CHAU, Dr Wesa, Executive Director, Per Capita D'ROSARIO, Dr Michael, Director of Econometric Research and Analysis, Per Capita [by audio link] LEWIS, Mr Peter, Convener, Per Capita Committee met at 11:11 CHAIR ( Ms Briskey ): I declare open this hearing of the Joint Select Committee on Artificial Intelligence. I begin by acknowledging the traditional custodians of the land on which we meet and pay my respects to their elders past and present. I extend that respect to any Aboriginal and Torres Strait Islander people joining us today. The committee will be conducting today's hearing in person and via teleconference. These are public proceedings being audiostreamed live via the parliament's website, and a Hansard transcript is being made. I remind all witnesses that in giving evidence to the committee they are protected by parliamentary privilege. It is unlawful for anyone to threaten or disadvantage witnesses on account of evidence given to a committee, and such action may be treated by the Senate as a contempt. It is also a contempt to give false or misleading evidence. Witnesses also have a right to request to be heard in camera. If a witness objects to answering a question, they should state the ground upon which that objection is made, and the committee will determine whether it will insist on an answer, having regard to the ground which is claimed. If the committee determines to insist on an answer, a witness may request that that answer be given in camera. I remind senators and members of their obligation under the Behaviour Code for Australian Parliamentarians to treat witnesses with dignity, courtesy, fairness and respect. I now welcome representatives of Per Capita. I understand that information has been provided to you on parliamentary privilege and the protection of witnesses. I now invite you to make a very short opening statement, noting that you have provided a written statement for the committee to table. But, if there are some brief opening remarks that you're able to provide us to set the scene, please do so, and then we can get underway with questions from members. Over to you. Dr Chau : Thank you. First of all, Per Capita is an independent think tank. We're committed to increasing economic equality. Our submission focuses specifically on how any benefits that accrue from the development and diffusion of AI are shared by all Australians. Many people look at Australia's gas industry and ask why so little of the wealth generated from our resources stays in Australia. If we look at a comparison, which is the Norwegian sovereign wealth fund as the alternative, the difference is that the Norwegians designed a system of public benefit at the point that oil was commercialised in the North Sea. We are at a similar moment today where we see the potential for Australia to be a global leader in training of AI models. At Per Capita, we recognise that this is a contested future and that it is contingent on many variables, but that does not mean that we should not be thinking about economic models that maximise shared values. We also accept that copyright of creative work remains a barrier to training of AI and urge the AI industry to negotiate appropriate licensing agreements with Australian creatives. Our submission doesn't claim to fully solve the complex economic and social challenge. What we have done is work with partners across civil society to develop a series of scenarios, linked to credible scenarios on the pacing of AI development, in order to model different value capture models. We also look at global attempts to capture value from both AI and other resource extraction industries, where governments have partnered with investors to develop new industries. Now I'll pass on to Per Capita's board member—and also a convener of the Centre of the Public Square—Peter Lewis, who convened the roundtable, to take you through some of the recommendations. We also have Michael, our director of econometric research and analysis, to answer any questions as well. Mr Lewis : Thanks, Wesa. Some of you might know I've got many views on AI, but today's contribution is specifically focused on designing a framework to maximise benefit to Australians in the event that a training capacity is developed in this country. Again, I want to put on the record that Per Capita stands with creators. A theft is being committed and, as our submission outlines, copyright issues need to be addressed on both economic and moral grounds. But this submission lays out what comes after the resolution of those issues. We started with a simple provocation: how can we be more like Norway in actively engineering arrangements for a new industry to maximise benefits for all Australians? I'd love to tell you there's an easy fix, but this submission is actually about how complex value capture is when we don't know what the future actually holds. Our submission models a series of seven different pacing scenarios, from slow burn to take-off, and then looks at how value would be captured under each scenario. We also looked at how other nations have approached value capture with AI and resource-extractive industries more broadly, and this is laid out on page 21 of our submission. What we came up with is a suite of interventions which could scale up or down depending on the velocity and the trajectory of the tech. For instance, it could be that