Oct 14, 2026 Episodes Compute Economics Technology

AI Factories in the Outback? — Tim Nelson [Compute Series]

57 min read

This is the third episode in a multi-part series called 'Compute in Australia'. Series announcement here. More episodes to follow over the coming month.

[Recorded 30 September 2026.]

Over the past few weeks, I've been trying to work out what it would take, on the energy side, for Australia to build 10 to 15 gigawatts of AI compute by 2030.

So I sat down with Tim Nelson, one of Australia's foremost energy economists. Tim chaired last year's independent review of the National Electricity Market, known as the Nelson Review. He was previously Chief Economist at AGL and is currently Head of Markets at Snowy Hydro (though he joins me in his own capacity).

We discuss whether data centres could eat half of Australia's electricity grid, how they could actually lower power prices for everyone, what smelters can teach us about building renewable generation quickly, what could bottleneck the compute build-out, and why the best place to train AI might be off-grid in the outback.

This episode is best watched on video, as we often refer to the calculations I wrote on my whiteboard.

Video

Transcript

JOSEPH WALKER: Today I’m speaking with Tim Nelson, one of Australia’s foremost energy economists. Last year, Tim chaired a major independent review into Australia’s National Electricity Market, commonly known as the Nelson Review, whose recommendations are being implemented as we speak.

He’s an academic and a PhD economist, but he’s spent most of his career on the business side of the industry. He ran a multi-billion dollar infrastructure portfolio at Iberdrola. He was Chief Economist at AGL.

He’s currently Head of Markets at Snowy Hydro. But today I’m speaking with him in his own capacity, as someone who has spent decades thinking about how energy works in Australia.

Tim Nelson, welcome to the podcast.

TIM NELSON: Thanks for having me.

Could AI data centres consume half the grid?

WALKER: So at the moment, there are only two countries doing [frontier] AI training: the US and China. Say Australia becomes the third, and say we build out 10 to 15 gigawatts of AI compute here by 2030. And there’s an argument around how much we should want to build out, but for the sake of argument, just assume that it’s 10 to 15 gigawatts. I want to talk about exactly what that would take on the energy side.

And I want to start with this back-of-the-envelope arithmetic I’ve put up here on our whiteboard, and just walk through it and then have you correct me and tell me whether I’ve made any mistakes or not.

So to begin with, let’s say a data centre is running at a gigawatt constantly. That would mean it’s consuming 8.76 terawatt-hours over the year, because there are obviously 8,760 hours in a year.

So taking that as an assumption, I want to start with AEMO’s 2026 ESOO. So this is the Electricity Statement of Opportunities, and they had some interesting data on data centre consumption so far, especially over the last financial year, but also some forecasts. In particular, what they call their step change forecast, which is their kind of central case.

So just to level-set: the NEM delivers about 200 terawatt-hours per year.

NELSON: Yep, roughly.

WALKER: And at its peak, it’s about 35 gigawatts.

NELSON: Yep. And those peaks happen at different times in different states. So as a rule, the peak happens when everyone’s got their air conditioner on or their heater on. So very cold winter nights, very hot summer days. In the evening, when everyone gets home from work, they switch on the heater or the air conditioner.

And generally speaking – this is a gross generalisation – but Adelaide and Melbourne and the southern states tend to peak roughly at the same time, whereas the northern states of New South Wales and Queensland tend to peak kind of more coincidentally. And you think about that just weather-wise: if it’s hot in Sydney, it’s likely to be hot in Brisbane. If it’s hot in Adelaide, if not right then, it’ll be hot in Melbourne. That just is a gross generalisation of the way it happens.

So that number of 35, it’s comprised of a whole bunch of peaks that don’t just happen at the same time. So New South Wales, 14 gigawatts. South Australia, three gigawatts. So yeah, they’re just different peak demands in different regions.

WALKER: So AEMO’s – and I should say for people, that’s the Australian Energy Market Operator, whose job essentially is to keep the lights on. Actually, one way I’ve thought about the different market bodies is: AEMO is like the executive; the AER, or the Australian Energy Regulator, is like the judiciary.

NELSON: Judiciary, yep.

WALKER: And then the AEMC, or the Australian Energy Market Commission, is like the legislature.

NELSON: Yeah. The simplest way to think of it is: AEMO operates the market, the AEMC makes the rules for how the market works, and then the AER is the cop on the beat that makes sure that everyone’s following the rules and regulates the monopoly networks.

WALKER: Great. So this is AEMO’s central forecast for data centres. So in the financial year to 2026, they consumed five terawatt-hours, and that was about 3% of the NEM’s total operational consumption. They project that by the 2030 financial year, data centres will consume 15 terawatt-hours, which is about 7.8% of the NEM. And then by 2036, that will have grown to 34 terawatt-hours, representing about 13% of the NEM.

So today we’ve got 2.2 gigawatts connected in terms of the data centre capacity, of all the different types – so that’s cloud and AI inference – and the average load is about 0.7 gigawatts. So that means they’re only utilising about 30% of their grid connections.

So if we wanted to build out 10 to 15 gigawatts of AI compute by 2030… I’m assuming 75% utilisation, which is obviously a lot higher than what we’ve had so far. And I think that’s probably the most controversial assumption here. The reason it’s a lot higher is that I’m contemplating a lot more AI inference rather than the stock standard kind of cloud stuff, but crucially AI training.

And so if you take a 75% utilisation: 10 gigawatts, multiplying that by 8.76, and then multiplying that by the utilisation rate, would give us 66 terawatt-hours per year; or 15 gigawatts would give us 99 terawatt-hours per year.

So that’s four to seven times AEMO’s central forecast – about a third to half of the NEM. Does that sound roughly right to you?

NELSON: Well, I think the way you’ve done the maths kind of makes sense.

I guess the fundamental proposition, though, is that it is a forecast, and forecasts are subject to uncertainty. The other key thing to note, though, is that the market’s changing, largely as a result of the technology revolution on the supply side.

So if you were to go back 15 years ago and say to people, “Oh, there’ll be 30 gigawatts” – which, remember, we’re comparing that to the peak of 35. There’s currently 30 gigawatts of small-scale solar in the system. Now, if you were to go back 15 years ago, people would have said, “That’s never gonna happen.” So the first thing I’d say is the market has changed materially in the last 15 years, and all of that extra solar energy is producing energy that we didn’t have previously.

Now, if you go back seven or eight years ago, people saw solar as creating issues for the grid, because it only generates in the day, obviously, and people use power in the evenings and overnight. Now we’ve got batteries coming along. So we’re in a world where the maths on the whiteboard there is showing that, yeah, there might be this need for extra energy, extra capacity.

But the way in which the market is changing at the moment is that what batteries do is they do very useful things across two facets of the way that we’re thinking about this particular issue. The first is that they allow more solar to come into the market. More storage equals more capacity for more energy to be built, and solar energy can be done relatively quickly. You’ve got companies now all over Australia that are looking at putting in PV on rooftops, matching what’s been happening on the residential side. So the energy scenario is kind of growing in the sense of what we can supply.

And we’ve often got that wrong and thought, “Oh no, the solar will slow down this year. How many more rooftops can take solar?” And year on year it keeps growing at two gigawatts a year.

Similar with the large scale: there’s improvements in technology, both large-scale batteries but also some of the energy generation, all of the time.

So is it true that we’re likely to see new loads wanting to connect to the grid? Absolutely. Can we generate more energy and use those batteries to effectively store it for evening and overnight purposes? Absolutely.

And I mentioned the two facets of what batteries do. The first is that they store all that energy from the daytime. But the second is, to the extent that you can use that energy at the premises where the energy was generated, you reduce the amount of demand on the grid.

And so the key planning challenge in my mind is not so much: can we get the energy? It's: how do we make sure we find the right locations where you’ve freed up capacity, so that you can get the right outcome with those new loads. And that’s as much of a planning issue as it is a market issue.

I think that’s where most of the activity from governments is going. That’s, you know, how do we plan for this so that it’s done in the least-cost way?

Can we power the build-out with only renewable energy?

WALKER: Yeah. So I want to get to this question of where in the grid can we unlock this energy. But let’s just keep level-setting on the scale of the build-out before we get to that.

