $7.69 billion. That’s more than it cost to build the data centre (based on estimates). This data centre hasn’t been able to source 400 MW of firm power – instead it can only find 300 MW. That has to impact the bottom line.
Appendix 2: Background on AMS13 and AMS14
The company running the data centres in the header image of this story is Equinix, a company with data centres in many countries. It’s dedicated to using 100% renewable energy in Holland via Power Purchase Agreements (PPA), but in late 2025 has also signed a “letter of intent to explore” using Rolls-Royce-based small modular reactors as part of its future energy mix.
I should note that Equinix’s PPAs are the real deal. Many PPAs allow people to power data centres 24×7 with grid power, while only guaranteeing the same amount of energy from renewables; that’s a scam. For example, if you need 2,000 MWh of electricity a year and you get a PPA for this amount from a solar farm, then you are cheating if you claim to be 100% powered by renewable electricity. The actual energy powering your data centre at night isn’t coming from your supplier. Equinix isn’t running this scam but is actually matching the MWh it is using at the time it is using it with a renewable MWh. Perhaps the difficulties of achieving this have helped move it towards nuclear power.