everyone's using LLMs. It could be a diffusion of open-weight models. There'll be different ways of capturing value from those. That's outlined on page 28. So we don't support one simple instrument or policy lever, but we advocate what we call an economic partnership framework with the AI industry. This would include elements of public ownership, revenue sharing and access fees to ensure the benefits are shared between all Australians and targeted to meet specific needs. I'll quickly go through some of those levers, and then, if you want to ask specific questions of Michael, he can go deep. The first idea is shared ownership and governance of the fixed assets—data centres and renewable generation—through an existing or novel entity such as the Clean Energy Finance Corporation—that sort of model. The second would be revenue apportionment, linked to the global value generated by the model. If we're training these models that are going to be dispersed, how do we not only get a one-off payment but follow the tail? So we look at ideas like a token tax, which isn't the answer but gives you some ideas of how you could build an ongoing value train. We also talk about an access fee for the training privilege to underwrite a public endowment fund that would specifically look at areas that are going to be structurally disrupted by AI, such as journalism, research, education and, of course, creative work—again, without undermining the current copyright process. I would particularly point committee members to appendix B of our submission, which highlights not just the value of the role creative work plays in the quality of models but also how, if the industry fails either because of the unresolved copyright issues or because of displacement, it actually creates the effect that the models aren't going to be as strong anyway—that idea of Habsburg AI. We also argue for an allocated reserve sovereign compute, guaranteeing access to government and public institutions. Finally, a training and apprentice contribution is needed to address the skills gap—probably delivered through the TAFE network. This all might sound a bit greedy, but the point is, if we take these companies at their word that they want a genuine partnership with Australia, then we also need to design a genuine and enduring upside for all of us. We know we can't trust these companies to properly log their Australian billings and pay tax, so we need to be a little bit more assertive in how we build the national interest into any business model from the outset. As our report points out, the best time to lock in the models is now, before the industry develops. That's the real lesson from Norway. Whether the future growth is steady or transformational, ownership, investment and a shared revenue model should be central to our national interests. We're happy to take questions. Michael's the brains of the operation, and he's on the call, but feel free to go wherever you'd like. CHAIR: Thank you, Mr Lewis, Dr Chau and Dr D'Rosario. We really appreciate the quite fulsome submission that you've made to this inquiry about a very critical area, how we actually genuinely share the benefits and that they aren't just concentrated in certain hands and in certain places around the globe. Increasingly, that is a live issue, as is how we genuinely leverage off the opportunity that is in front of us at the moment, noting, as you did, Mr Lewis, some of the policy challenges that we've got to get through. I'm particularly interested in your thinking around sovereign compute capacity and how that would work. As you say, there are different elements across the stack in terms of value capture, from the actual infrastructure to the compute, the potential models and a range of other things, particularly capacity for public institutions, universities—that opportunity really to elevate our leading research. We've got some pretty incredible people in Australia. How do we get them the access that they need to that? How large do you think that reservation would need to be, and how would we ensure that it's supporting genuine public-interest outcomes? Dr D'Rosario : Excellent question at the outset, and it's certainly not one that has been raised enough. Fundamentally, the challenge will relate to available compute capacity. In an ideal world, we would have sufficient capacity allocated to the national science agency and potentially both private and public institutions to engage in very advanced SLM developments. Many of us currently working on SLMs and lagging edge frontier models are largely doing so through foreign partnering. As a case in point, I'm working on SLMs with groups abroad that are, for all intents and purposes, trying to provide support to undersupported languages, a really valuable thing. It would be wonderful to be able to do that in Australia for our incredible First Nations populations and languages. Having reserve compute sufficient to ensure in the short-term the development of really viable agricultural SLMs, culturally focused SLMs and in particular R&D aligned SLMs would be incredibly valuable. To the extent that there is sufficient capacity, having some allocation to the national science agency to engage in the development of a national focus for the base model for the public benefit would be great, but, principally at this point, we do not know how much compute