So: four to seven times the central forecast is a huge delta, right? Now, if we wanted to supply this just with renewables, which is what the government has said is its goal, then if you take the capacity factor for wind at about 35%, for solar at about 25%, as a rough rule of thumb, for every one megawatt of average load that the data centre build-out needs, you would need to build 3.5 megawatts of renewable capacity…

NELSON: Depending upon how flexible the load is. But yeah, let’s go with that as the assumption.

WALKER: Yeah, okay. So that’s an important point. But just assuming for the sake of argument: for a 10-to-15-gigawatt build-out, you have 7.5 to 11 gigawatts of average load. Multiplying that by 3.5, that’s 26 to 39 gigawatts of new renewable capacity, which is 6.5 to 10 gigawatts per year over the next four years.

Now, just to put that in perspective, the best year we’ve ever achieved for large-scale wind and solar was 2024, where we built 4.3 new gigawatts. So I guess I’m just trying to put the scale of this build-out in context.

NELSON: Yeah, look, I think the two issues that I’d point to are: the first is the energy generation itself, which is what you’re talking about now, which is the gigawatt-hours or the megawatt-hours – so the energy throughput.

And I think that’s where it’s also important to think about the fact that we are now moving solar into that C&I – the commercial and industrial – rooftop space. So for anyone who’s flown over Melbourne, Sydney, Brisbane, Adelaide, Hobart: you will see warehouses that have just got rooftops that are just perfect for solar. Flat rooftops – they look fantastic. A lot of those rooftops, I think, in the next few years will then be covered with solar.

So you’re going to see quite a lot of behind-the-meter generation come through. For some of those warehouses, they’re actually going to become net generators. They’re going to be pushing power back into the grid. So I think it’s true that this is a big challenge, but do I think it’s a challenge that the industry can’t solve? Not necessarily.

I think it’s really important that the objectives that governments have put before both data centres and the industry – of, “If you want to connect to the grid, that’s great, but bring your own generation” – should stimulate quite a lot of activity. And one user that currently might be very happy with a large-scale solution might be quite happy to go to a small-scale solution and free that kind of generation up so that it can be used by other users.

So that’s the advantage of running a pool exchange in a grid: that we’re all using electrons, they all go into the hopper, so to speak, and then they get distributed through the system. There’s enough flexibility, I think, in the way that these gigawatt-hours come to market that we should be able to solve this problem if we’ve got the right regulatory settings. And I think that’s what governments are trying to do.

WALKER: Okay, so just to play that back and make sure I’ve got it correctly: one thing this picture I’ve put on the whiteboard here is missing is that a lot of consumers, even businesses, [are] moving to things like rooftop solar – that’s going to remove demand from the grid, and therefore free up the supply for things like data centres.

NELSON: Yeah, yeah. I mean, the revolution that’s occurred in solar has caught many, many people by surprise over a decadal timeframe. So the battery revolution is only just happening now. We’re seeing hundreds of thousands of batteries rolled out. What those batteries will do is two things. They will allow more consumers to use the grid less – less need for that large-scale generation, because they’re doing a lot of it themselves behind the meter. But the second thing it does is it frees up the capacity in the system.

[In] the NEM review that we did last year, we talked about: the old way of thinking about the electricity system was, “I just buy power, and that’s kind of the way I think about it.” I now break it into its three components, and the components are energy, shaping and firming.

And so what a data centre or any other customer really needs is: they need energy – so they need solar, wind, they need the gigawatt-hours. Then they need the shaping, which is effectively: how do I store that when it’s too windy and too sunny relative to demand? And that’s what the battery does.

But then they also need firming, and the firming is likely to be things like gas turbines that you can switch on occasionally, leave on for quite a long time. Now, those gas turbines are relatively cheap to build – I say relatively, because all of this stuff is pretty expensive. But they’re more expensive to run. And so that’s why they make sense for doing very short periods when there’s no wind, no sun, and we’ve exhausted the battery storage.

WALKER: Yeah. Dumb question, but the fact that we’re removing demand from the grid through things like rooftop solar – that doesn’t change the fact that, according to the government’s expectations, data centres still have to bring their own energy, right?

NELSON: We’ve already got, though, quite a lot of customers in the grid today, and those customers might at the moment have large-scale contracts for – pick a number, whatever it is – 100 terawatt-hours; let’s just go with that. If they can reduce their demand by, say, 20 terawatt-hours because they’re using behind-the-meter solutions, then that frees up a little bit of capability for other new entrants to come in and, whether it’s take some of the existing renewables or displace some of the ageing coal a little bit more quickly.

Because the other thing that we’ve got to do in the next five to 10 years is we’ve got to replace all of the existing ageing capital stock. So a lot of those older coal-fired power stations that are coming to the end of their life.

What could bottleneck the build-out?

WALKER: Contemplating a 10-to-15-gigawatt build-out in the next three years, what do you see as the biggest constraints to achieving that build-out? What are the big bottlenecks?

NELSON: The funny thing about the energy system is that the recommendations that we put forward last year were really around how do you economically incentivise people to build out more generation, because that’s what we need. We need the generation that produces energy, the shaping that stores it, and then the firming.

And the fundamental problem that we identified was that most customers – so let’s think of a fish and chip shop all the way up through to a large supermarket chain – they don’t want to sign a long-term contract, because they’re looking at the market and they’re hearing from the media about how, “Oh, everything’s getting cheaper.” So it’s an unusual kind of decision to make: “So I’m going to lock myself into a particular price when everybody tells me that the tech’s getting better and in a couple of years’ time it’s going to be heaps cheaper.” But on the generation side, because these things are so capital-intensive, I need to get a long-term contract, or otherwise the equity and debt won’t come into the project and I’ve got no project.

So what we said is that to solve that problem, you basically need to not interfere where the market can solve [it]. So if you’re a big supermarket chain, you sign a five-year contract or whatever it might be, but you need some form of mechanism to de-risk that year six to year 15.

The recommendations which we put in place to solve that problem are now being implemented by government. So we think that – when I say we, the panel members from last year – we think that that’ll actually solve a lot of the problems of bringing new generation to market, for both existing load that will replace the existing coal-fired power stations, but also for any new entrants – so whether it be a data centre or whoever else that wants to connect to the grid. So that problem, I think, is entirely solvable with the recommendations that are being implemented by government.

There are two other issues, though. The first is: do I have enough capacity in the market? Because I can bring extra generation, but if it just gets constrained because it can’t squeeze through the wires, then that doesn’t solve the problem either. So there’s two really important parts of that puzzle for which I think there are reasons to be optimistic. The first is that for every battery that goes in, that’s just less load at the peak, and therefore there’s a little bit of extra room in the market more broadly. But that comes down to location. Where do you build data centres or any other new big loads to take advantage of that capacity that’s being freed up by other people reducing how much they’re using it? Because remember, the average demand in, say, New South Wales is eight gigawatts, compared to the peak demand, which is 14. So you can see straight away that there’s a lot of capacity there that doesn’t get used for huge amounts of the year. So you only have to reduce it a little bit, you can free up quite a lot.

But then the second issue is really around: if you’ve exhausted all of those opportunities and you still need to build more, who funds it? I think that’s where the government has said, whether it’s data centres or anybody else, if you want to connect a large new load to the system, then you have to pay for the transmission and distribution augmentation that goes with that. And that seems pretty reasonable.

But I do think there’s probably opportunities, with better working together between the energy industry and the data centre industry, to best locate in the parts of the grid where that capacity is available today. And there are parts of the grid where we do have generation that gets constrained off. So pick an area in the country where you’ve had lots of generation built [and] the transmission’s not big enough to get that generation to the load centre, which is a big city.

Now, historically, the way that the energy industry thought about solving those problems was: we will build more or bigger transmission lines.

But to the extent that you can take the new load that would have located over here and bring it close to the generation, that’s a win-win for everybody. The generators that are constrained behind that point, they’re very happy, because more of their energy is being sold. The data centre is very happy, because it’s now getting quality supply and it doesn’t have to fund as much of the transmission upgrade. But also communities are probably pretty happy as well, because co-location of assets means that that energy generation and that load – it kind of makes sense that they’re there together.

WALKER: Yeah. So just thinking of some other constraints for the build-out, if you need gas backup for firming, my understanding is that the three big manufacturers of gas turbines are booked out into the early 2030s or something like that. How much of a bottleneck is that?

