Ep 245: Long Listen with David Keane on Sovereign AI & Project Magpie

>> The primitive forms of artificial intelligence we already have has proved very useful. >> AI is a golden opportunity. >> How do we ensure that this technology is regulated? >> The long arc of technological history is now in an extraordinary new phase. >> The AI6 podcast with Joanne Sweeney. >> David Keen, hello, first of all, you're in Australia, I'm in the west of Ireland, but we met earlier this year in Cambridge at Harvard Business School. Very good to see you, my friend. >> Good to see you too, three interesting places where there are a lot of Irish people. >> Oh, there absolutely are, and I think I told you in Boston, I went to Australia and within a couple of hours, I really recognized why the Irish love Australia so much. But it's great to reconnect and to see you albeit online, and I've been listening to a podcast that you were on recently and it's fascinating. But first of all, let's introduce you to the viewers and to the listeners of the podcast. >> Wonderful. Thanks so much for having me here, Joanne. My name is David Keen. I am the CEO and founder of a business called Southern Cross AI or SCX.AI. What SCX is doing is we're establishing the sovereign AI infrastructure for Australia. What do I mean by that? Well, what I mean is we're building the capacity to be able to enable all Australians, whether they are businesses, whether they're part of the government or even individuals, to be able to access the power of artificial intelligence. But to do that in a way that gives them the confidence of where that's being made, how their information is being used, and also a bit of an opportunity to use AI that has been fine-tuned for Australia. I'm sure we'll talk about that, Joanne, and it's our project Magpie. But we really believe that for AI to be successful for all Australians, it needs to be ubiquitously available. It needs to be cost-effective. It needs to be made right here. So yeah, that's what we're doing and really excited to talk more about it. >> Well, listen, I'm a European, you're an Australian and we know that the owners and the makers of AI and who's powering the world's AI are largely US or China-based. And so when I heard you drop sovereign AI into our WhatsApp group, I was right on it because the conversations that we had at Harvard were very much around that, like who's going to own our data and how can countries take control and have autonomy over AI. So how long have you been working on this and how far into the process are you now? >> Well, I've been working with AI technology now for a number of years, right back to even before the pandemic, when we saw that its ability to make non-deterministic decisions could really change the way humans work and the way they communicate. But yes, you're right. If you think about it, the power of that technology to be able to do things in ways that were simply impossible before. And if you call that intelligence, if you like, we can call that for this podcast, Joanne, but if you think about it like that, the idea that countries like Australia and other countries around the world will simply outsource the creation of intelligence to the USA and China is maybe not acceptable to many citizens of those countries. And no matter how much we enjoyed our time at Harvard and how much we love Americans, you can see I have a nice little badge here on my lapel about the partnership we have, the idea that everybody wants some super woke Californian AI or an AI built in China, maybe with some political controls. You know, those are big things. And so I think that Australians are speaking out loudly and they're saying they need the power of AI, but they want the confidence that their information is being used correctly. So yeah, we think that this is an essential part of the infrastructure for our country. And I believe all around the world. And I think when I think about sovereign AI and even that that word definitely has an emotional connotation also, I think about culture and context because as countries, as nations, as cities and, you know, think about the culture of Australia, if I think about the culture of Ireland, that's all nuanced. And we all operate in places that mean something to us. And so having technology that's built around culture and context differentiates you, I think it definitely gives you a place in the market that makes people stand up and listen. But are you going to stop at Australia? Aha, well, I'll come to that second, but I want to say how spot on you are. And, you know, you said something very interesting. This is about not just, you know, how do we spell words? It's about culture. It's about the way we are as people and even in countries as diverse as Australia and I'm sure Ireland and all throughout the world in diverse worlds, each country can have things that make it unique. And we believe that if we can create the infrastructure to create that locally, that includes the machines that make the AI and it also includes the models themselves. And those folks who visit scx.ai will be able to learn about our project Magpie. Magpie is a Australian bird. And that's about fine tuning one of these large language models for, as you said, culture, information, legal, legal requirements, all those kinds of things. Now, where do we go? Well, we believe sovereign AI is needed in every country. And we're convinced that that is going to happen. I think it's inevitable that the train has already left that station. And we expect to see that around the world. But part of the message for your amazing viewers would be if there are folks out there viewing this who believe that their country would benefit from a sovereign AI solution, get in touch with scx.ai. We'd love to bring what we've done to other countries around the world. Well, you know, I work a lot with government and public sector. And there's a huge fear there around