JIN, Dr Craig, Associate Professor, University of Sydney KEAY, Dr Sue, Director, University of New South Wales AI Institute NICHOLLS, Dr Rob, Senior Research Associate, Centre for AI, Trust and Governance, University of Sydney [16:06] CHAIR: Welcome. I understand that information on parliamentary privilege and the protection of witnesses giving evidence to Senate committees has been provided to you. Do you have any comments to make on the capacity in which you appear? Dr Jin : I'm also a member of the Centre for AI, Trust and Governance and I work with AI at the university. CHAIR: Do you all have a short opening statement? We'll start with you, Dr Keay. Dr Keay : Australia has a plan but no AI strategy and nothing that articulates our ambition and place in a future dominated by artificial intelligence. I believe we can afford to be ambitious and will face penalties if we are not. I'd like to challenge the narrative that Australia is too small and too late to the party to develop its own artificial intelligence. There's plenty of evidence that we have the talent. We're just not backing it up. Australian researchers publish in AI's most competitive venues at close to six times the global per capita average, ahead of the United States, Europe and China. We produce top talent, but we fail to provide them with the funding and access to compute required for them to succeed and build AI models and companies here in Australia. So we lose that talent and those future technologies offshore. I'd like to start with how AI and data centres intersect with our ability to conduct scientific research. Increasingly, every researcher is now a GPU user. AI is becoming foundational across all scientific disciplines. Researchers who cannot access compute in the form of GPUs cannot compete internationally. UNSW is a global, top 20 ranked university, and, if we can't attract and, importantly, retain world-class talent, our rankings will drop, threatening Australia's $55 billion education export sector. If we lose our scientific talent, we also lose our ability to conduct scientific research. Without that, we lose the ability to discover and innovate and the economic opportunity that entails. To me, this inquiry is an important opportunity to recognise the gap Australia currently has in AI infrastructure support, and the dearth of funding available for Australia to develop its own artificial intelligence, and tying that to the regulations imposed on new data centre investments. Our thinking on this has evolved since our submission. We propose that, in the absence of public investment infrastructure, developers reserve a defined share of installed AI capacity for public-interest use. This needs to be matched with something like a levy on infrastructure investments, which would be directed towards the funding and support of the development of AI capability and models on our shores, which currently receive zero effective support. I believe this also has the potential to positively impact our region. Australia exports energy to many of our neighbouring countries, who will struggle to secure the energy required to support the development of data centres and the compute that they contain, which is essential for artificial intelligence. So this could be the start of an important export industry for Australia in the form of sustainable compute powered by renewable energy and allowing companies to both use AI and still meet their net zero targets. But the benefits of this only accrue if Australian companies and government own some of the compute that is forecast to be built on our shores. Dr Nicholls : Thank you for the opportunity to give evidence today. With me is Associate Professor Craig Jin. We're both members of the Centre for AI, Trust and Governance, but we come from different perspectives. That reflects the fact that the university, like UNSW, approaches complex, contested problems like the impact of AI and data centre governance on a multidisciplinary basis. Our appearance together reflects that together—the University of Sydney and the University of New South Wales—we bring a range of expertise to bear on what are genuinely wicked problems. My focus is on regulation and governance—how we operationalise Craig's insight at a government level. And this is a rapidly changing environment. The landscape has shifted since the committee's terms of reference were established, not least because of the essay and recent calls for regulation last week by the CEO of Anthropic, supported by the CEOs of the other major frontier labs. AI and data centre governance can't be addressed by a single discipline or profession. I'm here to assist the committee in whatever way is most useful, including, potentially, answering the question raised by Senator Ghosh in the last session. CHAIR: Dr Jin? Dr Jin : I want to address two questions: how Australia should manage AI data centres that consume scarce national resources and what public value we should seek in return. My proposal is that government should decide first how much electricity, grid capacity, water and land Australia is prepared to allocate to AI infrastructure. Industry should then compete for that capacity, partly on the public value it returns. This reverses the present sequence in which industry sets demand and government accommodates it. These are more than large industrial facilities; we may be witnessing a new infrastructure layer above the internet, which I think of as a cognitive port. A task enters and is routed through a distributed hierarchy of models, specialist systems and computing resources. In time, it may resemble our electricity network—a network supplying an increasingly important resource. If that is where AI is heading, the question is not whether we host