NELSON: Look, it’s pretty tough to get gas turbines at the moment, but there are lots of credible players in the Australian market that have sites ready to go. And some of them, I think, would be pretty confident that they’ve got access to turbines if they can move towards financial close pretty quickly.

So again, what the government is doing with all of the state governments in getting those recommendations from our review last year up and running – the reason that’s so important is that that’s the firming part. The sooner that we can get those recommendations up, run those auctions for firming assets, they can then reach financial close, the turbines can be ordered or delivered, and you’re in a world where that frees up some more opportunity, I guess, for customers.

But it is fair to say that it’s a challenge, and that’s because globally everybody’s in the same situation, where they’re modernising their power systems. Globally, air conditioning is actually one of the biggest drivers of energy demand, and air conditioning is often quite peaky. And peak demand historically around the world was largely met by things like gas turbines and hydro. Increasingly that’s batteries. But I think your statement that it’s a tough market to get turbines is pretty reasonable.

WALKER: So do you have any sense whether that would be a binding constraint on getting to the 10 to 15 gigawatts in the next three years?

NELSON: Well, the other interesting aspect of these peaking assets is that, by definition, by being peaking, they don’t get used very often. And so globally what you’re seeing is a lot of data centres realise that, “Well, actually, we might want to do that ourselves with some type of backup generation on site.” And that might be a much longer-duration battery – so it might be more like an eight-hour battery – might be diesel gensets, whatever it might be. So I think there’s probably a combination of both.

There’s going to be a need for gas turbines to deal with what the industry calls dunkelflaute, which – for your listeners, it’s effectively just a nice German-sounding word for cold and dark nights where there’s no wind, there’s been no solar, and it’s just miserable conditions. Think of winter. So those dunkelflaute – you need some gas turbines that occasionally can get you through those periods.

But then on the reliability firming side, a lot of the data centres will probably bring their own backup generation anyway, because some of them certainly articulate that they need redundant capability. The reliability of the NEM, which is pretty good, is not enough. They need to know that they’ve got something behind the meter that they can rely upon as well.

WALKER: Will they be allowed to have diesel backup?

NELSON: I guess it depends upon what the regulations are for where they choose to locate. Again, this is where to be fair to the state governments and the Commonwealth Government, I think these are the issues that they’re trying to work through with these new requirements on data centres, which is a very sensible response, which is to say: bring your own generation, make sure you’re complying with all of the relevant planning and environmental laws. And if we do all that well, we should get a good outcome.

I’m an optimist by heart. I know some of those numbers on the board look like they’re challenging, but I’ve learned the hard way in the last 25 years that at every point where we’ve said, “Oh geez, that’s a big number, I’m not sure we’ll be able to do that,” we not only do it, we exceed it. And I think the biggest challenges are probably in the areas of social licence, and that’s where I suspect both data centres and the energy industry have got a lot of work to do.

How data centres could lower power prices

WALKER: So what would it look like if the data centre build-out actually lowered energy prices? Could you step me through the economic logic of that?

NELSON: Think of three components to everybody’s electricity bill. The first is how much it costs to generate the electricity. The second is how much it costs to transport it – so think of that as your transmission and your distribution expense. And then the final one is your retail. So that’s your retailer. Their key function in life is to manage the risk of buying energy for you. And I think most people wouldn’t understand just how risky that is: you want energy 24/7; the generation that’s always come through has never been constant at a 24/7 [level]. It’s always been going up and down. And so managing the risk of buying that every five minutes is actually what retailers do on your behalf.

Now, if we think through the way that those cost structures look in that world of growing demand. So the first one is the generation. And generation is largely – there are a couple of key things that drive that, which is: well, how much does it cost to build stuff? Then how much does it cost to run stuff? And once you know what those parameters are, you can kind of say what that cost looks like. So solar: cheap as chips. Wind: it’s probably more expensive than it once was, but…a bit like the gas turbine story, there’s some constraints there about how quickly you can get things done.

But what people don’t realise is that the prices we had in our market 20 or 25 years ago were largely a result of very low-cost coal and gas, because the boom that had happened in commodity markets hadn’t happened yet. So we all remember in the 2000s the huge growth in China and India drove a big demand for commodities. And so coal prices and gas prices are at least a couple of times higher than what they were 20 years ago. So when people talk about the cost of generation: it’s absolutely true that generating a megawatt-hour of electricity today is more expensive than it was 20 years ago. But a lot of that is due to the fact that we’ve got these really high coal and gas prices internationally now. But Australia benefits from that, because we export so much of it.

So if we take the generation story, whether it’s coal, gas or renewables, we kind of know what that cost is, and it should stay roughly at that cost of building new things. The transmission [and] distribution is where the big opportunity is. So we talked a bit earlier about, in New South Wales you’ve got eight gigawatts average, roughly 14 gigawatts peak. Now, economists love to talk about capacity utilisation: if I build this big piece of infrastructure, how much I use it determines what the unit costs of it are. So if I build a big freeway and then no one drives on it, then the unit cost for the car that’s going through is actually really, really high. If everybody’s using it, then the unit costs are really low. It’s the same for electricity.

So what data centres and batteries and some of these other more flexible loads can actually do is they can lift how much energy goes through the network. And if they do that successfully without adding as much to the peak, then you actually see unit costs fall for everybody. And that’s the energy economist nirvana. That’s kind of what we want.

Because if we go back 20 years ago, we built out all of this grid to largely deal with air conditioning. Then we went and put solar in, so we reduce how much energy goes through the grid. Then we went and put in batteries, which reduces how much we use the grid at peak times. So you’re seeing that we’ve got this capacity in the grid that if we use [it] more efficiently, we can actually see unit costs go down for everybody. So I actually think that if it’s the right type of load growth and it’s accompanied by the right type of battery storage, this could actually be a good news story for consumers longer term.

WALKER: And just to interrupt you there: is the way that would actually be operationalised through demand flexibility?

NELSON: Absolutely.

WALKER: Maybe you could just, for people’s context, break down exactly what that means. But does that concretely look like AEMO sending out instructions to the data centres, saying ramp down during these peak hours – and maybe if they’re doing AI training, they can do that because there’s not an end user who needs a Netflix movie then and there. Just how does that work?

NELSON: So the beauty of the National Electricity Market is that it’s actually a very volatile market. Now, I know that scares people, who kind of see volatility as, “Well, that’s something to be frightened of.” But it’s actually really, really useful for sending price signals to get the best behaviour to reduce cost longer term.

So at the moment, the market price cap in our market is a little bit more than $20,000 a megawatt-hour. The market price floor is negative $1,000. And the average price is probably in that $100-ish range, depending upon where you are in the market. Now, again, that scares a lot of people. They think, “Oh, what if it goes up to 20?” But no end users largely see that, because that’s what the retailer does: it manages it on their behalf.

Now, the beauty of those price signals is that if I’m a retailer, now I can think about who I contract with to manage that price risk. So let’s say you’ve got a flexible load, and then there’s another opportunity to build, say, a new gas turbine or a new battery or whatever it might be. If you’re willing to do it at a cheaper price than what the new supply option is, then I’ll actually contract with you.

So the AEMO dispatch instructions – all they really do is reflect what the financial position of those particular entities is. And so it all comes down to what’s most flexible, what’s the cheapest way of solving the particular problem. And that’s where I think not just data centres but any industrial user that’s got flexibility in the future should actually be able to monetise that and do quite well out of it.

WALKER: I guess it’s an important empirical question here as to how flexible are these AI training loads. Maybe a question I should put to a data centre operator. But I do know that they oftentimes have quite rigid service-level agreements with their customers. So I’m not sure how much scope there is there, but that’s something I should look into.

NELSON: But the beauty of having redundant capability on site is that there is a huge opportunity then for those moments of peak.

WALKER: Ah, I see. So you can just switch to your backup.

NELSON: Yeah, yeah. So, I think anyone who’s able to bring flexibility to the market of the future, they should actually be able to do pretty well out of it financially. But in addition to that, it’s one of those rare… Economists talk about externalities, and most people think about externalities as they’re always a negative thing: like, “Oh, it’s a bad idea.” But you do get positive externalities, and this could be an example of a positive externality, where somebody who’s doing something valuable for both themselves, but in the grid it kind of makes it better for everyone. But again, I’m an optimist at heart, so…

What smelters teach us about building renewables quickly

WALKER: So I’ve heard you may be thinking a little bit about smelters at the moment, and smelters are sometimes raised as a useful analogy to data centres, particularly because they have these very large loads.