will actually be committed. We've seen a lot of commitments in terms of infrastructure, and a number of players have suggested that they will relocate capital to make this happen. It is hard to conject as a percentage how much that should be until such time as we actually get to a point of being able to do so. For all intents, it would be wonderful to have access to several thousand GPUs annually for the public good. That's not an inconsequential cost, but I would imagine, for the frontier providers, it's a relatively minuscule allocation of compute. That would be the ideal allocation as a starting point—3,000 to 4,000 GPUs available all the time with an allocation but, beyond that, a substantially larger tranche for discrete SLM development as well. It's purely a conjecture at this point and based on a very modest interpretation of what these partners are likely to bring at this point. I would certainly like it to be vastly more than that in in the future. To use Peter's phrasing, we don't want to be greedy at the outset, but I happily will concede that the vast majority of the work that we're doing presently with SLMs and lagging-edge work in the context of frontier development is through partnership. I work with the London School of Economics and am able to collaborate with teams there to do this kind of work. It'd be amazing to be able to do this kind of work locally. CHAIR: Stepping back out—that was assuming that we can do that—what's the opportunity cost then if we don't? I'm trying to wrap my head around taking the opportunity, leveraging the opportunity that we have now, but, if we aren't successful in that and if we do see these hyperscalers decide to take their infrastructure, compute and training investment elsewhere, what does that then potentially mean? What do we then lose out? Are we genuinely losing out on that potential to get sovereign compute capacity, or do we have other ways? What's your sense of the opportunity cost in front of us at the minute? Dr D'Rosario : There are costs, the costs of material, both in terms of our capacity to drive national R&D programs and innovation and in terms of industry. I need to concede at the outset that we are seeing one of our scenarios presented emerge—perhaps one that wasn't in the original go science framework. We talked about edge computing and the emergence of edge computing. Yesterday, we saw Nvidia and Microsoft make a major announcement that will allow people to run their frontier operating compute on local machines. Desktops will potentially have the power to do some really powerful things over the course of the next three to six months, allocations notwithstanding. It needs to be acknowledged that we are seeing some dramatic changes in this space, and increasingly we will be able to do very, very powerful things using local AI models. At the frontier, there is undeniably going to be a need, in terms of sovereign defence capability and our capacity to rapidly improve and iterate on national infrastructure, to develop cybersecurity strategies to ensure the defence of our creative assets. There is still a scope for genuine loss. Yes, we can co-opt larger AI models and build on top of them, and the work of Mistral in particular in recent times is compelling in that respect, and Europe, as of a week ago now has a frontier model player, Le Chonk, being released. It's a rather funny name but a really impressive model. We are now in a position where we can no longer say that we're not the only ones that don't have access to these technologies. There are still certainly a number of others, but increasingly, we're seeing other continents develop this capability. Operating as a laggard, given the impact of these models on cybersecurity, would be challenging, both from a learning standpoint and an advancement standpoint. It is a matter of national security in that sense. I will also happily concede that we are seeing immense progress when it comes to open rate models and what we can do locally, and the ability to tune those models to benefit our farmers and to benefit our R&D programs is very, very strong. We would undeniably be on the back foot in terms of future R&D and in terms of the opportunity to drive innovation. CHAIR: Thank you very much, Deputy Chair. Senator DUNIAM: Thank you for a very fulsome submission, as the chair said. I want to go to that issue about reserving compute capacity. I think you mentioned a range of sectors, but I wanted to focus on government. Who in government do you think compute capacity should be reserved for? Are we talking the tax office? Are we talking the defence department? Who do you envisage would be the beneficiaries within government of such a reservation scheme? Dr D'Rosario : I'm happy to share my personal views to this, but, Peter, would you want to volunteer any thoughts first? Mr Lewis : No. You go. Dr D'Rosario : I think you've got it in one—the department of infrastructure, defence industries. Not everyone will agree with Defence having that initial allocation, but I think it's paramount that Defence is afforded sufficient resources here. Our national science agency, obviously—we have seen something of an attenuation of capacity there by virtue of the reduction in the size of the national agency. CSIRO is a wonderful agency with immense capability and incredibly talented individuals, so bolstering