more data centres; it is who controls this infrastructure, who has access to the capability it produces and what Australia receives in return. Cost recovery is not public value. The productive resource being created is cognitive capability. Australia should secure meaningful access to that capability by building enduring expertise, compute capacity, research capacity and sovereign AI capability. We should also retain the ambition to build a frontier model of our own, not for the sake of an Australian chat bot but to develop the people, engineering knowledge and institutions needed to shape frontier AI. I do not believe it is too late. CHAIR: Thank you. Senator Ghosh, Dr Nicholls has already proposed that they could answer one of your questions from the previous session. Senator GHOSH: Perhaps we could do that. Then they'll give that answer and I'll leave my questions. CHAIR: Okay. Dr Nicholls. Dr Nicholls : I think one of the key things is that regulation—law—is never going to keep up with technology, and that's actually a good thing. Regulation needs to be fairly close to the edge but has the risk of running too far behind. One solution is one proposed by a joint select committee of the parliament, which is to have a digital duty of care which is broader than merely the Online Safety Act, essentially creating an obligation which can be then converted into a regulatory response so that the issues raised by both Dr Keay and Associate Professor Jin can actually be adopted into a regulatory framework rather than try to catch up from a regulatory perspective all of the time. CHAIR: Thank you. Senator Ghosh, do you have more questions now, or do you want me to come back to you? Senator GHOSH: Please carry on. I have no further questions. CHAIR: Senator Henderson. Senator HENDERSON: Can I ask the basis for the naming of your department? It is talking about trust and governance. Is that because you don't have trust and you don't believe there's appropriate governance? I'm just trying to understand the origins or focus of your work. Dr Nicholls : The basis of the centre's naming was simply to ensure that we continue—or commence—having trust and governance in the AI space. It also reflected the fact that it's a multidisciplinary institute. Trust in an engineering perspective might be quite different from trust in a humanities perspective and different from trust in a legal perspective. Senator HENDERSON: I want to ask a broader question. Isn't trust in AI a bit like saying we need to trust the internet? Isn't it all about understanding what exactly it is you're talking about? Dr Nicholls : Absolutely. The trust in AI also has a variety of things. Do I trust an AI chatbot? That's a sort of trust-but-verify position. Do I trust that the regulation of different AI players is working the way that's expected? That's a governance and regulatory issue, which might be quite different from 'Do I trust AI?' Senator HENDERSON: Where are you at in the body of your work? Where do you think the biggest challenges are in terms of both trust and governance? Dr Nicholls : I think the biggest challenge is that we've got a situation where, as you've heard before, we've had rapid rollout of AI data centres with a focus, as Doctor Keay mentioned, on the data centre. But a lot of the value stack—the value proposition—is actually: What are you doing with that AI data centre? How do you deal with it in a way that provides value to, essentially, social licence? Perhaps it is an actual licence. It provides value. So it's not enough, as Craig mentioned, just to say, 'Oh, well, if a data centre is energy neutral, provides a couple hundred bucks off local people's energy bills, is water neutral and doesn't impact the water supply.' That's not enough. It's actually about ensuring that there is a capability shared, whether it's the cognitive capability that Craig mentioned—some form of sovereignty in the compute—or goes further than that. The idea of having trust and governance in the centre's name is to track and recommend on those issues. Senator HENDERSON: Can you explain how you then extend that to understanding what the value proposition is? If you've got a local data centre in your community and you want to understand what value it delivers, because it's maybe not impacting on water or power—I'm trying to understand how this makes a contribution. Doesn't a data centre make a contribution by reason of the information it computes, which should technically serve all Australians? Dr Nicholls : I'll hand to Craig because he's the expert on this. But before I do that, I'll try and answer the question, and the answer is: it depends. If it's a data centre that is doing inference to support local use, yes, there's a contribution to the Australian economy, and it might be a very important one. If it's doing a training run, where that newly trained model is used or deployed is entirely in the hands of the person who's organising the training. And if that isn't made available to Australia or to Australians, then there's a potential mismatch, and it's getting that issue right. Dr Jin : I'd like to expound on that. We're at an inflection point at the moment. It's quite interesting in terms of this industrial development. The industry needs our scarce national resources. This gives an extremely powerful bargaining point for the government currently. So the question really is: what is that public value bargain? What do we get in exchange for our scarce national resources? And what you can see is it's really building a cognitive capability. It's an ability to predict, infer, design and generate and increasingly to reason and to act. And so the question is already being asked by government. There are privacy issues, there are copyright issues, and