DENNETT, Ms Belinda, Chief Executive Officer, Data Centres Australia ROHL, Professor Andrew, Director, National Computational Infrastructure STICKELLS, Mr Mark, Chief Executive Officer, Pawsey Supercomputing Research Centre [11:56] CHAIR: Welcome. I understand that information has been provided to you about parliamentary privilege and the protection of witnesses giving evidence to this inquiry. I'd now like to invite you to make a very brief—as brief as you can—opening statement. If there's anything that you'd like to provide in writing further to that, please do so through the secretariat to give us capacity to get straight underway into questions. Over to you for some short opening remarks. Ms Dennett : Thank you, Chair, and good morning—it's just about to be afternoon—to the committee. I have a couple of points. We've been building data centres in this country for over 20 years. We're home to some of the world's best data centre operators, founded and headquartered here in Australia. It's one of the reasons Australia has been so attractive as a market for data centre investment. The other thing is there are different types of data centres that perform different workloads and have different requirements with regard to location, energy, cooling and connectivity. There are also different business models. A key point of difference is co-location data centres and owner-built data centres. In a co-location data centre, the compute is not owned or controlled by the data centre operator. That's a very important point in terms of policy design. In Australia, 87 per cent of our operational capacity is co-location. Data centres support digital applications and services across the whole of society, from small business productivity to home entertainment. They are critical infrastructure and run the cloud services that support banking, air traffic control, hospitals and robotic surgery, and government and business applications. Increasingly, they support AI services as they are infused into every part of our economy. There's no sovereign AI capability without Australian hosted data centre infrastructure. The country that hosts the infrastructure has a far greater say in how the governance of AI is done, whilst also capturing more of the supply chain value. Australia is one of the very few countries that can credibly host this infrastructure at scale, and the decisions that determine whether we do that are being made right now. The contribution is already real. Our existing data centre capacity has generated more than $13 billion in construction activity and supports over 9,000 ongoing operational jobs, most of them skilled trades in the suburbs. Building the forecast growth to 2030, it adds $20 billion in construction activity and a further $5.1 billion in ongoing operational activity. If we get the settings right, we can also attract a significant portion of the Asian region's latency-tolerant AI work as a new export commodity. I'm happy to take your questions. Mr Stickells: On behalf of Professor Rohl and myself, we do appreciate the opportunity to present to the committee. I would also like to extend my acknowledgements and pay my respects to the traditional owners of this land, the Wurundjeri people, and extend that acknowledgement to all of our First Nations peoples. We're very pleased to be able to have this opportunity to present to you and acknowledge our submission and the short opening statement. I'll just reiterate, on behalf of both of the two national high-performance computing centres, Pawsey and NCI, we provide significant public capability through our national supercomputers, Gadi and Setonix. They enable research across health, agriculture, resources, climate and weather, and, increasingly, AI itself. This work has national value even when there is no commercial customer to pay for it. It's a distinction for us to make that private compute and public compute are complementary pieces of Australia's AI infrastructure puzzle. One, perhaps, distinction to make is the private investment must be attracted, but the public computing capability must be decided and appropriately funded. I think that's a deliberation before government at the moment. My last point is just to reflect a little on what was in the earlier session. We were able to observe a robust discussion around private compute reservation and arrangements. From our point of view, private compute reserved for public interest use should not be mistaken as a replacement for public compute investment, and we're happy to speak to that. Thank you. CHAIR: Thank you very much to both of you for appearing before us today and for your submissions and evidence provided to this inquiry. I've got two questions, one to Miss Dennett and then to yourselves. Firstly, Ms Dennett, in terms of your work and the members that you work with in terms of the infrastructure and the importance, I appreciate that point that you made around sovereign capability starting with building and operating here on Australian shores and the opportunities that we have. Arguably, data centres and the construction of data centres are becoming more and more apparent in the public arena. This is going directly to social licence in particular. We've seen some problems happen elsewhere across other parts of the world where there is a genuine concern around the rapid increase of data centres and what's happening or the impact that they are having on local communities. Obviously, what we're seeking to try and do here in Australia is get ahead of that as much as we possibly can in terms of setting some standards around that, particularly on energy, location and water usage. What proportion of your members' future electricity demand could realistically be matched with new renewable generation and storage investments? Ms Dennett : I think there are a couple of things to clarify. Firstly, Australia has a very different regulatory framework to the US, and we've had mechanisms in place already that protect what has happened in the US. In fact, the US is now trying to adopt some of the regulatory measures we already have in Australia. For example, under the national electricity rules, all large-load users, including data centres, pay 100 per cent for their connection costs and any augmentations required for the network. That's seen $3.1 billion invested in energy infrastructure between 2020 and 2025, with a further $7.2 billion expected by the data centre operators themselves. That includes building new substations and building new transmission to the data centre. We've had those