NELSON: So there’s a couple of aluminium smelters around the country. You’ve got New South Wales, you’ve got Queensland. I think there might be one in Tassie. So yeah, there’s a lot of large loads. But you’ve also got cement loads that are very similar. So those industrial loads, be they aluminium or other types of industrial loads, they were foundational 30, 40 years ago for getting up new generation.

Which is really interesting, because that’s the world that we’re in now. So if you go back to the ’80s and the ’70s, a lot of the coal-fired power stations that were built were accompanied by big new things like aluminium smelters. Because if I want to get a coal-fired power station up, I need to know that there’s a certain amount of what used to be thought of as base load demand. So it’s kind of interesting how what’s old was new again, if that makes sense. It’s just a different type of tech.

WALKER: Is there anything specific we can learn from smelters about how to build out the renewable supply side quickly, as we’ll need to for data centres?

NELSON: So I think the couple of things that I’d point to there is: the first is that smelters, cement manufacturers, big industrial facilities are no different from any of those customers I mentioned before. They don’t want to sign long-term contracts, because they’re hearing from very clever people that things will get better in the future, because tech gets cheaper over time.

WALKER: Is this the tenor gap problem?

NELSON: The tenor gap problem, yeah. Which is interesting, because last year when we did the review, we came up with this term “tenor gap”. And again, it’s just an elaborate way of saying buyers’ and sellers’ expectations don’t match. That’s what it’s saying. So sellers want long-term, buyers want short-term; something needs to bridge the gap.

So whether it’s a smelter, a cement manufacturer or a data centre, the key thing that we recommended last year is: let that industrial customer talk to a retailer and agree on a price. And then the longer-term revenue that the new generator – whether it be a renewables generator or a battery or a gas turbine – try to provide some degree of certainty by contracting with that back end and then returning that to market as you roll forward in time. So almost like a warehouse-type concept. That was what we put forward to solve that problem.

And again, I don’t think there’s anything unique about other industrial loads relative to data centres about the way that that problem can be addressed. And I think a lot of the governments’ – both Commonwealth and states’ – response to the wave of data centre inquiries that they’ve seen is kind of in keeping with what we put forward last year. It’s a good outcome that those things are consistent.

How would Tim build 10-15 GW of AI compute in 3 years?

WALKER: So I know we’ve kind of touched on this implicitly already, but if you, Tim Nelson, were PM and you wanted to achieve this 10-to-15-gigawatt build-out in the next few years, how exactly would you go about it?

NELSON: So I think I’d probably do very similar things to what’s already been done. And this is where electricity – one of the other things that makes it so interesting is that constitutionally, electricity policy belongs to the state governments. But obviously the Commonwealth Government has a lot of skin in the game, because it owns all of the external affairs powers. It also has to deal with –

WALKER: It owns Snowy Hydro, for instance.

NELSON: It owns Snowy, but it also owns industry policy. It’s got regulatory bodies that oversee renewable certificates.

WALKER: Oh no, I’m sorry – Snowy was done under the defence power.

NELSON: Was it?

WALKER: Not external affairs.

NELSON: Oh, there you go. That’s something I’ve learned.

WALKER: Anyway, sorry – continue.

NELSON: So we’re in a space now where you’ve got the Commonwealth and the states having to work together, and they’ve had to work together collaboratively over the last 30 years to establish the market in the first place. And so one of the things that people don’t realise is that to get agreement to change the way that the electricity market works, you need all of the states and the Commonwealth to agree. That’s not the easiest thing for any state minister or the Commonwealth minister to achieve, because by definition all of the states have got different interests, all of that type of stuff. So we’re pretty proud that as a review last year, we were pretty much largely supported by all of the states and the Commonwealth. But if I was PM for a day, of course I’d implement the NEM review recommendations, because we’ve got to see that carried through.

The other thing, though, that I’d probably do is I would work with both industries to work out what are the strategic locations where these things make sense. Because it’s like a Venn diagram: we’ve got to get that sweet spot where it makes sense for a data centre to locate, and it has made or will make sense for the energy industry to locate. And that’s where, as I mentioned before, you’ve got a bunch of regions in Australia where you’ve got generation that’s wanting to get to market, but the transmission line’s not big enough. And so in the electricity world, we talk about them being constrained – they can’t get their energy to market. So the first thing I do is say, well, can we have a process whereby we try to find how we can get the industries to work together to locate in those pockets? So: NEM review, do that, and do all of the things that I think the states are doing, which is bring your own generation, which again is a bit of a tangent or an adjunct to what the NEM review suggested anyway.

But then the other thing that I’d really work through is: how do we get consumers, outside of those largely flat industrial loads, incentivised to use energy in the most efficient way? And I think that’s where things like batteries make a lot of sense – but then ensuring that the batteries are being operated in the most efficient way possible. So at the moment, you and I as households, we’ll put in a battery, and a lot of us do what I do, which is not the right thing to do – I should disclose I’m not the poster child for this. But I put in a battery, it soaks up all the solar during the day, and then I run the battery down slowly overnight. The more value-generating activity for me to do is, in that evening peak, help by putting that power back into the market. And then I might buy power again from the grid overnight.

The reason that’s so helpful is: we talked before about that 8-to-14-gigawatt difference between average and peak. If you could reduce how much households and other businesses are using that grid peak by putting power back into the market… Let’s say you could bring that down – I’m going to make up a number – let’s say you bring it down to 10 gigawatts. Now, the actual peak demand is still occurring; it’s just that it’s hidden now, because all the batteries are injecting power back into the grid, which is reducing demand very, very substantially. That four gigawatts – the difference between that 14 and that 10 – can be used by other consumers.

And so I think that’d be the other area, if I was PM for a day, that I’d focus on, which is: how do we most efficiently incentivise that? And I think we had some recommendations in our review that touched on a little bit of that, but I suspect that there’s probably additional things on the pricing side, the consumer side, that you could probably do.

Special compute zones in the outback

WALKER: Yeah, interesting. So on the first thing you do as PM – let me again play it back and tell me if I’m –

NELSON: It’s probably the second, because remember, I said I’d do the review recommendations first.

WALKER: That comes first. Yeah, you got me there. So the second thing you do… Is this the right way of putting it: you would try to find ways for data centres to build where there is renewable generation that’s currently being curtailed?

NELSON: Yes.

WALKER: Okay. So now my next question is: are those places essentially just renewable energy zones?

NELSON: I’d think of them as largely that. Because as a rule, the renewables that have been built so far, there’s been effectively two things that they’ve gone to try and find. One is great renewable resource – so the sunniest locations, the windiest locations. But the second is close proximity to existing transmission. And so you have had a lot of areas where you’ve had a lot of build, and sometimes that gets constrained as a result of there being too much power behind – the lines not being able to take it. So that’d be the first kind of point of call, so to speak. So I’d think of them as renewable energy zones. They may not be the renewable energy zones that state governments have identified, because they’ve already been built out. So think of them as kind of pioneer renewable energy zones, if that makes sense. The first generation of renewable energy zones.

WALKER: So let’s call them maybe special compute zones.

NELSON: Yeah, yeah.

WALKER: This is an idea that’s come up on the podcast before, which is inspired by some work the Institute for Progress in America did, when they suggested special compute zones for America. But basically like a renewable energy zone but for data centres.

So one reason to like having data centres in these places is, as you’ve mentioned, we’re unlocking curtailed energy. I guess the second reason would be we’re minimising the marginal loss factor, because if the data centres were on the outskirts of the big metropolitan areas and you’re transporting that energy, obviously you’re shedding some of it as you go.

Are there any other reasons to like having them in these special zones?

NELSON: So I think one of the blessings that Australia has is that it has renewable resources that the rest of the world – maybe apart from North Africa and the Middle East – the rest of the world’s very envious of. There’s very few other parts of the world where you get such great solar irradiance, relatively speaking, even through winter. But you also get great wind speeds. So I suspect that it’s not just the existing ones either – it might even be new prospective ones. And again, this is where it depends upon the use case for the data centre – or any new load, really – what it’s there to do.