then there's just the general knowledge of society. This is going into these machine-learning models. And it turns out that words are one of the most powerful mechanisms for making models. And so our words are producing some of the most powerful computation engines on the planet. And what is the public value exchange that we get for using our scarce natural resources to produce this cognitive capability? Senator HENDERSON: Can I just ask again just a broad question. What is the value of having data centres here in Australia, particularly when you look at data sovereignty, national security and our defence capabilities? Do you recognise the importance of having data centres on Australian soil? Dr Nicholls : Absolutely, yes. But it goes beyond that. We absolutely need it for domestic use, but we also have the potential, as Dr Keay said, for it to be an export capability. Having it here but having a level of control over what is done here is really critical. Senator HENDERSON: But isn't that an academic question? If there's a data centre nearby, how do we know that they're not engaging in theft of Australian creatives' copyright, for instance? We went to AirTrunk, and the security just to put your nose in the door was incredible this morning. So how can we possibly understand what is going on behind closed doors? Are they working for good or for evil? Dr Nicholls : That's a hard question. But the potential way of dealing with it is: are they working for us, or are they working for the interest of the multinational? Senator HENDERSON: Okay. So how do you answer that question, then? Dr Nicholls : By having a requirement to share some of the additional value that comes from the systems which are in that data centre—the compute and— Senator HENDERSON: So when you say, 'having the capability to share those systems'— Dr Nicholls : A mandated requirement that there is sharing. CHAIR: Can you just unpack what that means a bit more for us, though? Senator HENDERSON: Some people don't understand. CHAIR: Does it mean a value of profits? Are you talking about computing capacity? Dr Nicholls : Yes; I want that capacity. CHAIR: Can you unpack that? We are not— Senator HENDERSON: How do you share it? Just explain. Dr Jin : I think, as doctor Keay said—think of it as infrastructure. When you have energy infrastructure, you're not saying: 'Who owns it? Who gets the benefit?' There is energy infrastructure, and Australia needs to get benefit from that energy. We are now making a cognitive-capability infrastructure, and Australians need to get that capability. It would be very odd, I would have to say, if we didn't have local capability that understands how AI models work, how they should be generated and how the infrastructure should be developed. If we're going to have a distributed and hierarchical model supporting our country, we need the expertise to understand how that works, how the knowledge engineering should work and how the capabilities should be laid out. There are many, many open questions, and we need the expertise developed here, and that expertise stems from developing the capability here in Australia. Dr Keay : Can I have a go? CHAIR: Yes. Dr Keay : In terms of value, there could be zero value from having data centres here, depending on who owns what's in them. So it's the tenants of those data centres, and, as you have found from your visit to a data centre, it's very hard to get any information about who those tenants are and what they're doing in the data centre. And that might not change. However, on the point you made about data sovereignty and national security, at least if they're Australian companies under Australian jurisdiction, you don't have to worry about the US Cloud Act. You could have more guarantee that data sovereignty will be protected, which you are not if you are dealing with people that own the infrastructure that are not Australian companies. It is very difficult to unpick what is actually going on inside the data centres, but what we do know is how much energy they are using. You can make a lot of assumptions about the type of AI loads that are being generated just based on energy usage. If you were to mandate to operationalise data centre expectation 5 around making sure that some of the compute inside those data centres actually benefits public interest users, then you really could just go off the energy usage. Then, as long as you make sure that obligations flow from the data centre operators through to the tenants, you could be putting a hypothetical levy on the amount of energy that is being used and convert that either into cash or into GPUs that could be used for Australian public interest purposes. But, apart from that, it's probably really hard to unpick exactly what's being installed in the data centres and what they're being used for. The most effective way, and the way that already is essentially being reported on, is how much energy is being used. CHAIR: That's how their capacity is measured at the moment, obviously for that reason. Dr Jin : But you lack visibility. CHAIR: Yes. We did ask directly who their tenants were, and they were very reluctant to tell us today. Senator Pocock? Senator DAVID POCOCK: Thank you very much for your time and the work that you're doing. Dr Keay, asking that the government departments why we need these data centres for training here, they say it's because we need access to the frontier models. We already have that, they tell me, but we need more access to them, and this is the only way to ensure that we have that access. Does that argument stack up? Dr Keay : It sounds pretty circular. To protect ourselves from US frontier models, we need access to US frontier models—is that the argument? Senator DAVID POCOCK: Essentially, yes. Dr