Systems Rewritten: AI

Transcript

Systems Rewritten: AI

(00:03):
There was one day in July where semiconductor ETFs took in 7.1 billion, which is a staggering amount of money.

(00:08):
The International Energy Agency predicts that 20% of the current data centres that are planned globally, we will not be able to build them because of constraints in the power.

(00:21):
The largest data centre in Malaysia makes up about one third of Europe's entire compute capability.

(00:29):
And we've seen parallels in the dot-com era, and we had spare capacity for a few years. Some companies went bust, but at the same time, this then laid the foundation for the digital economy and internet economy that we have all experienced over the last 10 years.

(00:45):
Data centres are contributing about 0.8% of US GDP growth only in the first quarter of 2026. In 2025, we knew that if you stripped out the data centre investments, the GDP growth of the US, according to a Harvard economist, would be at about just 0.1%.

(01:07):
And the long semiconductors trade, 82% of them win it. That's the most long a trade in their surveys ever. In August, that's dropped to 53%, so I think the energy might be coming out of the trade.

(01:22):
Supply and demand, the foundational backbone of market economies. And we're in the AI era where demand and supply is being rewritten. Behind this AI revolution sits a monumental physical build out and energy demand. Globally, data centre power consumption will double by 2030 toward 945 terawatt hours, consuming more power than the entire nation of Japan. In some regions, the strain is already visible. Data centres accounted for more than a fifth of island's national electricity consumption in 2022 and predicted to reach nearly a third this year. Stateside in 2023 when less powerful models existed, roughly a quarter of electricity use in the US state of Virginia was sucked up by data centres. And yet, McKinsey estimates that there will still be $7 trillion in global data centre infrastructure investment by 2030. Then come the hyperscalers, Alphabet, Amazon, Meta, Microsoft, and Oracle, who are expected to spend around $670 billion on AI related capital expenditure this year alone and nearly $800 billion next year.

(02:45):
The nine largest AI companies have collectively forecast $4.1 trillion in CapEx through to 2030, a trillion dollars more than the total 2025 CapEx of all US non-financial companies combined. So looking at systems being rewritten, are we witnessing the greatest industrial retooling in human history or is it all too much to handle? Well, welcome to the Next Five Podcast. I'm Tom Parker, and this is episode two of our three-part series, Systems Rewritten. Today, we explore the infrastructure race powering the AI economy, the semiconductors, data centres, and power grids that make the digital economy possible at all. Joining me to navigate this are three leading experts. First up, we have Matthew Wright, chief strategy officer at TradeNation.

(03:46):
Thank you for having me, Tom.

(03:47):
Next up is Walter Pascuarelli, researcher at Cambridge University and advisor to the OECD.

(03:53):
Thank you for having me, Tom.

(03:55):
And finally, Bernard Maher, author and futurist.

(03:59):
So great to be here.

(04:00):
Well, Walter, let's start with you. In our first episode, we looked at the market concentration and multi-trillion dollar valuations of AI companies. The AI revolution is a story not just of code, but one more akin to industrial revolutions of the past, one of steel, concrete, and energy consumption. But the numbers behind this industrial build out are staggering. Let's dive into data centres. Can you give us some of the numbers behind this build out? Do we have enough data centres to keep up with AI's demand, and what about the power that's needed for them?

(04:38):
Thank you, Tom. I think that's really the crux of the issue, the enormous amounts of capital there currently going into this build out of data centres and computing power. Now, it's very easy to get almost a bit tangled up in these enormous sums, trillions and billions that are being invested left and right, but maybe to make it a bit more accessible to ordinary listeners, from what we know nowadays, data centres are contributing about 0.8% of US GDP growth only in the first quarter of 2026. In 2025, we knew that if you stripped out the data centre investments, the GDP growth of the US, according to a Harvard economist, would be at about just 0.1%. So they make up a majority of the US GDP growth nowadays. And that tells us as well that investors and the AI firms, they're sensing something. They're definitely sensing an opportunity for a major macroeconomic shift of the US and potentially the entire global economy.

(05:40):
The first thing to say is that data centres are not necessarily new. We've had some exposure to them already when we use the internet. We've had some exposure to them when we used to access our favourite website. But the whole thing is that with artificial intelligence, something fundamental has changed. And the thing that has changed is the density and the way that these data centres are built. AI consumes significantly more electricity when on the one hand being both trained, when we create the models, the GPT-6, the GPT-7, the next big model, but on the other hand also when we want to ask it queries, when me as a user, I ask it, "Hey, where's the best coffee shop near my place?" And it wants to scale that up and you think about it of millions of users and companies running large scale operations, that number becomes a lot larger.