WALKER: How important is the latency…

NELSON: Latency – and maybe it doesn’t even need to be grid-connected, full stop.

And so if you take, say, the western part of New South Wales: you’ve got great wind speeds way out west – I’m talking west of Cobar. Great wind speeds. I grew up in regional New South Wales and spent a lot of time in far west New South Wales – my parents live in Dubbo; I spent most of my high school years in Bourke, in far western New South Wales. So I can tell you it’s very sunny out there. There’s some great, great resources.

But if you go west of Cobar, for instance: great wind resource, great solar resource. Depending upon the use case for the data centre and the other infrastructure it needs, there are probably locations out there that make a lot of sense, because they’ve got great renewable resources. They’ve got communities that might be very welcoming. I say might because that’s obviously a key part of this: any data centre or energy company needs to make sure that it’s being welcomed, not opposed, because the social licence is as critical as any of the economics or the technical parts. But if you could get that alignment between resource and data centre type – or any new load type – and the way it could work, that could be a fantastic opportunity.

WALKER: Yeah, okay. So I want to talk about having these special compute zones in those areas like west of Cobar specifically, as a subset of this idea of special compute zones. Because social licence has become a huge issue for data centres in the US, and the same thing will happen here. And so one question you might ask yourself as a policymaker is: what if we could just bypass that issue entirely by putting the data centres in very remote areas? And so I want to ask you about what the economics of that could look like.

But just to set this up a little bit more: if you did try and put them in, say, existing renewable energy zones in the regions a bit closer to the seaboard, I think social licence is still an issue in those areas. I mean, take for example – and you’d be across the details of this more than me, so if there’s anything you want to add to this, I’d be curious to hear it – but say the reaction to the wind farms in places like New England, or the reaction to that renewable energy zone. I don’t see why a gigawatt-scale data centre would be any more popular than a wind farm in those kinds of regions.

NELSON: Well, I suspect there’s probably two issues that go on – there’s a lot of issues, but you can kind of broadly put them into a couple of key categories. And again, this is way outside my wheelhouse, but it’s kind of what I hear, so I’ll give you a feel for it.

There’s the water side, which I think is a really legitimate concern: is this going to be drawing on precious water? Water is a scarce resource – how does that look? But having spoken to a few data centres, some of them do have closed-loop systems which don’t actually use any water at all. So I think the tech appears to be there to solve particular problems.

And so I think that’s really then incumbent upon the industry, in the same way that it’s incumbent upon the energy industry, to work with communities and explain that and kind of take them through it.

Then you’ve got the energy side. Now, the energy side’s a really interesting one, because people use energy today. So it’s not new in the sense that people don’t feel that there’s a scarcity of energy in the way that they feel there’s a scarcity of water.

So then it comes down to amenity, these other issues, which are really, really important. And again, that’s where I think the energy industry and the new large loads, including data centres – they’ve got the same objective function there, which is: you want to locate in areas where you’re being welcomed, not told to jump on a bike and get lost. And that’s where I think it kind of makes sense for them to work together.

And then you’ve got the economic benefit of the point that you were making before, which is: the closer I locate the load to the actual generation source, the better off I am economically, because I’m not losing it across transmission lines over a long, long way. But in addition to that, I can then kind of plan my redundancy, and whatever it might look like, much closer to home.

WALKER: Yeah. So say for the sake of argument, say we did want to just bypass the social licence [issue], and instead of putting the special compute zones in the regions, we went another 100 to 200 kilometres inland, to places like – I mean, you said even past Cobar, but I guess I’m thinking in New South Wales local government areas like Cobar, or in Queensland areas like Hughenden – where the land is much more arid, it’s not used as much for agriculture, local communities would love some industrial activity, they’re not connected to the grid, the EPBC approvals are quite different. It seems like those could be really interesting places for a large-scale compute build-out. So we’re talking, I guess, on the fringes of the outback.

Just at a high level – and then I’ll ask you some more specific follow-up questions – how would that work? I mean, should we extend the grid to those places, and then the data centres would be grid-connected, or would we just have them running on behind-the-meter generation?

NELSON: It’d be big behind-the-meter generation. We’d be talking about some big wind and big solar. But look, I think it comes down to –

WALKER: So you’d have to extend the grid, in other words?

NELSON: Possibly, but I don’t know enough about the flexibility of the different use cases for the data centre. If they’re flexible, then it would make sense for them to actually go to those really great renewable resources, assuming communities are welcoming them on – they’re the very, very remote communities. I think that would actually be a really good outcome: if those communities were welcoming and you had the right flexibility, then that could be a great outcome. There’s those two ifs, though. Have you got the flexibility? Have you got a welcoming community?

Whether you grid-connect or not, I think, comes down to a couple of other considerations. The first is: the more connected the grid is and all the generation is, the more flexible it is. And the more flexible it is, the lower cost you have. Because if it’s really windy over here and it’s not so windy here, and that generation can’t get from there to there, then that means you need more capability here to store it for longer, or whatever it might be. So the more connected the regions are and the more connected the system is, the better the outcome, from just the way that the energy can flow through the system. So that’d be the first point, and that’s why you’re seeing interconnectors built. So you’ve got the new interconnector between New South Wales and South Australia that just allows that diversity of geography to kind of play a role.

The other key consideration, though, is how much easement do you already have, and how many existing power lines do you already have. So if you’ve already got a big transmission line going from point A to point B and you’re talking about slightly making it bigger, then that’s probably a lot less of an issue than if you’re building something completely new from scratch. And that would be an observation: that where you’ve got existing easements, existing power transmission infrastructure, [it] kind of makes sense to use those to the extent that you can. And that’s where, if you think about it just logically speaking – and again, I’m a bit outside my wheelhouse now, so I’m kind of freewheeling, which is always dangerous.

WALKER: This is what podcasts are for.

NELSON: So if you take, say, western New South Wales – you go way west, way, way west.

WALKER: So you’re talking beyond Cobar?

NELSON: Well, up to Cobar at the very least. We know that Cobar and all those places in the far west, they’re all connected to the grid today. So there must be an easement that connects.

WALKER: Oh, so Cobar is connected?

NELSON: Yeah, yeah, Cobar’s connected, Bourke’s connected. I remember the lights going out periodically when I was a kid in Bourke – I won’t deny whether I got up to no good or not. So if you kind of think about those easements: well, can you build more infrastructure there to – instead of taking the generation that way, perhaps it could go the other way? Now, I don’t know the answer to that question, but again, going back to that “if I was PM for the day”, that’d be the type of thing that I’d be asking. Could this be done? And what does that look like?

WALKER: And sorry, when you said instead of the generation going that way, could it go that way – what’s “that way”?

NELSON: So if we think about at the moment, the generation generally is on the eastern seaboard – so the big coal-fired power stations – and they’re sending it into the regional areas… When I say regional areas, the western regional areas.

But perhaps there’s a scenario where you could see it generated in the very far west of New South Wales… And I’m using this as a hypothetical example, by the way. There are great wind speeds out there, great solar. But it’s more just the point that there might be other parts of the eastern seaboard where you’d get a more welcoming community. I don’t know whether that’s far west New South Wales, but it’s certainly something which I’ve heard people talking about wanting to develop. There’s a group of developers out there that have done some work on what wind would look like. And so it certainly sounds like a sensible thing to do, but the question really is: what do the easements look like back to the eastern seaboard, and how would you do it?

WALKER: So if you did put these data centres in the far western parts of New South Wales, maybe Queensland as well – and I guess, as you’ve just said, it sort of depends exactly where – but what would it require in terms of network augmentation? And very roughly, on what order of magnitude are the costs here? Is this billions of dollars?

NELSON: It could be. But remember, we’re talking largely about – if that connection was to be made, the governments have made it very, very clear that it’s the data centres that are funding the connection. And that’s really a cost-benefit for that new user. They’ll think about it and go, “Well, actually, it just makes sense to get the energy that I’ve got here, sort out my reliability here, and then remain thinly connected or not connected,” or whatever it might be.

WALKER: Because there are other costs for them as well – like the fibre optic cables, et cetera.

NELSON: All that stuff.