measures in place. We've always paid for our own energy use, so there's no free ride in there, and there are no tax incentives to build data centres in Australia. That has put us in good stead. I talked in my opening statement about the compute being owned by the tenant and the data centre operator providing, if you like, the shell; 80 per cent of the energy is used for the compute and about 20 per cent for the shell. What tends to happen is that the energy is procured separately. But, between the customers and the operators, they are already investing in new renewable energy generation through power purchase agreements. Some 40 per cent of energy use is already offset by power purchase agreements and 30 per cent through renewable large-generation certificates, so around 70 per cent is covered by renewable energy. We know the government's new rules push us to 100 per cent. Everyone is on board with that. What we have to do is get the mechanisms right to ensure that those renewable energy projects can be built at the same pace as data centres can be built. Because, at the moment, there's a mismatch in timing. We need to make sure we speed up our renewable energy progress rather than slow down our data centre progress. Because, if we slow down, that capital will move elsewhere and we will miss the opportunity. CHAIR: Do you see an opportunity for us to use this to increase investment into renewable capacity and structure? Is that what you're saying? Ms Dennett : Absolutely—we say that the data centre industry is a catalyst for the energy transition. We would be needing to go through the energy transition with or without data centres. The opportunity we have is that data centres are investing in that infrastructure and underwriting renewable energy. If we don't attract that investment, the question is who pays for that energy investment. CHAIR: Professor Rohl and Mr Stickells, in the conversation we had in the previous session we talked about compute and sovereign compute and directing that to potential research capacity. What potential proportion would be beneficial to researchers and our research capacity around that compute access for public good? Prof. Rohl : I'll start. I think that's a difficult question to answer—which is what the previous group also said—because we don't actually know how large that infrastructure is. What I can say is that some of our partners are the Bureau of Meteorology and Geoscience Australia and they have plans for building AI models based on the data that they hold, which is unique data to Australia. But, at the moment, they don't have access to anywhere near enough GPUs to be able to do that work at a remotely affordable rate for a government department. I think it's pretty clear that some access would be invaluable, but I wouldn't want to put a percentage on it— CHAIR: But we don't have a number on it yet. Prof. Rohl : if you don't know. Is it $10 billion or $100 billion? These are the sorts of numbers that are being thrown around—about how big the AI revolution's going to cost. CHAIR: Is there a percentage rather than necessarily a quantum figure? Prof. Rohl : I'll give some context, based on an answer given before. CHAIR: Sure. Prof. Rohl : Today, we have 120 next-generation GPUs sitting in NCI. South Korea has just bought 8½ thousand for government. They have the same mission we do. This isn't for the whole country; it's for that research sector. So that's 8½ thousand, and they spent $350 million on buying that. Fortunately, for them, they actually signed the agreement two years ago, so they haven't been really smashed by the increase in prices. But Singapore, only a few weeks ago, opened their new centre with 1½ thousand GPUs, and that's $300 million. We have 120, so we're very far behind. So, yes, some additional capability would be welcome. Then, you have all the issues now around if GPUs are somewhere else. What does that mean? We have petabytes of data for both the geological and the climate and weather forecasting. Do you want to move that around and then compute on it and bring it back, or should the GPUs be located next to the data? We would argue that's to our advantage—the GPUs need to be next to the data when the data is that big. CHAIR: Yes. Given the evidence that was just provided by Ms Dennett, the ability to co-locate and the opportunities that may be available to us mean it's much better—I would ascertain—to have it here than access it elsewhere, because then you've got the impact of data or the compute happening elsewhere and under another jurisdiction's regulatory model as opposed to us having control here. Would that be right? Mr Stickells : Yes. CHAIR: Deputy Chair, I'll throw to you. Senator DUNIAM: I just wanted to go to Data Centres Australia. The comment was made that we've got to get the settings right. I just want to get as close as possible to a prescriptive view of what 'the settings right' is. The Commonwealth government, to whom we will make recommendations through the parliament, has only so much capacity to set rules, and obviously then it's the states and territories. But your guidance, as clearly as possible, on how to get those settings right would be helpful. Ms Dennett : Thank you. The things that we think are the biggest constraints to Australia making the most of this opportunity go to energy and to planning approvals, and the speed at which that happens. Every country is grappling with the energy transition and how they build their energy infrastructure to support this. One of the reasons Australia is so attractive is our potential abundance of renewable energy. That not only makes energy cheaper but is also just attractive to the people who are building out compute that want to use renewable energy. So, in ensuring we can find the way to do this—'quickly' is probably not the right word—in the right sequencing, some of the challenge comes when people look at the pipeline and the forecast for data centre growth, and there's a panic that that's all happening at once. Data centres are built in phases. The capacity comes online in stages. We can do this in terms of the energy transition. We can get it right. There's a lot of work happening across AEMO, the AEMC and energy companies to work this through. So that's the first one. The other one is around approvals and planning processes. When we launched back in November last year, we said the major call from us was to streamline our planning processes. States have had different ways in which they've gone about approving data centres. Victoria really was best in class