(06:29):
With these electricity requirements, we see that nowadays because these data centres need to be built on the one hand next to electricity grids, but also next to local populations, the demands have become much larger and that has also sparked local opposition. Most of these oppositions, rightfully, they ask themselves, "Well, what are we talking about when we say this compute gap? What is actually the thing that we want to establish? After all, when I ask GPT or Claude a question, it's already working." The question is kind of intuitive. It's also somewhat misguided. Fundamentally, the build out of the data centres is a bet that artificial intelligence will be integrated across every layer of the economy. Think of it for knowledge workers or running more complex financial operations, augmenting humans if we want, but there's also inevitably going to be conversation about the displacement of workers. And then fundamentally also the second area, which is still remaining more under explored, is the topic of robotics.

(07:31):
Here we have self-driving vehicles, here we have drones that deliver parcels. All of this requires massive amounts of computing power backing up this emerging economy.

(07:45):
Yeah, absolutely. And Bernan Walter said there that this is a bet on AI being a part of every sector of the economy. And I want to look at the power element of this discussion even more because power grids are becoming the ultimate ceiling on AI development. We've got grid infrastructures that I interviewed a guest a year ago from HSBC who said that Thomas Edison would still recognise 95% of the grid today. It's not changed that much. Can utility providers scale fast enough or does energy and the data centre availability become the defining competitive bottleneck over the next few years?

(08:23):
Yeah, I absolutely believe that the energy grids will be one of the biggest bottlenecks over the next few years. What we are seeing is that we're building this enormous data centres that needs so much energy. We're now building data centres that need a gigawatt of energy, and that is the equivalent of the same that a million households would use. So Meta is currently building a one gigawatt data centre, XAI is 1.2 gigawatts data centre, Anthropic two gigawatts, and OpenAI has already, I think, scaled up to 10 gigawatts. So if we put this into context, it's very easy to build a data centre. It's very hard to build the power grid to upscale the system and all its components. So a data centre can be built in one, two, or three years. The power grid will be upgraded in five to 10, 15 years. So that is a tricky position to be in for most countries today.

(09:26):
And actually the International Energy Agency predicts that 20% of the current data centres that are planned globally, we will not be able to build them because of constraints in the power. What is interesting is that we look at, you said Edison would still recognise their current power grid. That's true. But what I now see is I work with many of the hyperscalers and what they are doing is they're bringing their own energy. This is quite exciting, especially renewable energy. So when they build data centres, the plan is to say, okay, let's build solar power stations, wind, geothermal next to the power stations. And particularly exciting is potentially nuclear reactors. We now have the capacity and the ability to build small nuclear reactors are very, very different to what we've had in the past. So companies like Google and Amazon are now building these nuclear reactors alongside their data centres.

(10:24):
And I think over time there are other things we could do. We could look at fusion technology, for example. What is interesting is people always say fusion technology is always five years ahead, but we've made some huge progress over the last few years and we've now had the first fusion reactor that actually produces more energy than it is needed to produce that energy. And Google has actually signed up to buy that fusion energy once it's available for its data centres. And then we have other cool things like solar power in space because space has solar power 24 hours a day and we can now capture this and we are currently building huge solar farms in space. And the amazing thing is that we can get this energy to earth via microwaves. So we can build huge antennas next to data centres and actually capture the energy coming from space.

(11:20):
And I guess in the future we could even build data centres straight in space because we don't need them down here and we can compute even up there.

(11:27):
Well, yeah, Matthew, I mean Bernard's kind of given this futuristic look at potential solar panels in space or even building your data centres in space, but then at the moment we still need to build nuclear power stations in order to create that one gigawatt that's needed for an open AI data centre. And a nuclear power station is small or medium sized comes about one gigawatt. Matthew, let's bring in the market and trading perspective. You see active investors pricing in these macro shifts in real time. For a long time, the market focused almost exclusively on the software and model creators. Are you seeing traders shift their risk exposure towards the infrastructure plays and where is active capital positioning itself right now?

(12:11):
The trade that's on at the moment, I'm just going to produce some numbers here, how much capital is actually flowing into this trade. So US semiconductor ETFs took about $46 billion in this year, which is an incredible amount of money. There was one day in July where semiconductor ETFs took in 7.1 billion, which is a staggering amount of money. When you look at ETFs and tracker funds, it gives you a really good idea of where this money is flowing. And when you look at the returns on the S&P 500 over the last year, it's been about 17%, right? And within that you've got the magnificent seven, the hyperscalers you work with Bernard, they're only up about 13% over the year. And where the money has gone is it's gone down the layers. So you've got silicon and you look at silicon ETFs, people like NVIDIA, these sorts of companies, that's up 78%.