that capability with this allocation would be exemplary. We've seen their capacity for innovation for decades, but they have not been afforded sufficient resources to benefit from this most recent wave. Undeniably, defence industries and our peak science entities would benefit most. Certainly, thereafter, the university sector should also be afforded sufficient allocation, but let's also acknowledge the benefit to assigning some allocation to TAFEs as well. The TAFE sector and TAFE aligned industries are going to be critical players in the context of this intelligent transformation. In the context of SLM development, TAFEs working with universities to develop viable agricultural SLMs, viable R&D SLMs, would make a great deal of sense. But, principally, without a shadow of a doubt, it would be our defence industries and peak science bodies. Mr Lewis : The one other thing I would add is that a reserve compute would also provide the opportunity for public-interest innovation and building local industries around things like health data, which is vested in Australia but underwritten by access to that compute. There's a bunch of great Australian companies looking for opportunities to build off the back of public databases that they just don't have the opportunity for at the moment. Dr D'Rosario: That's right. I work with several really, really talented Australian groups that are developing some remarkable technology and that are finding footholds in both Africa and Europe, but much of their capacity to develop models is being curated and supported within European contexts, even though the innovators are Australian. This is particularly true in the context of computer vision and the work that some Australian health tech innovators have been able to do in that space. Undeniably, there are significant spillover effects for everyone if we can get that compute allocation. Senator DUNIAM: The reason I ask that question is that, ultimately, we will have to make recommendations around good ideas and how government might administer them. It's one thing for us to say that there should be a reservation set aside for government, because the next step is government figuring out who within government might benefit from that. I daresay there are members of the community out there who don't think that the ATO, for example, should be a beneficiary—but that's another matter. I'll move on. My final question is in relation to the investment fund. Who specifically is on the list of people—entities and individuals—that should be able, prescriptively, to access this endowment fund you talk about? Mr Lewis : Michael can go into more detail, but we identify particular areas where there will be foreseeable disruption that will impact the public good. Part of our proposition is, even if you sort out copyright with some sort of licensing arrangement, there is still going to be industry disruption at the end of it. If we are a nation that doesn't have a vibrant media industry or a vibrant cultural industry, or universities that are resourced to do their job, then that will be—I don't know. You were talking about opportunity cost before. That's just a 'cost' cost of this industry emerging. At the beginning, we need to think through how, as the industry scales—if it scales—those benefits are accrued across those specific sectors, which are critical for our liberal democratic values and not allowing them to be undermined. Again, I think our paper is more a provocation for you guys to think through the complexity of the frameworks, rather than a prescription saying, 'It should sit here or there.' With respect, it's probably your job to come up with the model for taking the ideas and then operationalising them. Senator DUNIAM: With respect, let me tell you that we're not the subject matter experts here. This is why we have these inquiries to get this information. Thank you very much. CHAIR: I understand Senator Pocock has some questions. Senator DAVID POCOCK: Thank you very much for your submission. You talk a lot about value capture, which is great, and we've seen the federal government take an equity stake in big projects like PsiQuantum. You argue that they should be doing the same thing here. Could you talk us through what that would look like. Who would they partner with? How much would it cost? How urgent is it to do some deals like this now, as you said, while things are getting started, not down the track? Dr D'Rosario : That's an excellent question as well. Specifics in terms of calibration are very complex. We could go down the path that we have in the past, in terms of the CDC being a passive holder and generating dividend. There's certainly nothing wrong with that, but there's uncertainty if it brings with it a lot of benefits. But taking a more precise and potentially more considered and measured approach in terms of plausible JVs with hyperscalers or engaging in directive holdings through a new direct holding entity could potentially be really beneficial. It really depends on government's appetite for risk in this context. Holding through a fund brings with it some benefit, but we certainly can't bring it under control with outcomes that you want to, so some direct holdings in terms of that compute and inference would be really, really compelling. There's also a lot of knowledge transfer that happens through a JV, through a line