and globally really competitive, with a four-month process and a concierge system, and getting through that quite well. New South Wales had been slower. They're both looking at different mechanisms now of how that works. But getting that streamlined was the thing, through the federal expectation of, 'Can we get more uniformity to make that faster?' The jury's out as to whether we've got there or we've got it more complex at the moment. It's perhaps looking more complex. But they're the two foundational opportunities, we think, in making sure Australia doesn't miss the opportunity. Senator DUNIAM: Thank you very much for that. Just to be clear on the planning and approvals component that you've just mentioned, the best streamlined model—noting all states and territories have different planning laws—and the best sort of approach remains the Victorian approach, as you say, with the concierge service. Is that correct? Ms Dennett : Yes. Just to be clear, it wasn't any less rigorous a process. But they have people trained that understand the data centre work and were helping it through the different agencies in state government. So things weren't getting blocked on the desk of someone who doesn't deal with data centres particularly often. And we find that: they've got to go to Transport, or they've got to go to Environment. It kept the process moving. So that had been working particularly well. I think it's perhaps paused with the election, which is a shame because, certainly, the global data centre market hasn't paused. It is moving at pace. Senator DUNIAM: Thank you very much. Very helpful. CHAIR: Ms Chaney. Ms CHANEY: To the Pawsey: in your intro, you talked about private compute being reserved for public use, which mustn't be mistaken for public compute. I think that is what you said. I was interested in hearing you talk about the differences between those and their relative importance. Mr Stickells : My question was framed to try and ensure that the considerations of government were not to replace the public compute investment with the reservation arrangements that are being contemplated. We actually need potentially both, and both would be a complement and help strengthen the advanced computing ecosystem that is being really scrutinised at the moment. Through the Pawsey Centre and working with NCI, we have had more than a decade of experience of managing, if you like, a reservation policy for Australian research computing. We've been providing a portion of the national computing facilities capacity in both Pawsey and NCI to our nation's researchers and to government agencies. That process has given us experience in that area. We understand that the demand now is heavily oversubscribed and that many of the researchers are seeking to use it for AI related developments in their science and supporting areas. We think that it's an important complement to bring private investment into this space to help support this market, but the fundamental need for public investment and public support so that we can govern, secure and maintain this in the national interest and serve communities rather than just customers is the thrust of our presentation here. Ms CHANEY: Sorry, I'd like to know a little bit more about the pros and cons of both and how they're different. Prof. Rohl: I'll start and then hand over again to Mark. One thing that's not really perhaps understood in the wider community is that the AI revolution wouldn't have happened if it weren't for supercomputing centres. Supercomputing centres have always been looking out for how we optimise the infrastructure that we've got. We've got a demanding audience, and so Nvidia saw an opportunity about 15 years ago. They said, 'We've got these chips that are great for video games, but we actually think, because they're doing video games, they're going to be great at a whole bunch of mathematical operations.' And it turns out that large language models just use these basic matrix operations as most of what they do, so they are exceptionally good at it. But they were developed to help other areas where matrix operations are important, like finite element analysis. That's when we're looking at engineers or—if you've got a car or an aeroplane and you want to see how it will react in a wind tunnel without the expense of a wind tunnel, that's the method that you use to do the in silico calculation. So that was the first part. When we start talking about AI, a lot of people here will instantly be going to large language models, which have, let's say, been around for about three years. Our researchers use a wide range of different AI techniques. AI has been around for a very long time. One example of what they would have done is 20 years of simulation of Australia's weather. Could we now train a model so we don't have to run these simulations, which take weeks again and again? We could now run them in an hour, but they would be trained on all of that data that has been calculated on these supercomputers. The way you configure that is different to the way you configure something to do a large language model, which is what these centres are being built for, just to develop large language models. So the emphasis here is that AI is wider than just large language models, and a lot of people are depending on those other aspects of AI to move forward in their research. Ms CHANEY: So reservation of compute in data centres built by hyperscalers will only actually help in large language models, but there's a whole lot more public interest need that needs to be done in purpose built centres. Is that right? Mr Stickells : I also should have added that there's an area of skilled support and the management of the data and the science and the expert support that goes with these projects that we support nationally that I think is likely unsupported or may not be well supported through a simple reservation arrangement with hyperscalers. That's an area that we invest heavily in, the training and skilled support environment, to provide the best computational resources to tackle some of these major scientific and national challenges. A concern that I have is that that would not be a feature of a reservation policy—if it wasn't carefully supporting skilled jobs and the control and management of the data environments that we may need to manage with sensitivity and with sovereignty. Those sorts of conditions are important considerations. I don't want to discount it, but I think it's a carefully considered process. Prof. Rohl : Can I directly answer your question. I think you need