(12:54):
Then you've got networking and optical. And when you look at those ETFs, about 36%, they've gone up. So you carry on going down the line, almost like a waterfall effect. You got the real estate investment trust up 41%, then you've got the kit that goes in them. So these are things like getting from the power supply to the server rack, companies like Veritiv up 70%. Power Industries, which does clever switching gear up about 70%. Then you carry on down, you get the electrification of it, that's up 26%. The only place it falls over is when you eventually hit the power companies and they're down 10% on a year. And the reason they're down 10% on a year is because in the first year, I said it was a two-year trade, in the first year they're up 50%. So that's really where the money has flowed.

(13:33):
It's been a staggering trade. Now, all I'd say on that is traders look forward, not backwards, and that's what's happened over the last couple of years, but I think the energy in the trade is fizzling out, I believe. Like I said, traders don't look in the rear view mirror, appetite is turning. And there was a Bank of America survey that surveyed a lot of fund managers around the world and the long semiconductors trade, 82% of them win it. That's the most long a trade in their surveys ever. In August, that's dropped to 53%. So I think the energy might be coming out of the trade.

(14:01):
Well, look, not too far from the front lines or power grids and data centres sits a geopolitical and regulatory battleground. Walter, you advise governments on improving their AI infrastructure. There seems to be an economic urgency to just build, build, build and keep up with AI. What are governments around the world thinking strategically? As I said at the top of the show, data centres consume nearly a third of island's electricity. Are their strategies driven by industrial FOMO, if we will, and where does this sit when it comes to say other priorities, net zero objectives, for example?

(14:39):
Yeah, that's an excellent question, Tom. So there's a famous scene in Ellis in Wonderland where Alice speaks to the Red Queen and says, "To get to places, I need to run fast." And the Red Queen answers to her and says, "If you run fast, you only stay still. If you want to get ahead, you need to run twice as fast." And that whole scene has been taken up by some other commentators that it has been turned into something called the Red Queen theory where acceleration means progress. But fundamentally, it all relates to what has more, on a more popular level, been known as the AI race. Now, I'm personally not a huge fan of that because an AI race implies that there's a finishing line and the finishing line, as has been told by so many experts and leaders of AI firms would be AGI, but AGI is kind of more like a theoretical concept that we haven't really agreed any definition to that.

(15:29):
But more importantly, the global supply chain of AI development is a lot more diversified. So if we put ourselves in the mind of, let's say, Trump and the Chinese president, there, the logic is still really one of either domination or economic improvements and trying to essentially become one of the leading global superpowers. We see that in the US being really one of the largest investors in data centres and infrastructure, but also really the provider of some of these models and China wants to catch up with that. But critically, both countries have made it part of their national strategies to really create this AI narrative as where they see their future. And I think as we shown with the numbers on GDP growth in US, I think that's clearly evidenced by that as well. Europe has what I call a bit of AI identity crisis. On the one hand, it has all the capabilities.

(16:27):
If you look at the cheering award winner, so Joshua Banjo, Jeffrey Hinton and Jan Lakoon, they all went to European universities. They all had European education. Now, part of the problems are a little bit more structural in the sense that you have 27 member states, you have 27 jurisdictions, but there is also really an issue of mindset and investment mindset and a bit of the risk adversity that we've had through decades of comfort and of interdependence with other countries. If you look at, for instance, the topic of compute infrastructure and data centres, the largest data centre in Malaysia makes up about one third of Europe's entire compute capability. And that tells you quite a bit about the investment that Europe is making for its future. And I'm not really saying that what we should be doing is engaging in some kind of reckless race dynamic, but there's probably a difference between racing and wanting to stay competitive in order to protect those things that you hold so dearly, such as for instance, competitiveness, economic advantages, and that leads to social safety nets that might become more of a topic as we talk about job displacement and unemployment, particularly of entry level jobs.

(17:36):
There are some changes happening. Mario Draghi, who wrote the competitiveness report on Europe a few years ago now has established a Ryan Group in which he talks as he is known very openly and very straightforwardly about the issues. Another report was published just a few days ago in which academics, so again, not like startup founders or some of the builders of these technologies came out and said some real changes are needed, but some traction is happening. And I always say Europe needs a big cup of coffee to be able to wake up. There's some other regions in the world, for instance, the Gulf States, their logic hasn't really changed much. So they do have their own AI models, but really what they're trying to do is to control everything else, such as we see this in the build out and establishment of some of their compute facilities, and they have the advantage of having both capital and energy.