WALKER: I mean, maybe labour costs are higher when you go remote.

NELSON: Yeah, I’m not sure. But you’re right, all of those other considerations are going to be really important. And whilst I can talk a little bit about how that might look on the energy side, obviously there’s all of the other kind of issues: does it make sense for the data centre to be far away from the city, or whatever it might be?

WALKER: And is it TransGrid they’d be dealing with?

NELSON: This is where the system is incredibly complicated. Like, one of the cool things about working in electricity is you do realise just how complicated this marvel of engineering is.

WALKER: It’s amazing that it even works.

NELSON: It is. It’s a testament to all of the amazing people in the energy industry: the engineers, the operators, the people who do all of that hard work. But TransGrid operates the transmission system in New South Wales. And there’s a different transmission provider in each state. So Queensland’s Powerlink, South Australia’s ElectraNet, you’ve got AusNet in Vic, you’ve got TasNetworks. So there’s a whole bunch.

WALKER: Transmission is a single monopoly in each state.

NELSON: Yep, more or less. There’s some slight variations, but let’s go with the simplification that there’s one in each state, and they’re natural monopolies, so they’re economically regulated by the Australian Energy Regulator.

Now, those transmission networks, they can’t just go and build something. They have to get permission to build it. And that’s where they go to the regulator. They ask the regulator, “Can I build this?” The regulator runs a cost-benefit analysis. If it gets built, then those charges go through to all consumers in the form of regulated transmission charges.

Then you’ve got the distribution networks. So the key way to think of the difference – and this is a gross simplification, but let’s go with it: the transmission network is the big steel – those big steel-looking towers. The distribution network is the poles and wires in your street. That’s the best way to think of the difference.

WALKER: In New South Wales, the three bodies that do distribution are Ausgrid, Endeavour and Essential, and they look after different parts of the state.

NELSON: Yep. So you’ve got Ausgrid on the real eastern part of Sydney, then you’ve got Endeavour west in Sydney and down to Wollongong. And then you’ve got Essential, which does the country.

WALKER: And so with distribution, you’re moving lower-voltage energy to the ultimate consumer. And that obviously passes off from the original transmission, which is, as you said, the huge steel structures carrying much higher-voltage energy over longer distances.

NELSON: Yep. And some large loads – so the really, really big loads, like, I’m talking the big industrial loads, so things like the mega, mega big data centres – they might connect into the transmission network directly.

WALKER: Directly, yeah. So for some of these big gigawatt-scale clusters (data centres) – I don’t know if you know this, but are they potentially only connecting to transmission? Or…? And I mean, if you were putting them in remote parts of the state, would they need distribution as well?

NELSON: Possibly. It depends how big they are and how big the grid is where they are. And that’s where I think that this is the interesting opportunity for co-location and planning. And this is where I think there’s a big opportunity for the energy industry to work collaboratively with the data centre industry to try and make sure that they’re working together to identify the least-cost solutions.

WALKER: Are there any existing price signals in the market that would pull data centres towards those areas?

NELSON: So as a rule, anywhere where you’ve got kind of MLF issues or congestion issues, there’s a big economic signal to come and locate load there, because you’re going to get a better price from – so let’s take a generator. Let’s say I need to get $100 per megawatt-hour for all of the generation that I’ve got to make my original business case, but I’m being congested, or I’m curtailed by 20% – let’s make up that number. Then, by definition, to the extent that a data centre is going to locate and help me solve that 20% loss of revenue, then I’m going to be prepared to share some of that, logically. So I think the economic signals probably already exist.

WALKER: So another interesting idea I’ve heard in terms of places to put these big training campuses is in brownfield sites. I’m [particularly] thinking of old mines, where there’s often renewable generation and transmission [already built]. So places like the Hunter Valley, the Latrobe Valley, Central Queensland.

Am I missing any other promising locations for the build-out?

NELSON: I keep coming back to that Venn diagram of great resource and a great place to build a data centre, or locate any new load… The one thing that I’ve learned in this industry is that I’m sure data centres are going to be connected, and I’m sure we’re going to see growth in that part of the market. But I’ve also learned that the things that people thought were going to grow and were going to be the biggest kind of growth turned out to be – yeah, they were going to grow, but there was something that came from left field. And so we know that there’s lots of new mines producing precious metals and some of the stuff that goes into the very tech that we use day to day. It might very well be that there’s a lot of the same principles that we’ve spoken about through this conversation that apply to some of those other industries as well.

WALKER: Yeah. If you had to place three one-gigawatt data centres tomorrow, can you tell me where specifically you’d put them?

NELSON: Oh, that’s such a hard question. I don’t think I’ve got enough knowledge on the spot to say, well, point A, point B, point C. But I think the things that you’d point to are anywhere where you’ve got the prospect of new renewable resource or existing renewable resource that’s constrained – that’s a great outcome. Access to transmission that already exists – and that’s where I think those clever companies – and again, giving a plug to the EnergyAustralias and the AGLs of the world – they’ve made it very public that they’d love to work with data centres to locate them on those types of sites where they’ve got access to transmission automatically. So I think, you know, renewables. And then I think it comes down to: does the data centre need to be close to its customers?

And when I say its customers, I’m going to think about customers being people in large metropolitan areas. Now, I’d like to think that I know what the answer to that is, but I don’t think I do. I think I’d be foolish to try and offer a view.

Renewable energy certificates and the time-matching debate

WALKER: So in terms of how the government is going to actually operationalise its expectations of data centres – like, for example, they don’t drive up prices, they don’t impose net costs, they’re 100% renewable-backed – the AEMC produced some advice to the council of energy ministers, part of National Cabinet. And the first policy recommendation the AEMC provided was to create REGO obligations for data centres.

NELSON: It’s funny, your “REGO” – because within the industry everyone calls them “regos”. Rego.

WALKER: Oh, really? Yeah, see, I’m a newcomer to this industry, so –

NELSON: Well, it might take off.

WALKER: Bringing a fresh perspective on pronunciation. But essentially renewable energy certificates. How well do you understand that policy recommendation, and can you tell me how you think about the problem it’s trying to solve?

NELSON: So the issue with electricity has always been: if you generate a megawatt-hour and I consume a megawatt-hour and it goes through the network, I can’t actually say hand on heart that the megawatt-hour that you produced is actually the megawatt-hour that I consume, because once it all goes into the wire, it’s all the same thing. And all the engineers out there will say even that’s wrong, because it doesn’t go into the wire, it just kind of transfers through the wire, so to speak. And even that’s probably wrong. It’s the danger of an economist trying to pretend that they understand engineering.

But what the REGO obligation is trying to achieve is – the market for renewables is largely solved through an accounting metric. So if you’re a generator and I’m a user, the way that we contract is not just for the power; it’s actually for these REGOs. Or, in the old world, it was large-scale generation certificates, or LGCs. And it’s really just a supply and demand issue: if my requirement as a large energy user is to procure enough REGOs to satisfy 100% of my consumption – let’s say that my consumption is 100 megawatt-hours – then I need to buy 100 REGOs from you. And what that means is that you’ve generated 100 megawatt-hours and I’ve consumed 100 megawatt-hours.

Now, how it all washes up through the grid electrically is irrelevant, because we know that what went in also came out, and that’s the accounting way that it gets sorted. So it’s a solid recommendation using an existing piece of accounting architecture, and it achieves the outcome, which is: you generate 100 megawatt-hours more because I’m consuming that 100 megawatt-hours. Does that make sense?

WALKER: Makes sense.

NELSON: Yeah. Because that’s the way that it’s historically worked, and it’s not unprecedented. So if you look at a lot of the major commercial businesses in Australia today – take Coles, Woolies, Aldi – they’re all buying their power from renewables today. So a lot of people who are listening to the conversation we’re having today probably don’t realise that there’s a huge number of commercial businesses now that voluntarily buy these things, because they want to demonstrate that they’re buying renewable energy. And the only way that they can do that is buying these REGOs or LGCs.

WALKER: Yeah. And so there are a few different ways of implementing this model. One is that you just buy and then retire an amount of REGOs equivalent to...

So one REGO is one megawatt-hour?

NELSON: One megawatt-hour, yeah.

WALKER: So say you consume a gigawatt-hour across the whole year, then at the end of the year you surrender 1,000 REGOs.