to turn that around and say that, for fitting large language models, you have a very specific set up, which is very good for that but not for other things. We have a more general way of setting things up. We can train and certainly do model inference on our infrastructure, but it is not as efficient as if you did it on a hyperscaler infrastructure, because we need to set up for all these other activities. I don't at this point want to get too technical, but, for us, floating point—not doing calculations with integers but with large floating point numbers—is incredibly important for most science. It turns out that it's not very important for large language models. They're going in one direction while we're trying to go in the other. So, yes, it would be useful infrastructure to make available to the research community, but it wouldn't satisfy their demand to a massive extent, because of those differences between the workload that they're set up for and the workloads that we traditionally run. Ms CHANEY: I have one more question. How much does demand for public interest compute currently exceed supply, and how do you expect that to change over time? Prof. Rohl : Right now, it's, say, approximately three. The applications closed last Friday. We had a quick look again, and, averaged across the two systems, it's about 2.9 times oversubscribed. Last year, we were three times oversubscribed and had to turn a lot of users away. We now need to do some analysis to understand whether a generation has disappeared and it's just the people who were successful coming back to us that has had that massive impact, or whether it's just been steady for the last 12 months. The other thing I should add is that, at the moment, we're in the midst of a procurement to replace the one at NCI, Gadi. What is very clear is that we will not be replacing the capacity that we've already got, so that's going to be even more problematic for our researchers getting access to that merit. Ms CHANEY: What sorts of things are not able to be done because we currently can't meet demand? Prof. Rohl : The answer is a lot of things. I'm happy to take that on notice and give you an analysis of those that were unsuccessful. It's about 100—no, it's about 200 applications. Ms CHANEY: That would be great. Just at a high level and a paragraph would be fine. Thank you very much. CHAIR: Senator Dolega. Senator DOLEGA: Ms Dennett, I thank you for appearing before us today. Can you tell us about who the members of Data Centre Australia are? And why do developers or operators become members of Data Centre Australia? Ms Dennett : We represent around 86 per cent of operational capacity. That's everyone from the Australian colo providers, which you might have heard of, like AirTrunk, CDC and NEXTDC; the hyperscalers, like Microsoft and AWS; some of the newer players that are currently building but not yet operating, like Goodman and Zerra DC—I've gone blank—to some of the global companies like Equinix and STACK Infrastructure. The majority of the colo providers, but also some of the hyperscalers, also have affiliate members that are part of the supply chain, which is an incredibly important story that often gets missed when we talk about data centres. They're manufacturing here in Australia. Companies like Drac Mechanical, which does all the steel fabrication here in Victoria, and Codecom, which does cabling in Geelong, are expanding and exporting to the region and manufacturing here locally, employing hundreds of Australians in great jobs. Senator DOLEGA: Are there any notable players that are not members, and why wouldn't they be members of Data Centres Australia? Ms Dennett : We have the majority of capacity, and some of the newer names you might have heard around neoclouds. IREN is a member. Firmus is not a member. You'd have to ask them why they're not members. But we've only been around for 10 months, so we're pretty happy with our broad membership at this stage. Senator DOLEGA: That's great. I'm a very staunchy trade unionist, and I think that everyone should be a member of their union. Ms MASCARENHAS: Fairness, come on! Senator DOLEGA: Where I'm going and what I'm trying to understand is that—you've got a really good picture of what AI factories and data centres look at based on the broad membership that you cover. In your opening statement, you were talking about representing some of the best operators. I'm wondering how this committee and communities could really test what makes an operator the best operator. What is the gold standard of some things that we should be looking at and communities should be looking at for players in the industry and particularly those that are setting up AI factories? Ms Dennett : AI factories are a little bit different, and they're new and perhaps not yet actually built, but, in terms of how we've thought about data centres for the last 20 years, it really is around security and around uptime. Data centres run every part of our modern life. That we are talking to people that are not physically in this room is going through a data centre. If you work from home, you're using a data centre. If you're watching Netflix or if you're using your mobile phone, it's going through a data centre. That 99.59 per cent critical uptime is incredibly important. Customers have choice of who they go with and where they locate. The Microsofts, the AWs and Googles of the world are very critical customers. They have choice. If you look at where those customers are choosing and who they're choosing to build with—we recently had our inaugural awards, and we talked about AirTrunk, CDC and NEXTDC really being the companies that built Australia's reputation as a data centre leader. They've been building here for 15-plus years. They are highly regarded by those global customers, and they've set the standard for the new companies coming along that are following and meeting that. All those things matter. Communities have said they want to have more say and they want to engage here, and we are hearing that, and that's something that I think we can improve on and we can do better on. I think we'll start to see that. I've been out to some of our community consultations out in the west of Victoria, and they've been really well run. This is new. We're trying to find out what the right process is, but I think our members are certainly stepping up there. Senator DOLEGA: In one of your responses to the deputy chair, you said some of the biggest constraints were energy and planning approvals, and you would encourage your members that they