data retention, intelligence, and where our data is stored even this morning. And so this is a pre-recorded interview. So by the time it goes out, it'll be a couple of days. But this morning, the breaking news in Ireland is that the European Union is rowing back somewhat on the harshness of our GDPR rules and our regulations around AI, there's a recognition that they're slower to market and they're kind of clipping the wings of innovation. So I think the mood is shifting on this side of the world. But talk a little bit about Project Magfi and how you've used open source, but then refining, as you say. Well, that's one of the interesting things is that the market is moving so quickly. And one of the areas about AI that I think is really helpful and pleasing is that the open source community has really embraced AI. And there are a number of open models now available where organizations like SCX can take those models and we can build on top of them. And that's what we've done with Project Magfi. We've actually created Australia's first fine-tuned large language model that understands Australian culture, Australian laws, Australian practices and able to help all people that use it to get answers to their questions in the ways that kind of feel right. And again, you said it, it's that feeling. It's not just, you know, how do we pronounce a word or how do we spell something? It's more than that. And Magfi is really about bringing that to all Australians and doing that, though, in a way that uses those open source foundations so that we can deliver not just an Australian aligned model, but we can do it in a way that really drives the performance of that model. We can build something that's good as well as being Australian. And I think that's what people are expecting now. They want the very best they can get from intelligence, from things like these large language models, and we expect to deliver that at scale for our users. Now, we have to talk about the infrastructure that powers AI, because without that, you couldn't have Project Magfi. So you have partnered with a company and let's talk about the chips, right? We're all talking about the chip value of NVIDIA and tokenomics, I heard you say on the podcast. So talk to us a little bit about what's happening behind the scenes. Well, I'm sure many of the viewers of your podcast will have seen these stories in the media about organisations building these massive data centres and the cost of those data centres. And, you know, NVIDIA, this year passed five trillion US dollars in valuation, five trillion, that's a tremendous amount of money. And, you know, that's because they've been selling a lot of chips into those massive new big data centres. But many people are getting very concerned about the power use that those data centres take. And I can tell you the forecast here in Australia is that data centres currently take about 2% of our power load. And the forecast, if nothing changes, is that by 2020, that goes to 12%. That is a significant issue for a country that has made a net zero commitment. If we don't fix that, we're not going to make a net zero commitment. But it's the same with water. Many of these modern data centres use water. And I've heard as crazy numbers as half a litre of water, chat GPT completion, your water is being used. So these are significant environmental concerns, both in power use and in water. So we chose to implement SCXAI's infrastructure using a next generation ASIC-based chip. So that's an application-specific integrated circuit. We're not using GPUs. So I'll say that again, we're not using GPUs for this deployment. And the reason is we're able to get at least one order of magnitude more efficient AI outputs. And it simply means we can create more intelligence per watt. If you think of the energy watt we're putting into this, more intelligence per watt. And what that means is that we can deploy this without impacting Australia's net zero commitment, without taking away water from citizens who might want to use that water. And particularly in a country that's subject to drought as Australia is. These are real environmental concerns. So by building, maybe a way to think about it is building an AI factory for the future, not building it with the past approach, build it with the future looking approach. And by using these next generation chips, the vendor we're working with is called Samba Nova, also a US company making amazing chips. But yeah, this transition towards ASICs that deliver high performance and the right level of energy efficiency that we need to make sure that this can be a viable deployment is what SCX is all about. So if other nations are listening to this and are a bit curious about how to develop and rule out sovereign AI in their own country, how long is the process? Well, we've been working on that throughout most of 2025. That's been our focus on building it out. And there's a lot of work to do. So first of all, we're incredibly fortunate to have an amazing team at SCX. The people here are incredible. They have executed in time frames I just didn't think were possible. I really didn't think they were possible, but the team has built the entire infrastructure to the level by which we will have this year 2025, Australia's largest sovereign inferencing cloud, right, the largest one built very rapidly with the highest performance and highest reliability. So look, I think it's all about the right technology, the right people working together to build something like this. But if you approach it with a modern view with this idea that you can create something using the next generation of technology, you can create things quickly. So I think, again, as you talked about, look, there are people out there watching this who are thinking, this is that sovereign AI is important for their country. We'd love to share some of the ideas. But interesting, Joanne, is that someone told me this