Keay : Obviously, it would be good from a security perspective to have access to a whole range of models. But you would be putting all of your eggs in one basket if you were to rely on only being able to access US frontier models. You actually need to be able to access all types of models. You can have access to open weight models and the US frontier models, if available, but we know that they can be closed off anyway, so we probably have to plan for the eventuality that we may not have access to them anyway. But, from a pure cybersecurity perspective, this is never just a software issue, because you require people to make sure that infrastructure is safe. In terms of protecting our critical infrastructure, it is people that have to go around and put in the security patches that make sure equipment is properly air-gapped and that will protect our critical infrastructure. And, yes, they need to understand the potential threats from those AI models, but the answer is not solely those AI models on their own. I think, if we had to, we could manage without access to those US frontier models. At the moment, the US frontier models are just making it easier and quicker to identify vulnerabilities, but we haven't seen a huge uptick in cyber threats such that you might anticipate if it really was at the AI-extinction level of risk that seems to be the idea flowing around. Senator DAVID POCOCK: The second part, they say, is that, if we get training data centres here and these big AI companies potentially basing teams here, then we have an uplift in capability. But it seems like that's only going to happen if we mandate it, and otherwise we'll just host the GPUs, and they'll be used from Silicon Valley. What do you say to the capability side of things? Dr Keay : How do you mandate collaboration? They don't have to give Australian researchers access to have a look under the hood to really evaluate their systems and get a good understanding. There'll probably be commercial reasons why they wouldn't want to open that up. We've had quite productive discussions with some of the alternative-model companies like Mistral in terms of legitimate or seemingly sincere collaboration opportunities, where they will send their research scientists out to work with us if we can find the right mechanism to make that happen, but to date I'm unaware of any opportunities that foreign researchers have been given to really have a look under the hood at how those US frontier models are being developed. I'm not sure how you would put in place any sort of enforceable obligation to make that happen. In terms of how that uplifts AI capability, it's unclear. Just because they train frontier models on our soil doesn't mean it will uplift our capability. Dr Nicholls : What does uplift our capability is what Craig was proposing: us having a capability locally. That is important for us—by 'us', I mean not just academics but Australia—to be able to understand what the cyber threats that Dr Kay mentioned are and what they are likely to be not now, because zero day is too late, but in a few months time to allow for proper patching to prevent those issues. Having local capability and local access to capacity are both critical. Senator DAVID POCOCK: Why do you think there's such a disconnect between researchers and industry that I hear from here in Australia who say 'we can absolutely build capability here, and in fact we've got some amazing companies who are working in this space' and the federal government, who just has big eyes for Microsoft and Anthropic and OpenAI and Google. What's going on there? Dr Nicholls : I think you can actually do both. In order for us as a nation to be able to understand what we are doing in negotiations with Anthropic or OpenAI or any of the others, we actually have to have the capacity to have that understanding. To be able to answer Senator Henderson's question, we actually need that capacity locally and that capability locally. Dr Jin : I'd like to respond. Just like Australia's Opal nuclear reactor—we had Argentina help us build it. For frontier AI capability, you need hundreds of thousands of GPUs. Australia needs its own frontier AI models. Your peer governments in the UK, France, EU, Canada, India and China are all pursuing such capabilities for similar resilience and innovation reasons, so it can be justified as a capability-building instrument—people, systems engineering, safety research, data governance, and the ability to train or adapt high-end models under Australian control. CHAIR: There's wanting Australia to have our own frontier capacity. In terms of what we actually need—I'm talking about the way it's currently measured in terms of energy use—what is the understanding of what Australia actually needs in terms of capacity? To come back to your original point, Dr Jin—which was that we need to be asking ourselves what we're prepared to power and how much power we're prepared to spend on these facilities and all of this other stuff that therefore flows out of that—just for Australia's use, what is that capacity? Dr Jin : I think the best way to think about this is again of an infrastructure. I'd like you to think about doing some work. A cognitive question goes out, and it gets routed. You would want that routing to stay within government AI perhaps and not go out. Let's say you have privacy concerns. A question about health goes out. It could be tagged on a layer above the internet. This question goes out. It could be controlled. Let's say you have a question. You have a natural disaster that's happening, and you want some solutions. You would want urgency for that compute. You would want the energy to be used for that compute to go out. You would like to be able to regulate the infrastructure