should actually follow the rules. What are the rules? Especially when it comes to planning, you wouldn't recommend starting to build a data centre and getting out some bulldozers without a planning approval, would you say? Ms Dennett : All our members abide by the regulations. Planning approval is critical, working through—it is a long process. Often, engagement with authorities starts many years before the application is submitted. Yes, we, of course, would support that. Senator DOLEGA: Part of good consultation, which you were talking about in some of the Victorian examples—communities are wondering, 'How is it going to affect me? Am I going to be able to have the water that we need in our communities for agriculture? Are we going to be able to have energy security ourselves?' and hoping that these are not going to be at risk. We've had evidence that this is a new gold rush in Australia, and communities are rightfully asking, 'What's in it for us?' in having these centres that will have potentially big global players providing lots of dollars. Do you think that that operators should be consulting with communities, in particular local government at the local level, as to how benefits to the community should be shared? Rather than just thinking, 'We as a company might like X cause or B cause,' or, 'This might be good PR,' you could actually go to a local council and say, 'We'd like to partner with you as we know through some groups that there might be a new bridge that's needed for a walking track or something that would benefit the whole community.' Do you think there are opportunities for greater collaboration with all levels of community, including local government, to talk about how benefits can be shared? Ms Dennett : I think firstly it's important to recognise that already happens. Under state planning regulations, data centres, like any developer, commit to community development infrastructure levies that pay for schools, roads and the like. On top of that, there are also voluntary agreements done with local councils to support. There are numerous things that have happened in the west of Melbourne, whether that's a shared bike path that's been funded, a basketball court, a community cafe, tree planting, waterway cleanup—there are myriad examples. That has predominantly been done with local council. I think what we're starting to hear is that communities want more say directly. What our members are doing—I saw a fantastic example recently where they went very early to the community and said: 'Give us your ideas. What do you want to see? What would be helpful?' They've taken all those ideas back, and that will again be worked through with the council to make sure that there's agreement. I want to be clear that those things are already done and already funded in data centre development. Senator DOLEGA: It's helping to build that confidence, right, when people understand this is the buy-in for us? We're hearing lots of great example about research and medical opportunities. We're hearing opportunities for greater efficiencies and improvements through these centres, but when people have that local buy-in as well, I think that will help them. Finally, in your submission, you highlighted a skills shortage. Are you aware of any members that are, say, contributing in regions to local TAFEs to be able to say, 'We need more people to upskill from having their sparky certificate to getting another qualification'? Is there opportunity for developers or operators to be able to work with TAFEs to be able to upskill workers, especially in regions where centres may pop up, so we see more of that community buy-in and workers in the community having new jobs of the future to be able to, again, build greater confidence and buy-in. Ms Dennett : That's certainly happening. In fact, Microsoft have the Datacentre Academy that's run with TAFE NSW and now Victoria University down here, skilling and building up the pipeline for people to enter our market. A lot of our members have partnered in that process. They take on work experience. They're taking on people for on-the-ground training. Some of our members do full four-year apprenticeship support. We're working here with the Victorian government around their skilling programs. There's certainly more we can do here. We've spoken to the ETU about working with them. Electricians are going to be a shortage, not just from data centres but also from the energy transition. They find data centres to be fantastic training bases for electricians. They're safe. We expect there's more we can and will do in that space. Certainly, as we now start to see the data centre footprint move more into regional Australia, which it traditionally hasn't, we know members are really engaging there, not just around the electricians and the trades but around STEM skill training, AI training and the like. We see that as part of our role. People that are doing the high-performance computing are also really critical to our industry, and we want to make sure we're supporting those and really shining a light on the opportunity of the jobs in our fantastic sector. Senator DOLEGA: I might just make a suggestion; take it or leave it. TAFEs are critical in regional Australia. So many people can transition to careers, especially with fee-free TAFE as well. If there's buy-in from developers and operators, I think that would be a good opportunity for the region. Thank you. CHAIR: Mr Watts. Mr WATTS: Thank you all. I think this has been a really useful hearing, adding some granularity to what we're talking about here. These conversations can often be very abstract, and I've walked through many data centres before, trying to work out how to convey to people outside them what's actually going on in the bits on those server racks. I sort of conceptualised this as how we can give Australia's world's best researchers the world's best tools to stay world's best in a changing world. I just want to tease out the important points you've been making about public compute and private investment, where there's overlap and where there are differences. You've made a really important point about the need for what you call tier 1 public compute for research provided by Pawsey and NCI. There's also obviously a lot of public compute provided by the universities themselves. In my electorate, for example, Monash University co-locates its MAVERIC supercomputer in a CDC data centre in Laverton. That's not tier 1, but it's working in parallel with private investment. Then you see these proposals for compute reservation