of ownership that doesn't happen as much through more passive and indirect investment. That is not to say in any way that the existing approach is poor, but it would make a great deal of sense for any of the existing holding entities, potentially even an entity like NBN Co, to actually get involved in some way or in some capacity. We have the vehicles. We have the models. It's just a question of where it should sit and the extent of investment. I would imagine that taking the approach that we took originally with the CDC, in terms of equivalent percentage holdings, would be compelling, albeit now they represent a really modest fraction of the overall sector, so trying to acquiesce that level of investment is not inconsequential in terms of budget. So one would hope that there is a response from the hyperscalers that they do want to provide us with a slightly discounted position or, alternatively, that we can utilise some long-term fundraising mechanisms to make this possible, because this is a long-term play and the investment shouldn't be exclusively in compute. Everyone's talking about compute, but, with the emergence of these distributed models and what I would describe as being hybrid models, we need to think really strongly about inference holdings as well. I think there's a real opportunity there because the returns are going to be strong and consistent and potentially not like anything we've seen since the 2013 mining boom, in terms of the opportunity both from an investment and return standpoint. Again, it's an imperfect answer, but I guess the response would be: do we hold directly, do we hold passively or do we use an existing vehicle through some sort of JV model to be able to acquire some intellectual property through the process as well? And I would advocate for any of those that could be subject to government's appetite. Senator DAVID POCOCK: We've heard about one of the challenges. There are obviously companies that are building data centres [inaudible] space where the GPUs are used by mainly multinationals, or international companies. I asked AirTrunk about this. They said that they had no local tenants. I have two questions. Is it investing in the companies that have the data centres or in the companies that are utilising them and essentially owning GPUs and compute as tenants? Dr D'Rosario : I'm not going to deny the fact that the researcher, the academic and the educator really would like us to have dedicated ownership of the infra and the GPU, so it can be used for so many compelling purposes, and certainly startups and scale-ups could potentially access an apportionment or a term. That would be absolutely incredible. But we know the costs are really consequential. The most likely outcome would be that we try to invest in the value chain in a couple of ways. One through a modest amount of direct holdings, and then the other through a modest amount of co-investment with both local and international ventures that want to invest and develop. We do have some really amazing players in this space, but they have not been able to raise sufficient capital to co-invest in a major infra project. So, to your point about AirTrunk, yes, we don't have anyone that's ready to do that yet, but perhaps with co-investment and government presenting as a plausible co-investor we would see groups come together and we would see a crowding-in effect of local capital as well. To answer the question in short, undeniably all the way along the value chain it would be wonderful, from a research standpoint, to own some of those GPUs and to have a [inaudible] program. From a pure investment standpoint, it would make a great deal of sense to facilitate some degree of co-investment with these partners, and they appear to be open to that. Certainly within other jurisdictions they are investing heavily with an Australian [inaudible] as well. Senator DAVID POCOCK: One of the concerns raised yesterday was that, if you did have some sort of compute reservation policy or agreements where the big hyperscalers had to provide compute to Australian researchers and, potentially, startups, you'd undercut sovereign owned compute where they can then not sell their compute because people are getting a discount on compute that's owned by multinationals. What do you say to that? Dr D'Rosario : That's a really good point. I'm not wanting to contest the logic of that claim, because there's certainly some sound logic that underlies it. At certain points in the future, I think that that logic would undeniably hold. I would say, at this particular point and over the course of the next three to five years, the logic perhaps doesn't hold in light of the significant demand that appears to be untethered within both the domestic and international markets. There are no players that are operating at sufficient scale to come together to access that compute locally. We simply do not have the enterprises operating that way. If they were able to access apportionments, there would undeniably be a strong level of interest. All and sundry within the market are aware of that. So I think the logic holds in the future, and I think we need to think really logically about future allocations, particularly as the cost of compute declines, as the demand for compute stabilises and as we see diminishing returns to scale—maybe diminishing marginal returns to