But another way of doing it is time matching. So we take hourly slots across the day, and then the energy you consume at those slots creates a liability for REGOs generated in those same slots.

Or a third, kind of hybrid, model that the AEMC proposed, which I think they favour, is you would have the day divided into larger chunks than the hourly periods.

Of those different models, which do you think is the best?

NELSON: Look, my personal view historically has been: REGOs are all you need.

WALKER: And so you just have the annual [approach]?

NELSON: Yeah, yeah. And the reason for that is that people talk about, “Oh, if there’s an oversupply of renewables, then the REGO is kind of overvaluing the contribution of the renewables.” But our wholesale market for the electricity already does that. So one of the things people don’t realise is that the price in the electricity market can go negative. And it generally goes to the negative of the REGO value. So in the middle of the day – say, for instance, in South Australia historically, 70% of prices are negative, and they can often be very negative, all the way down to negative $1,000. So renewable producers already have an incentive to stop generating at times of the day, or to build projects that produce renewables at different times of the day.

Where I think the AEMC’s recommendations, or the three options, are really interesting is that there’s already a movement in the electricity market to break the day down into chunks and trade those chunks. So, like, trade the evening between 5 and 9, or the morning between 6 and 9, and breaking it down that way. And the reason that’s sensible is that with storage now, it makes a lot of sense to say, “Well, I’ll buy my power between 10 and 2, and then I’ll sell my power between 5 and 9,” because that’s what I can do now with a solar system and a battery. So there’s a platform called Time of Day Markets, or TOD Markets, that allows participants to do that today.

So I think if you’ve got the annual REGO – if you’ve got that kind of obligation, and then you’ve got the electricity market providing that sharp value on when renewables are valuable, when they’re less valuable – and you’ve got the opposite for storage: the storage becomes really valuable to charge when there’s too many renewables and really valuable to discharge when there’s not as much renewables. The market solves those problems.

WALKER: So you don’t like the time-matching way of doing it?

NELSON: Look, I think it’s a really elegant solution, and if we didn’t have such a dynamic wholesale market for electricity, I’d say it’s probably a really good idea. I just think it’s probably duplicating what the electricity market already does.

WALKER: Yeah, interesting.

NELSON: But I wouldn’t say that that’s a universally held view. There’s a lot of people who really like the idea of timestamping. It’s just something which I think kind of duplicates what the wholesale market does.

WALKER: So here’s another problem: there’s a timing mismatch between getting the data centre built and then whatever renewable generation you’re building along with it coming online.

NELSON: So I know in the Australian context – because there’s been quite a few projects that have been built under the New South Wales LTESA framework, [and] the Victorian kind of approach, that you’ve seen a lot of renewables built – I’m pretty sure that governments are saying, “Look, we recognise that you can build a data centre probably more quickly than someone can build all of the energy infrastructure. To the extent that you are actually committed to doing what we’ve asked you to do long term and bring new infrastructure in, we’re happy for you in an interim sense to use existing renewables that are already selling into the grid,” because we’ve seen quite a significant increase in renewables in the last few years as they’ve all been built out under those underwriting policies.

WALKER: And then how would that work? You would do, like, a make-good promise or something?

NELSON: I think you’d use something existing – REGOs and certificates – until your new stuff’s available. Which is really sensible.

WALKER: So you would be buying those REGOs?

NELSON: Yeah, which is really sensible. Because you want to make sure that you’ve got flexibility in timing. Timing is where the flexibility becomes so important – that really rigid, “Oh, by the 1st of September 2028 you need to do this.” It’s: “By the 1st of September 2028, the goal is this, but we’re going to give you some flexibility about how you do that, so that if for whatever reason it’s not ready until the 1st of January 2030, that’s okay – but only if you can demonstrate you’ve actually been able to find those renewables that aren’t contracted today.”

Nuclear; unintended consequences; Jevons

WALKER: So does nuclear energy figure anywhere in the data centre build-out? Or does it just take too long to come online?

NELSON: So nuclear is an interesting one, because it comes up all the time and then it goes away for a while; then it comes back up again and then it goes away for a while. I think fundamentally, my kind of simple take as an energy economist is: you want to use the cheapest, most sustainable form of generation. At the moment, the cheapest form of generation is a combination of renewables and some storage and some firming – so whether that’s gas peakers or whatever it might be. Numerous studies have been done that give that perspective. So everyone talks about the CSIRO work. You’ve got AEMO, who’s constantly looking at it. You’ve got academics that have had a good look at this.

So a colleague of mine, Paul Simshauser, and another colleague, Joel Gilmore, did a really interesting study where they went back and said: if you had to rebuild all of the generation today and you had these options, which one would you pick? And the one that they found was cheapest was renewables, storage and gas. I had a paper in the Australian Journal of Management nearly 10 years ago where what we did is we said, well, look, let’s run the numbers on mid-2000s-level costs of generation, and we’ll run the numbers as what they were in 2016, 2017. And what we found is that if you were to replace the grid at that point in time based upon the mid-2000s, you’d build out more coal and gas. But by 2017, the answer was actually renewables and some gas peakers. Batteries, not so much by that point, because they were still pretty expensive.

So it’s funny: you fast-forward 10 years, you’ve shifted to renewables with some gas peakers. You fast-forward another 10 years, you’re in renewables plus batteries plus a bit of peakers. And they’re just within [a] 10-year period.

So nuclear on the cost side of things is pretty high. But it’s also the uncertainty of how long it would take to build. Because we’ve already highlighted that we’ve got a big challenge to build out the generation capacity to deliver on expectations of new load, but also replacing the existing coal. We don’t have enough time to develop a whole new industry if that’s going to take a long period of time. So I think before you even get to that nuclear [question], you’d need widespread government support to remove the prohibition. And then the second thing you’d need is a really hard look at: can you get it done quickly enough, and what’s the cost of it?

WALKER: Yeah. I’m curious whether there’s a way in which, if the data centre build-out does accelerate the energy transition, we could be like the dog who caught the car, and it causes some kind of unintended problems? I don’t know exactly what that might look like, but maybe you suddenly have a supply side that’s even more variable than it is today, for example. Or just how could this go wrong?

NELSON: I don’t see it necessarily as going wrong. I see it as more of a risk to be managed. And one of the things which I think we’re really proud of through the NEM review last year is that focus on… The dynamic in the Australian energy debate has always been around: do you go coal and gas, or do you go renewables? And our approach to that was, well, you want to do the cheapest thing that’s the most sustainable, but you also need to go into it with your eyes wide open that whenever you’re seeing tech change, the risk profile changes.

And so it used to be that we would store energy in big coal bunkers and big gas pipelines, then we would generate electricity to meet variable electricity demand. We’ve kind of now flipped that on its head, where we take the resource when it’s available – the wind and the sun – then we store it in a battery, or we store it in a pumped hydro or whatever it might be, and then we meet that variable demand. Now, that’s not a better or a worse system. It’s a system that has a different risk. And so we therefore need to make sure that we’re dealing to the risk, rather than dealing to the public debate about which one’s better. We just need to move that debate into: everything in life is about risk management, and it’s just about ensuring that we’ve got the right tools to manage that risk.

Where do I see things over time? Well, there’s the possibility that the social licence thing is a much bigger issue than what we’re anticipating now, and that might cause issues on the supply side. But my gut says that if it’s as big an issue on the supply side, it’s probably going to be as big an issue on the demand side for new loads. So I think there’s some – to use the economist terms – some automatic stabilisers there, in the sense of how that might move forward.

I think, though, the other thing which I’m very optimistic about is that data centres will get more efficient. By definition, there’s really smart people working in those industries who are going to be looking to – well, rather than building a gigawatt, can we develop a more efficient way of doing this so that it’s only 700 megawatts? And then, oh, can we do a more efficient way where it’s only 500 megawatts? It’s not my wheelhouse, but if you look at the energy efficiency of things that consume electricity over the last 30 years, they’re so much more efficient than they used to be. The lights in the roof where we are now [are] probably using a tenth of the power that they did 30 years ago. Air conditioners use a lot less power. By the time we shift over to electricity for cars, the efficiency we’re getting out of that is extraordinary. So humans get more efficient over time, and I suspect that we’ll probably see growth, but we’ll also see great leaps in efficiency as well. And that’ll again be a great outcome, because it means you probably need less infrastructure compared to what people might think today.