as part of private investments in these frontier AI labs. You made the point about the kind of compute necessary for non-LLM research. You make the point that you're three times oversubscribed for that. How much of that research is shifting to more AI-style research, given recent advantages? What do you forecast demand over coming years looking like in that respect? Prof. Rohl : I'm actually doing a study at the moment that we're getting towards the end of. If I'm allowed to, I'll provide that. Mr WATTS: That'd be useful. Prof. Rohl : Interestingly, because we're doing it in collaboration with the US national labs, they've seen that their— Mr WATTS: Who have a reservation policy? Prof. Rohl : It's one of their national labs' merit allocation schemes, and they're seeing about 40 per cent AI. Scanning through keywords and titles, we look about the same, but I'm trying to do a bit of a deeper dive to understand if it's the traditional communities that we've served now starting to take these AI techniques and implement them and increasing what they can do by combining with the models they've currently got, or if it's actually new people coming in and saying: 'I can do something that I've never been able to think about before. AI now makes it possible.' We want to bring that to you. Our challenge is we only have 120 GPUs. I think MAVERIC is 144, so that gives you some idea of scale. Our challenge is that, even though we're oversubscribed, there's a massively long queue to those GPUs, because there are so few. Mr WATTS: A short way of answering that is we need more of both. Prof. Rohl : Yes. Mr Stickells : Mr Watts, I was going to say it's not an either/or. It's a change. It's an 'and'. There are many, many more domains now, and the world of medicine is increasingly turning to computational needs to accelerate drug discovery and to support that. That's a system that hasn't used computation as much as other fields in the last decade or two. AI is now adding another layer of productivity or workflow changes to science and to research. It's an 'and' not an 'or'. There are large instruments that can do highly accurate calculations and modelling like our national supercomputers, but the entire sort of workflow for science and in any other industrial domains is embracing AI, which needs both high-intensity training capability and the delivery capability of other types of computing architectures. It is a changing ecosystem of which we are one important part. The ability for us to ensure that we could secure the best of a benefit to the nation around private investment in this area is obviously something we are keen to support, but we reiterate the need for those public pillars of capability and sovereignty and matters that direct towards our national needs. Mr WATTS: Ultimately, both of those categories of compute will have to be hosted in a data centre, a physical location. If we are going to meet that expanded demand for research compute, we need to roll out data centre capacity as well. Is that right, Ms Dennett? Ms Dennett : Absolutely. People often ask why we are seeing the data-centre growth. I come back to three things. Firstly, we are still shifting on-prem compute that is in the basements of buildings out and into purpose built data centres. We are using more digital services every day, and then we've got AI, which is fuelling this opportunity, to use technology to solve problems that we haven't been able to solve before. That's the reason for the growth in in capacity. It's not speculative. In fact, 97 per cent of our co-location capacity is already leased in Australia. Much of what we're building is already preleased. We need to be building more just to keep up with our local demand, and then there's the export opportunity. Mr WATTS: Are you able to provide some granularity about the kinds of research we're missing out on here? There was a good example about Professor Barca's Gordon Bell Prize team being forced to conclude its research overseas because of a lack of compute in Australia. What are the two-thirds of the research projects that you don't allocate compute to about? What kind of research projects are not being served at the moment? Prof. Rohl : I think I've already agreed to give you a summary of those. Mr WATTS: You can take it on notice and provide it in detail. Prof. Rohl : I've asked, at a one-paragraph level, if you would like more detail. Mr WATTS: I think on notice, if you could provide as much detail as you would like to, we'd be interested. I have a specific question on how we shape the future and try to meet this demand. I understand the request you have for public funding of this public compute. The other mechanism that we're considering is this reserved capacity type model. If the government did write a reserved capacity condition into a deal with a frontier lab, what would it need to specify for it to be useful for your purposes—for example, what hardware types, network or storage where data is held, support stuff? What are the kinds of levers that governments should be trying to pull to shape the future on this research front? Prof. Rohl : The No. 1 thing is going to be very fast network connectivity between where we are and where these hyperscalers are. Mr WATTS: When you say where we are, where? Prof. Rohl : Canberra and Perth. The challenge is—and we have an example at the moment of trying to move two petabytes of data out of our data centre into another data centre, and it takes months. It takes that long to do that transfer. These are 100 gigabits per second. They're not superfast, but five years ago they were superfast. We need that connectivity. Several of my members use AWS in addition to NCI. We've already started to say, 'How do we chain these workflows together?' What we're suited for, which primarily is this massive amount of simulation data generation and data reduction, then transfers a small amount, and then they build services on top of that to serve to their clients. In terms of that specification, I think, other than the network, it is pretty straightforward. It's what number of GPUs and how much data. We would then work things out based on that. The other thing with hyperscalers in particular is that often they have network charges. If you're moving large amounts of data, that can be a crippling cost. That would need to be factored into account. Mr WATTS: What about the nature of the data you're dealing with in this research? Are there categories of data and research that we aren't able to share overseas for either legal, regulatory or sensitivity