scale—on future frontier model training. In the short term I think that there wouldn't be no challenge associated with allocating that capability. We've already heard about some of these multinationals leasing capability from one another at fairly exorbitant amounts. If the compute is available, certainly the returns will be there in the cycle of investment that we would look at now. Senator DAVID POCOCK: I know you said that you have concerns about copyright—this is very much if that can be dealt with—but you also talk about a fund for copyright. I've got a bunch of questions. How do we ensure that there's enough money there? How do we ensure that enough of it stays in Australia? We had testimony a couple of days ago that 90 per cent of that would potentially have to flow overseas to overseas creators because of the way that copyright works. We had concerns raised about Australia becoming a data-laundering hotspot. I also have concerns that you may have Anthropic and OpenAI paying into this fund, but if things don't pan out for them, if it turns out that people more and more are using open-weight and open-source models rather than paying through the nose for access to frontier models, then what? How do we ensure that we're not just throwing creators under the bus to potentially— Dr D'Rosario : Agreed. That's a really excellent point. I should mention at the outset that what we're proposing is over and above any kind of copyright claim. Essentially, we're noting the fact that, with the emergence of these models, we're going to see an inordinate amount of content generated. That's going to result in a crowding-out effect relating to human content, which will make things more challenging for human creators. That's something that we just have to acknowledge. We're already seeing it. We have seen some modest countering with music through reducing the amount of content that can be produced on platforms like that, but the floodgates in some respects are starting to open. So this is not a recompense for copyright, per se; it's a mechanism acknowledging a market failure that will arise, a crowding-out of human creativity that will need to be supported in some way. But it's not just creatives. We're talking about all manner of human activity that will continue to need to be supported in the future. We think that the copyright mechanism that is in place—that's the mechanism that the government has spoken about negotiating on, and so our stance is to continue to stand with creatives. To Peter's point earlier, I think he was talking about how this corpus could be used. This is essentially just an acknowledgement in the same way that—how do I put it best? We've had funds in the past that have existed, and, in acknowledgement of the unique challenges associated with a timely transition, then we would like to think that an apportionment separate to the core copyright compensation could be put in place. Again, this would be something that would need to be negotiated carefully with hyperscalers, but I think they too see immense value in human creativity, and they too want to see a continuance of that because it actually benefits models, particularly at the fringes and at the edges. Otherwise, we end up with a really peaked distribution as opposed to having those wonderful outliers in terms of content that creative humans can come up with. So it's not actually a copyright compensation mechanism. It's an acknowledgement of the human crowding-out effect. It's bad, but— CHAIR: I just need to move on to some other committee members who've got some questions. Senator Sheldon. Senator SHELDON: I just want to be helpful. Again, thank you for the weighty submission. Obviously, it's particularly helpful for the committee. You've touched on a number of aspects contained within your submission. It might be helpful for the committee but also for those that are listening in just to go through and look at it from this aspect—your report and submission—particularly to Mr Lewis. Last month, the Department of the Prime Minister and Cabinet released its consultation paper called Getting it right: building AI infrastructure that works for Australia. Those submissions actually close tomorrow, but the paper echoes themes present in your submission—access to compute for Australian researchers using AI infrastructure investment to fuel training programs and skills development in this emerging sector and regulating foreign firms who want to train their frontier models in Australia. It would be helpful to us if you could tell us which parts of that layered dividend capture model the government already covers in this paper and what else you would like to see, if you could capture it in a clear way from your submission. Mr Lewis : I'm not across that paper. I don't know if you've read it, Michael. But, from what you've just said, I guess the missing pieces are those initial pieces around the idea of co-investment in the infrastructure and, secondly, the value capture. So the real concern, I think, is that—and you hear it a lot—even the most optimistic vision of the development of training in Australia is that we're going to build some data centres and we might get a little bit of tax revenue and a bit of economic activity, and that's it. So, if we are going to be a base for global training of data, how do we follow the tail of that benefit? The analogy