of that cognitive layer. Having not sufficient regulation and not sufficient control would be the same thing as if you didn't have sufficient control over how energy is distributed within this country. It's going to be cognitive capability you can start to think of as—your constituents will be asking for it, as they ask for clean water, for energy and for the internet. CHAIR: Yes. We can't just tell people to log off. But it's still not— Dr Nicholls : It's a difficult question to answer; otherwise, we'd answer it. You're going to need some capacity which is local—some, hyperlocal. So I wouldn't be at all surprised to see lots of Telstra telephone exchange buildings hosting small inference-level data centres. What are they going to be? Maybe a megawatt? They might be two, but probably one. There happen to be 483 local exchange buildings which would be within four kilometres of a bit more than 85 per cent of the population. That looks like a gigawatt, but it doesn't look like a gigawatt data centre. So the answer is quite difficult. You need large data centres for training. The power that's used is really, really high—in the gigawatt range—while the training runs there. But one of the real problems for grid capacity is: when the training run ends, what's the demand? Pretty close to zero. It's a very hard thing to answer, because it's not stable, it changes, and it depends on what you're trying to do. CHAIR: I guess that's why I asked our previous witness whether we actually are at risk of ending up with a whole lot of stranded assets. Dr Keay : Potentially. Sorry, was your original question around our ability to potentially build AI models? CHAIR: I understand there's a pitch for that, but I'm trying to understand: even without that, what do we think Australia's need is now, and will be into the future, for what Australia needs rather than for just training US models? Dr Keay : From a purely research perspective, we already have more than three times the demand for the national computing infrastructure that exists, and we know that that demand is probably reduced, inasmuch as people don't bother asking. All up, researchers in Australia probably could access a maximum of the equivalent of 2,200 GPUs. To do model training at scale, you probably need to have experience on clusters of 10,000 GPUs, and, to do that, our researchers go overseas. So they're not even asking for access to that type of compute. We could give rough figures, based on what we know now, but, increasingly, because science discovery requires access to artificial intelligence, that's likely to only increase. But then how you map that against the proposed data centre build-out is probably a little bit hard to say. We haven't invested in upgrading our computing infrastructure here in Australia since 2018, so we have a lot of ground to cover. Even if there were to be excess supply here, and if we were able to get access to that onshore, it's hard to see how researchers would be complaining about having too much. Dr Nicholls : The stranded-asset risk is difficult to quantify. I'm, sadly, old enough to remember the dotcom bubble, and when that burst there was the potential for having all of these stranded assets. But the infrastructure part remained; the capability, in terms of subsea cables, remained and became really important for the next stage. So what might look like a stranded asset today might still look like a critical part of the supply of services tomorrow. CHAIR: Senator Pocock has a final question, and then we're going to have to wrap. Senator DAVID POCOCK: One of the things that the government has started talking about when it comes to the need for data centres, and the need for a lot more of them, is national security. If you look at some of the national security investments from the government, they're spending a few billion dollars investing in quantum computing and they're spending $260-plus million on a few submarines. Where do you think this ranks in terms of an investment in national security and sovereign compute? Dr Keay : I don't think we have visibility of how much of that is actually going to be directed towards national security, do we? It's zero public investment, so I'm not quite sure how that translates. Senator DAVID POCOCK: Do you think there should be public investment in compute? Dr Keay : We seem to be the only country in the world that thinks you can go through the AI transformation with zero public investment, but that's where we are. I think perhaps we spend $30 million a year on artificial intelligence compared to an economy like South Korea, which has a smaller GDP than Australia, which is investing the equivalent of A$212 billion. I think we have access to about 2,200 GPUs; South Korea is investing in 250,000 GPUs. Like I said, without having a strategy in place about exactly what this means to national security, what's required for our national security agencies, what that looks like and how we would be building up our own capability, it's hard to answer whether it's a sufficient investment. At the moment, it appears to be an investment of zero and hoping that somehow, by these foreign investments coming in, we will obtain some benefit. Senator DAVID POCOCK: Thank you very much. CHAIR: Thank you, everybody. We appreciate your time. If you've taken any questions on notice or have any further information, please send it through. That concludes our hearing today. Thanks to all the witnesses who appeared, to Hansard and Broadcasting for your assistance and for dealing with our tech issues throughout the day, to our secretariat for all your coordination and to the local community who helped host us here today. Thank you very much for your time and patience as well. Committee adjourned at 16:41