issues—health data, genomic data or maybe security research data? Are there additional reasons for us wanting to have compute locally versus being forced to go overseas? Prof. Rohl : A good example of that would be the data particularly from the mining industry, who have a lot of consultants going out and doing a lot of reports. We actually have a lot of information about the continent that no-one has ever pulled together because it's in lots and lots of reports from individual consultants, both on geology and on the environmental side. That's a massive resource, and we would not want that to. Mr WATTS: We wouldn't want that going overseas. Prof. Rohl : No. Mr Stickells : I want to add perhaps a slightly more sombre note to your question, Mr Watts. If the only approach was to adopt the reservation policy, I think we'd also have to consider what research, which that won't be able to migrate to that sort of model, stops. I believe there are some specific sovereign or technical requirements for certain datasets or computational needs to be managed in the specialist facilities that are offered through public compute. It's a follow up to the demand question. I think there would be some risk of areas of public concern that would that would be lost if we went down that path. Mr WATTS: That's a further reason for the yes-both-more-type frame. Mr Stickells : Yes. CHAIR: Senator Pocock. Senator DAVID POCOCK: Thank you for your time. For Pawsey and NCI, I appreciate all those answers. I'll chuck a few more things on notice to get some more detail. I've got some questions for Data Centres Australia. How do you respond to concerns that your members are taking construction workforce away from the desperately needed task of building more homes in this country? Ms Dennett : What we tend to find with data-centre construction is that there isn't a crossover in terms of existing workforce. The construction firms that are building data centres move from project to project, phase to phase. Often, the workers are employed for many years on the one data centre build-out. They are very specialised. We have recently put out a new report talking about the skill sets of those jobs. They're quite different construction jobs to those doing residential housing building. We do acknowledge that working on data centres is attractive, and we believe we are attracting more people that want to work on data centres into the industry. We recognise that we have a role to play and that our construction firms have a role to play in ensuring we are building out and training up that workforce. Senator DAVID POCOCK: Thank you. At the start of your answer, you said that you aren't having an impact, but then, at the end, you said you acknowledge that you are an attractive place to work and are pulling workers to data centres. I assume they're coming from somewhere, presumably building houses. I've heard from a number of construction companies that say that they are finding it harder and harder to find sparkies in particular. I think this should be a concern, and it's strange that you can't acknowledge it. What about the RBA's concerns that data centre investment is inflationary? Ms Dennett : I'll take your first point. I think, perhaps, you misunderstood me. I was suggesting that we don't believe we pull people off residential projects onto data centre projects, but we do believe we are attracting people who may be starting out their careers and want to work in the data centre sector. Senator DAVID POCOCK: That's certainly not what I'm hearing from anyone else in construction, but sure. We'll take that evidence. What about the concerns that data centre investment is inflationary? We've heard that straight from the head of the RBA. Ms Dennett : There is certainly a lot of investment coming into the country in data centre builds. Around the world, we're seeing that, if it weren't for data centre construction, economies would be in recession. We think this investment is incredibly important. We think there are a lot of factors that go to inflation. Data centre investment may be one of those, but it is also incredibly important that our economy continues to grow, and that's what data centre investment is doing at the moment. Senator DAVID POCOCK: Senator Dolega talked about training and apprenticeships. How do you respond to concerns expressed by NECA, HIA and others that apprentices training in data centres are getting a less enriched training experience that potentially doesn't give them the skills they need as an electrician to wire a house or do more complex tasks? Ms Dennett : We've been working with NECA. We've been working with the ETU. If we continue to get the settings right in Australia, there's a lot of work in data centres for electricians. The people that are running apprenticeships in our membership are certainly looking for long-term careers for the apprentices that they take on—working with their companies. I think there's more work that we can do with organisations like NECA and the ETU to give people broad opportunities, including working with TAFEs, but, certainly, our companies that are taking on apprentices hope that those apprentices will stay and work with them in the data centre sector for a long time. Senator DAVID POCOCK: Finally, your members basically pay no corporate tax and are unlikely to pay any corporate tax any time soon. I know that most of their tenants are also multinationals who themselves don't pay much corporate tax. What's the ongoing financial contribution after the initial construction sugar hit? Ms Dennett : We've got research out that shows that that's $2.3 billion of economic activity annually in operation—past the construction phase. If we look to the 2030 build-out, that would be $5.6 billion of economic activity per year. Senator DAVID POCOCK: It's not a huge amount, given that a 0.25 per cent increase to interest rates, I think, is about $4 billion shouldered by Australians. Is that all we're going to get? Ms Dennett : I would think the billions of dollars of investment, the thousands of jobs, the supply chain economic boost and the exportability of many of the Australian companies manufacturing in Australia is certainly worth it. Senator DAVID POCOCK: What do you say to the fact that 70 to 80 per cent of the build of data centres is imported? Ms Dennett : That's not what our research shows. Certainly, our report looks at the construction of data centres and the contribution of data centres exclusive of the compute. Much of the compute is imported, but much of the build of data