is the way Saudis make oil. If we are going to be developing this, how do we ensure that the value capture is embedded at the bit of value that travels, which is the tokens? Senator SHELDON: Dr D'Rosario, it might be of assistance if you could add to that, just so that we can get a feel for what is common ground and what is ground that still needs to be covered, in your view. Dr D'Rosario : Undeniably, there's a lot of common ground and [inaudible] aspect— CHAIR: Sorry, Dr D'Rosario. We've just got some technical difficulties at our end; we're struggling to hear you. We're getting an echo. Dr D'Rosario : Is this better? CHAIR: I think we got rid of the echo, yes, thank you. Continue. Dr D'Rosario : Yes, there is a lot of common ground. Certainly, I think, government acknowledged a number of the key considerations. But, to Peter's point, in terms of where there is scope for additional uplift, it's certainly in the context of— Senator SHELDON: Sorry, Dr D'Rosario. It'd be of assistance to the committee if you said where the common ground was and also where the difference is, definitely. Mr Lewis : Would it be easier if we took that on notice and came back to you with a bit of a chart? We'll go through that and then identify those points of difference. Senator SHELDON: That sounds sensible. CHAIR: I appreciate we're right on time. I know Ms Chaney and Ms Mascarenhas both have questions, so we might quickly get them on short comment and maybe take the rest on notice. Ms Chaney. Ms CHANEY: We're an attractive location for data centres but not the only one. If we're too demanding, we could scare off foreign investment. I'm interested in hearing how you weigh up our bargaining power and our ability to be demanding. Dr D'Rosario : I would say at the outset we have some scope for negotiation. I think in the long term we have far more scope than many would volunteer. I think that at some point we will see this intelligence becoming very much a commodity, and we're going to see this commodity intelligence approaching the price of electricity. When that starts to happen, Australia starts to become an incredibly relevant provider of technology and of service and of our unique utility in light of our geographic advantages and geographic benefits. I think in the long term we're—and I'm sure that they have done that technoeconomic modelling as well, because I'm sure that they would be aware of that too. I think also we have some capacity to negotiate because we provide some stability and proximity to markets that are of critical interest to hyperscalers. But you are correct in saying we need to get the calibration right, because, if the economics won't work, it simply will not work. So this proposal is not for any one mechanism to be terribly onerous but for the mechanisms collectively to account for the success that any one individual model would have within its ideal scenario. Within each of these states of the world—slow burn, open frontier and augmented growth—there is going to be one flavour of calibration and a model that is best, but unfortunately it tends to not be as successful under other modelling scenarios and outcomes. So we just need to calibrate the outcomes such that we get a comparable return within each of these situations. It might be that the cost associated with registering to train is relatively modest, but it allows us to understand who is training and to potentially support them and curate them to the extent that their Australian players get behind them as well. So we're not in any way suggesting that we should deviate from a commercially viable economic model. In fact, I think extreme token taxes will cause people to shift to alternative models. That will just result in model shifting. We've seen a lot of proposals like that. We think that the calibrations have to work for the hyperscalers as well. That's why we're really open to this idea of co-investment and carrying some of the investment risk, in particular, looking to a more balanced investment pool that incorporates the inference as well as compute. Ms CHANEY: Thank you. CHAIR: Ms Mascarenhas. Ms MASCARENHAS: This will be a question on notice. I see lots of parallels between the mining industry and data centres and AI, but then there are some fundamental differences as well. For the resource sector, those resources are owned by all Australians or people within that state. In the mining industry, it employs people locally at a significant level. The thing that I'm trying to get my head around with the way that we share benefits is how communities that have data centres within their communities get extra benefits because they have these data centres in their own backyard. The other thing I'm curious to get my head around is that we see a boom-bust cycle and a two-speed economy, particularly in our resource-rich states. How do we prevent that from happening? We are seeing this massive boom in data centre constructions. That's on notice. Mr Lewis : We'll take that on notice and come back to you. We'll come back to Senator Sheldon as well. Ms MASCARENHAS: Brilliant. CHAIR: Thank you so very much. We'll bring this session to a close now. Again, thank you, Dr D'Rosario, Mr Lewis and Dr Chau. For the information that you indicated you'll be able to provide on notice for the committee, please do so through the secretariat in the next two weeks..