in a meeting. They said, hey, hang on a second. Why is sovereign AI just for a country? Couldn't sovereign AI be for a city? Couldn't sovereign AI be for a government department or organization? Sovereign AI be for a company? And I think they're probably right. As we start to think about this, why wouldn't a company or an organization want an AI that represents its way of being, its way of doing things? And some of the technology we've built at SCX is ideally suited to help those organizations. And we might see this breakdown into these smaller sovereign entities that need access to this in a way they can control exactly what happens to their information. Now, you're on to something because we know the future of AI will be domain specialisms where we have maybe the legal profession or finance any professions. People who are subject matter experts in their own field who are building up their own local model. So partnering and having your own sovereign AI makes perfect sense, right? Yeah, indeed. And we were talking about some of the countries in our region here in Southeast Asia where there are organizations that would love to have that capacity. You can imagine a sovereign model for doctors in Bali that have to deal with tourists more than other doctors would have to deal with. So it needs particular fine tuning or a particular model for lawyers that operate in Singapore where there is a significant cross-border trading process involved. You know, these are important and there will even be models that are built with cultural nuances that meet the way organizations and groups would like to interact with their AI. So you look, this is a natural, I kind of see it as a natural development. You start off with all these new technologies and some big wide slath of deployment. Then you get down to how can we make it really meet the specific use cases that people have? And I think we're starting to see that trend now and probably that will accelerate in 2026. And how do you think the end user, the Joe public or the small business will change their behavior from potentially being an everyday chat GPT user to being a magpie first user, right? Well, look, we want to give them an option. I think that's all about optionality for people. You know, we don't... I suppose one of the ways to think about it is that we would expect humans to use multiple AI tools. It's not a one tool fits all world. It's not that kind of world. We would see multiple tools being used in different purposes. And if you think about even how people use social media today, they're still using Instagram where it's a very visually centric tool and might be using an ex post where it's very text centric and then maybe a TikTok for a video approach. And I think something very similar happened with AI. You'll get different use of the tools based on where you want to see the output go. And so we would see many small businesses using magpie to be able to get a very Australian specific view. Maybe an answer to a question about some process they're engaging in to attract customers or to build their business structure. But then they may still turn to a chat GPT or a Google Gemini to create images. I mean, everyone's seen the famous nano banana. It looks pretty good. So I think you're going to see a combination of those things. But we see the job of SCX is is build the infrastructure so that when people want to be confident about how their information is being used, you know, who's doing what with what and get that level of confidence that they're being taken care of and their intellectual property and their information is being respected. There was one story, Joanne, I was going to share with your listeners and viewers is it was actually in the national media here in Australia a few weeks ago. It was a medical clinic and they announced that they were moving from human receptionists that would take calls and book sessions with the doctors to an AI based voice agent that would take the call and organize the sessions. And they were talking about very publicly how this was going to improve the service for their patients. But of course, two things also go along with that, even if that is true. The first thing is that for a reception staff lost their job, they were replaced by AI. That's a question for our society. And maybe we can have a separate conversation about that. That's a big question for our society. But the second thing was patients were calling up, talking to that voice agent, giving their personal medical information without knowing what was happening to that information. Where was that voice agent taking that data? Where was it processing it and how was it using it? So look, these are real world concerns that we need to understand. I think your point earlier that the EU is recognizing is we can't slow it down, but we still have to be responsible. And it's that combination that will give the best outcome for people. Well, magpie be multimodal and will you step into the arena of agentic AI? Well, we'll only be multimodal if we can have images recognized of the wallabies beating the Irish in rugby. That's about the only reason. No, of course, we plan to make everything multimodal. Our team, our science team is actually working on something really interesting, which I can forecast first on this podcast, which is we are working on an Australian voice model. And that is to help those people who do want to do their voice agent, but do it properly with security and governance and guaranteed sovereignty of the information. But they can have an AI that speaks like an Australian. I don't just mean the sort of monotone, ochre Australian. I mean the real diverse Australian that lives in this country. And I think that is what makes us so strong and powerful. And when we can build unique models like that, we're going from text chat completion and image recognition to a world whereby we can treat these AIs as true agents, as