centres, whether that's the steel, the cabling or the switch rooms, is manufactured here in Australia by Australian companies. I think our report puts it at around 55 per cent of that investment that stays here in Australia. Senator DAVID POCOCK: That's different to what we've heard from some of your members, but thank you for that. CHAIR: I appreciate we are right on time, but I know Ms Mascarenhas has a question. Ms MASCARENHAS: Yes; probably a couple. I'll try and do a speed date version! This is a question to NCI and Pawsey. We've talked a lot about public-interest compute. Can you explain the investment cycles and when you guys would need to make your next investment, because it's not a 'set and forget' system. Computing continues to improve. Then, my additional question is, we know that you're currently oversubscribed three times. Have you also done predictions as to how much public-interest compute would be needed in, let's say, five or 10 years, from a timeframe perspective? I'm thinking about this from a government perspective of how we get the investment into what you guys are doing. Mr Stickells : There's a lot to that question. I think in our submission we provided some information about the various roadmaps, the level of support that has been provided to date and the critical decision that is before government, and governments, in this area. In terms of timing, last year would have been better than this year, and next year is absolutely critical. We're right in the middle of a decision cycle that, if unmade, sees us move towards a winding down and a closure of this system. There is a period of transition that requires, if an investment is made in a new system, that the procurement, building and commissioning of that system does take a period of time. We're right at that particular juncture now. Ms MASCARENHAS: In dollars? Mr Stickells : Dollars in this area are sensitive to the market that we're in at the moment. It is an order of magnitude more than what we had to build the system several years ago. We commissioned a system—these are public figures—with a $70 million grant from the Commonwealth in 2018, and that enabled us to build the system that we have now. It was in the world's top systems when it was launched about three years later. It would be four to five times, I would estimate, to replace that sort of infrastructure again because of the changes to the computing market that we've talked about and some of the questions the committee has addressed today. It is of that order of magnitude. Ms MASCARENHAS: A question on notice is: I've sometimes heard data centres and supercomputers described as looking like windowless prisons. This question might be to both of you. Can give examples of how this infrastructure can actually look beautiful? Pawsey is probably one of the most beautiful supercomputers in, at least, the Southern Hemisphere. Mr Stickells : Through you, Chair—there's a little bit of local knowledge there from the member for Swan. CHAIR: Yes, I appreciate that. Mr Stickells : I acknowledge that as the question. I'm happy to take that on notice, but a very quick response is that it's very important that, as a facility—we are also a data centre. Pawsey and NCI operate small data centres and, as Ms Dennett referred to, we are part of that ecosystem. We've been operating for more than 15 years in the Perth suburban area. We've got decadal studies around the impact on our use of the aquifer, which is novel and seen as a very positive way of ensuring sustainability for this sector. We have been voted by our peers, alongside NCI, as the most attractive supercomputers in the world on the basis of the wonderful Indigenous art that we have on display. I think that's a wonderful thing for Australia. Prof. Rohl : If I can quickly respond as well, NCI is on the ANU campus and is just across the creek from all of the student housing. We've never had a complaint about noise or emissions. I think it looks not as nice as a policy centre, but it still looks quite nice. We'll share some photos of that. Ms MASCARENHAS: This is probably a question on notice for Ms Dennett. I find it fascinating that our public has a disconnect between loving cloud computing and not recognising there is a physical footprint associated with it. I come from a mining background, where the thing I sometimes see is that the worst player is the way that the whole industry is judged. It's great that you've got 86 per cent members. Hopefully, it'll be 100 per cent after four weeks of this committee hearing—let's see. The thing that I think is critical is the way that social licence is created, not only in community consultation but also in community benefits. Some of the lessons that we've seen for resource companies, where money gets handed out to communities, are that people are dressed up for balls but have nowhere to go, or community clubs have amazing uniforms, but no volunteers. So, it's okay for money to be given out to different community groups, but what would be great to see is Data Centres Australia—or it could be the Australian government—having a mechanism to make sure that the local community actually gets infrastructure that benefits them in the way that they want it. I think it will be place based in each scenario. That's just some thoughts on that. Ms Dennett : And they do. Perhaps they don't realise it, but it's the energy infrastructure, the water infrastructure, and building out recycled water infrastructure that can be used for parks, gardens, market gardens or golf courses, because data centres, for much of the time, don't need water. There is a lot that is already happening in this space that perhaps we haven't talked about enough. Perhaps people weren't interested. Certainly, we have reports out about that happening, in terms of the infrastructure, and the community funds and the like are also there, and those mechanisms exist. Certainly, our members are looking at what more they can do. But if I can leave you with one thought, it's that we hear a lot about the data centre sector funding a lot of things. There is a point where Australia becomes uncompetitive, globally, to attract that investment. We just urge the government to consider where that tipping point might be. CHAIR: That will now bring this session to an end. Thank you so very much for making yourselves available to provide evidence and answer questions of the committee. Those that we've indicated you'll provide on notice, please do so through the secretariat within the next two weeks. Proceedings suspended from 12:58 to 13:41 .