WALKER: Yeah, just on that: say data centres do get dramatically more energy efficient – and I’m sure they will get way more energy efficient, but say it’s dramatic. Is there a kind of Jevons paradox thing here, where now we use compute a lot more? The demand response just completely overwhelms the productivity improvement. How do you think that might play out?

NELSON: It’s hard to say, but you’re right. Like, when prices fall, people tend to use more of a product, depending upon what its use case is. So I don’t know. But the great thing about the way in which our electricity market works is that if there is a supply-demand imbalance, the price will shift. So if we build out infrastructure and then the large loads for whatever reason get more efficient and they use a lot less, then in the wholesale market, for instance, prices will be lower, which will be a net benefit for every other consumer.

So one of the things people don’t realise is the whole point of the microeconomic reform process in the mid-’90s was to actually ensure that consumers benefited from oversupply. So in the late ’80s, early ’90s, there’d been this huge build-out of all of the coal-fired power stations. You had so much excess supply that in, say, Victoria, that didn’t clear until 2012, I think it was. So the whole point of having a market is that if that happens again, then you get into a situation where prices benefit consumers. But I’m hopeful that we get this right, and if we get it right, then, as I say, on that network side of things, it should be a good outcome for everyone.

WALKER: Yeah. This has been really fun. Thank you, Tim.

NELSON: Thank you. Thanks for having me.

[The interview officially ended here, but then Tim and I kept chatting, and I decided to publish it. We keep pointing at the whiteboard where I wrote my calculations, so this section is best watched on video — you'll find it on YouTube, Substack or X.]

Post-interview chat

NELSON: I’m just so impressed that in the time between we spoke, you’ve gone away and researched all of that.

WALKER: I’m going to try and write up something more substantial, which I’ll share with you. But am I crazy in thinking that this is –

NELSON: So who said the 10 to 15 [gigawatts]?

WALKER: The 10 to 15 – okay, let me fill you in on that. Let me go from first principles, and then let me go from, like, there’s actual demand for that much. So there’s a set of geostrategic arguments for why Australia should want to become the second home for training compute for companies like OpenAI and Anthropic outside of the US.

The problem is that there are only two countries building frontier AI – like, the most cutting-edge systems – and that’s the US and China, because you just need enormous amounts of capital and talent to build them.

So Australia is not going to build frontier AI. So how do we maintain access and prevent ourselves from being cut off? Especially in a world in which ANZUS is less reliable than it used to be.

At a very rough first approximation, you would need maybe 10 to 15% of global [AI] compute to have that kind of leverage. At the moment, 75% is in the US. And the best forecasts have global AI compute at about 100 gigawatts by 2030. And that factors in the efficiency improvements, like the chips getting more energy-efficient. So say we wanted 10 to 15% of that 100 gigawatts.

NELSON: Gotcha.

WALKER: That’s 10 to 15 gigawatts.

At the same time, it’s been reported that Anthropic had been looking to put five gigawatts of training compute in New South Wales, with long-term plans to have 20 gigawatts in Australia. And it wanted that five gigawatts by 2030.

So we’re doing training here. Yeah, 10 to 15 gigawatts seems to be the kind of minimum – super back-of-the-envelope.

I think that whole “leverage” argument is very tenuous.

But I just wanted to kind of bracket that and assume that was the goal. Then what would that actually take?

And the reason for this discrepancy [between AEMO’s forecast in the 2026 ESOO and my back-of-the-envelope calculations], I think, is they’re just extrapolating the kind of stock standard cloud stuff with a bit of AI inference on top.

NELSON: Yeah. This is your strategic kind of training model. This is your business-as-usual using model.

WALKER: Yeah. And this is training and a lot more inference. This is cloud and more inference. Because that’s getting you to a third to a half of the NEM by 2030.

NELSON: So I think the interesting thing for me, though, is that if you flow this through to here, this is where those kind of left-field ideas about going very regional probably make a lot more sense. I want to say regional – I’m talking very isolated, remote parts. Because if you have got a little bit of flexibility, in the sense that you don’t need 100% utilisation, but in addition to that you’ve got a lot less need for continuous connection back to the major metropolitan areas, building this type of stuff is probably a lot more achievable if you don’t have all of the social licence constraints. That’d be my simple take.

WALKER: But isn’t this kind of build-out in the remote parts of Australia just crazy expensive, if you think about transmission?

NELSON: It depends if you connect it up, and I guess that’s the million-dollar question. Do you connect it up, or do you –

WALKER: Just do behind-the-meter?

NELSON: When I say behind the meter, think of it as a really, really big meter. So it’s almost like an embedded network.

WALKER: Yeah, like a local grid.

NELSON: Yeah, like a local grid. And I don’t necessarily think that… If that’s a prospect for some of these end use cases, then that’d be where I’d be kind of thinking it makes a lot of sense. And as we know, there’s people who are developing these concepts. But historically they’ve developed the concept with a view that, “This is a great resource; I then need to build this big line over to here, because that’s where the load’s going to be.” But if the load’s quite happy just to go there, then you may not need the line at all.

Because the other thing to remember is a lot of the gas pipelines are also in those far-flung parts of the various states. So you’ve got the Moomba to Sydney pipeline, you’ve got the Moomba to Adelaide pipeline, and then you’ve got the Queensland connection up here. That whole area is rich in solar, rich in wind, very sparsely populated, and has access to gas.

WALKER: Yeah. [My assumption of a 75% connection utilisation rate] is the thing that makes me nervous – that could be too aggressive.

NELSON: Yeah, it might be that it’s only 50% or whatever it might be. And, you know, again, this is not an area –

WALKER: That is weird that that’s all it’s been so far. Actually, I think in the ESOO it was 27%, in the last financial year.

NELSON: The thing that I’ve kind of looked at – and again, I’m no expert in data centres and their various uses – but as a proportion of their overall cost, energy is not immaterial, but it’s by no means the big… Well, and that’s where it’s very, very different from a lot of other energy-intensive users. A lot of other energy-intensive users, energy is actually a huge proportion of their overall cost structure.

WALKER: That is really important. And one thing I wanna understand around the economics is just, like: if all that matters is just time to power, and just getting the chips up and running as soon as possible just swamps everything else – maybe a lot of these costs, and some of the transmission network augmentation costs, etc., if you are building them remotely, maybe those are just kind of rounding errors.

NELSON: I don’t know.

WALKER: I think rounding errors is probably putting it a bit too strongly.

NELSON: But that’s where I think the clever way in which they’ve said, look, bring your own infrastructure, is that if you’re in this world, that probably is a little bit more of, “Oh look, we’ll just connect to the grid, fund the deep augmentation, it all kind of works.”

But if you’re in this world, particularly with that type of utilisation, you might then form the view, well, actually, that isolated approach – where you’re in a fairly isolated part of Australia, but you’ve got great solar, great wind, great access to gas pipelines – that actually might be your blue sky thinking, which is actually the least cost, depending upon what your needs are with this type of model.

So AEMO might be entirely right, because they’re talking about a different type of growth compared to what this strategic stuff that you’re talking about looks like.

WALKER: Yep. Cool, all right. Thanks, Tim.

NELSON: Thank you.

AI Factories in the Outback? — Tim Nelson [Compute Series]

Video

Sponsors

Transcript

Could AI data centres consume half the grid?

Can we power the build-out with only renewable energy?

What could bottleneck the build-out?

How data centres could lower power prices

What smelters teach us about building renewables quickly

How would Tim build 10-15 GW of AI compute in 3 years?

Special compute zones in the outback

Renewable energy certificates and the time-matching debate

Nuclear; unintended consequences; Jevons

Post-interview chat

← Previous

Video

Sponsors

Transcript

Could AI data centres consume half the grid?

Can we power the build-out with only renewable energy?

What could bottleneck the build-out?

How data centres could lower power prices

What smelters teach us about building renewables quickly

How would Tim build 10-15 GW of AI compute in 3 years?

Special compute zones in the outback

Renewable energy certificates and the time-matching debate

Nuclear; unintended consequences; Jevons

Post-interview chat

← Previous

Related episodes

Episodes

#Tim Nelson

Compute

Economics

Technology