true assistants that can help us do our work. And what about the the agentic mode? Will you go there eventually? Well, I think our customers are already doing that. So as well as providing access to things like Project Magpie through our chat bot and folks can visit chat.scx.ai and try it today. We're also offering the API access. And we're already talking some of Australia's largest businesses that are looking to use the API endpoint to build agentic applications. And I'm seeing amazing innovation here in Australia as entrepreneurs as well as larger companies are investing the time in building agentic workflows that can perform tasks. And that's where they're going to get the real benefits in their operation or depends on how they achieve the productivity of their teams is making these agentic applications that work. So our infrastructure is ideally suited for powering those agents. Now, there's a lot of people building agent SDKs and agent toolkits, but all of those things need the tokens to power them and SCX plugs into those. And that's wonderful because I mean, you can build with whatever agentic toolkits you like, you can build a very comprehensive system, but then we can give you the sovereign high speed efficient tokens to power those workflows. And I'm looking forward to seeing more of those applications next year. How has SCX and Magpie been received in Australia? What are the media people saying about you? Well, we're already seeing a lot of folks really interested in this. And there's so many people certainly in regulated industries. I mean, there's a lot of regulated industries that would just tell back from using it at all. They literally couldn't use it. And this is going to blow your mind, but before SCX in Australia, you could not buy AI APIs in Australian dollars. You had to buy it in US dollars. You could not even buy it in a local currency. And the big hyperscalers would not guarantee it was being performed in Australia. So literally, there was nothing. I just couldn't believe it myself, but it's obvious. And you're going to see a whole bunch of people come into this market and offer something very similar. I think it's going to be a global trend. It's going to happen all over the world. But we've been engaged with universities, looking at using this for their student populations. We're engaged with financial services companies who are trying to build those agentic applications to improve the productivity of their teams. You know, we're working with health care organizations that are really struggling because of the regulation around health care, care and patient records. You know, they need AI, but they need to make sure that they've got complete governance and control of what information is processed. Where, you know, we've got small business organizations who want to go to those kind of voice agents we spoke about before, almost as the virtual receptionist. So, you know, maybe it's a trade type application and they want to be able to have a receptionist answer the call, but not just answer it, you know, provide information about the service, even organize an on-site visit, give an indication of pricing, all delivered by the AI agent. So we're seeing those use cases now in a way that really wasn't happening here. I think Australia was a bit behind, I would tell you. I think it was, and there was the fear and the concern over the use of AI, I think was holding back innovation. So now we're seeing that, you know, reduce, and I would expect that 2026 is going to be a big year for SCX and really for the whole industry. That was just going to be my next point because I've just finished an episode on McKinsey's The State of AI Report 2025 and trust being one of the biggest barriers to progression and acceleration. So I think when you can answer astutely and give confidence around the trust question and who owns our data and do we retain our IP, I think that we will see organizations move because we know they need to move. We all need to move. And for somebody who's very keen interest in government and public sector, I want to see my institutions move, you know, for the betterment of citizens. So, yeah, that trust factor is huge. So maybe you've just again, differentiated yourself in the market with that. We really believe that's true. And it's one of those things that's necessary for countries. You know, we talked about this at the beginning of this session was you can't outsource your intelligence to just the Americans and the Chinese. You just can't do it. Now, you need to work with those models and those people because they've got some amazing technology and you need to make sure that you have access to that and you can't legislate your way out of that. That is a reality. But at the same time, you need access to AI that can be delivered with the level of control and governance that organizations need. And you're sure it's the sector organizations, but it's all and it's regulated industries. But as we've been talking about the last few minutes, it's the general general population. They want to know what's happening to their information. They want to feel confident that's being used in the right way. And from time to time, they'll use a bunch of these tools. But if they've grounded access to sovereign high performance and also energy efficient, and we could talk about that for a long time, we in our model, in our financial model, what we found was that if we didn't do something like this, it was just impossible to make money. The level of investments needed to drive the old way of doing it was so substantial that it would just be very difficult to make money. But this new approach using ASIC based hardware that is efficient, you know, in order of magnitude, at least maybe to more efficient in terms of building the AI capacity, changes the economic model. And that means we can bring this to everybody in a way that is going to empower them to build those next generation of genetic gaps, to embrace those technologies like a genetic flows and do that, but be able to afford to do it. Because we know that the pin is just sitting above this AI bubble. The hundreds of billions that are being invested by the few does not make it accessible to the many. So again, the environmental, the economic model, all these big questions are really important when we think about scalability to all nations. Well said. And for many humans who care about the impact on climate and care about the carbon offsets and the challenges they have, and certainly in Australia, organisations here of scale are required to report, particularly public companies, report to shareholders and markets about their impact on the environment. And we're able to work with those organisations and say, look, you can still do AI now, good news. But when you report, you'll be able to talk about the fact that you're not letting AI change your general outcome in terms of your carbon offsets and other kind of programs you're participating in. You can do it and feel confident you're getting the most efficient way of building AI that's possible on the planet. I think that's what will make a difference to some of those organisations that have those concerns, similar for some of the legislative and legal concerns about the way models answer questions. And we've baked those into things like Project Magpie and the whole SCX story. So, yeah, we think it's important, it's the timing. And we think that as we see people start to embrace these technologies more, we need to provide them access to this when they need it. So AI is mainstream adoption is accelerating. What predictions do you have for 2026, given that we're coming to the end of 2025? What a year, and in many ways, it was a very interesting year. If we go back to January, if you can remember, seems like a long time ago, was the deep seek moment for that. Back in just January, we talked about the Chinese models coming out being as powerful as the commercial American models and potentially very cost-effective to train. And that had an impact itself on not just Nvidia, but the whole industry. But then during the year, we realized that even with these new models that were being made available, open source, and even with new ways of training, the usage demand was so incredible. There was so much growth going on that we still needed to improve the investments and really get serious about building the infrastructure we need to create these AI use cases that will make a difference in business. And we get to the end of the year, and we've had two other moments happen. The first one was called the Kimi K2 moments. Some of your listeners may know what Kimi K2 thinking is. It's another Chinese model, fully open source, that really hit the leaderboards of the AI models just a few weeks ago. And it's claim to fame. It's able to process hundreds of what are called tool calls or those agentic steps that you build in those applications. Hundreds of tool calls without any human intervention to come to an answer. So imagine it's searching for this, it's looking for that, it's updating this, it's checking that database hundreds of times, and then to come up with an answer. That is something that was unimaginable in January 2025. So what are we going to see next year? Well, for sure, we're going to see more of that, more advancements. You know, we just had the recent Gemini 3 announcement come out. So we're going to see more powerful LLMs. But my view for next year is the two biggest trends are going to be, number one, those voice agents. By this time next year, we do the podcast next year, we can come and measure this. We're going to see voice agents become very common in the market. People will have personal voice agents, companies will have business voice agents. It's going to be something that they will use every day. It's going to be very, very, very standard. The second thing in 2026 is video, you know, the creation, not just of images, but the creation videos. People by this time next year will be creating real short form personal videos. If they want to tell, you know, Lord of the Rings meets the towering inferno, they can create that for themselves. And that's going to change the way we use AI completely. I think that's going to also change the way humans engage with each other and the way they get entertained. And there are two predictions for next year. But underlying all that, Joanne, is we need this sovereignty, governance and confidence in this AI. Because if we don't get that, then we risk, you know, you talked about the pin. If the bubble bursts and we don't have confidence in what these tools do, we may set ourselves back as a species for another generation. When we have an opportunity in front of us to really embrace this and get the outputs that true AI can deliver, better healthcare, better environmental care, you know, better economy, I think it can deliver that if we understand how to use it. David, it's been fascinating. Thank you so much for hanging on and extending your working day to have this conversation with me. I think we are going to have to catch up in maybe less than 12 months and just see how the landscape is evolving. But I wish you lots of success and with sovereign AI and SCX. And I will be keeping an eye on it from this side of the world. So thank you for joining me on ASX. And where can people find you, but also remind them of Magpie and where they can go? Indeed, they can visit scx.ai to understand more about what we're doing as a business. And they can find me on LinkedIn. I'd love to connect with anybody of your listeners. We'd love to work with folks in all countries around the world and share our experience in building sovereign AI for Australia. Well, let me try and get you to Europe in 2026 and get you on a stage or two. That would be great. Thank you so much. Good to see you.