(18:29):
Most of the hyperscalers and most of the AI firms, they're happy with usually having a fifty fifty split of renewables and traditional oil sources. So for them, that's obviously a very convenient position to be in. When it comes to net zero, the electricity demands, particularly the energy demands, bring up kind of a paradox. So it's very easy to say, oh, look at all these AI firms that are now building data centres and are ruining all the net zero promises and pledges that they made. But fundamentally, if we look at the four biggest investors and buyers of power purchasing agreements, the first ones are Meta, then you have Amazon, then you have Google, and then you have Microsoft. That said, Microsoft had about a 25% increase in emissions. Google had something similar with their supply chains also because of the compute infrastructure.

(19:20):
If anything, that means that the corporate net zero pledges are now under strain. AI still makes about just 1% of global emissions, which is different compared to, for instance, air conditioning or electric vehicles, which we still consider to be the future of our net zero promises. And most of the tech firms would say that ultimately these investments that we're making will stimulate new innovations that can reduce some of these emissions. The International Energy Association says that that's a promise that's not necessarily a guarantee, but clearly the paradox here is kind of heightening as we build out these data centres. And ultimately the question that this always brings us back to is whether artificial intelligence can ultimately fulfil the promise that it's currently making.

(20:05):
Yeah, some fascinating points there. Bernard, I want to build out on a few more of this hyperscaler investments, whether it's into PPAs and the energy infrastructure. The capital expenditure numbers are absolutely mind-boggling. Hyperscalers are spending $670 billion this year alone. I mentioned it at the top of the show. And by 2030, the nine largest AI firms are expected to spend $4.1 trillion, and that's a trillion dollars more than the total CapEx of all US non-financial companies combined in 2025. Is this level of infrastructure build out sustainable at all? And I don't mean sustainable in the energy point of view, I just mean able to do it. Or is this a pipe dream and there will be excess physical capacity leaving stranded assets before the downstream commercial revenue arrives?

(20:53):
Really good question, and it's such an important question. I'm lucky that I see this from both perspectives. I work with many of the hyperscalers on the one-hand side, but I'm also working with lots of companies and I see the increasing demand for AI. And for me, we can have both at the same time. We can have a bubble where we invest too much and maybe not everything will become economically viable. At the same time, we will have an increasing demand for a capital investment to build these AI factories that fuel what we will need in the future. So even if you look at the capital investment by all the big AI companies today, I think they're saying it's just over $4 trillion they're planning to invest. McKinsey actually said we need to increase this by 2030. We need even more investment, maybe up to $8 trillion.

(21:47):
So this gives us an idea of what is happening, that the demand is going up so much. I'm lucky that I work with many of the big blue ship companies across the world, so I see what they're doing, and we are literally just at the very beginning of this AI revolution. They're playing around with generative AI. They're maybe having a few pilots around agentic AI. Every single company I work with has huge plans to increase their use of AI. And what we'll see is all of these increases will also increase the need for more data processing. Simply if we move from generative AI that can answer our emails and do some research for us to agentic AI that can perform tasks and entire workflows for us, that increases the demand. The other thing we're seeing is that more of the AI interactions and outputs will become multimodal.

(22:45):
So we're moving from text to voice generation, image generation, video generation. And again, this ups the need for computing power. And then a point you've already made, physical AI. So what we are seeing is we now have AI embedded in robots, in machines, all of this increases demand. So yes, we'll have both. We will have a much bigger demand, but I also see many investments that might not pay off. And we've seen parallels in the dot-com era when lots of investors invested in dot-com companies, and especially when we build out the fibre networks, for example. In the beginning, all the telecom companies saw we need to invest in fibre optic technology everywhere. And we had spare capacity for a few years. Some companies went bust, but at the same time, this then laid the foundation for the digital economy and internet economy that we have all experienced over the last 10 years.

(23:44):
So what I see is we need all of this capacity. Will every project be economically successful? No, but it will still underpin this revolution that is coming in the future.

(23:57):
Yeah, no, it's a fantastic point. I mean, you saw this happen. It happened in railroads. They spent the equivalent of, I guess, billions in today's money on building out the railroads, and lots of companies went bust and everyone still benefits from today. You got it with fibre, the dot-com boom, and now you've got YouTube that's running it. What happens with creative destruction is the companies that end up winning these races end up picking up the capacity on the cheap. And then suddenly when you pick up something on the cheap, you can give away free services and it creates whole business models that you never even though were existing. I mean, there are some companies now we never even knew would exist at the end of the '90s or couldn't even conceive them. So it's definitely going on. I'm pretty sure it'll happen here again. You're going to get these data centres at some point when you get a slide down, especially when you have really high fixed costs and then something happens to your revenue that drops a little bit, these companies go and what picks them up will be really interesting.

(24:47):
Don't be too worried about it.

(24:48):
Well, I do want to dig in a little bit more to some of the energy geopolitics. How are you seeing some of that across the data that you have here? Are investors looking at different instruments and sectors? What are they doing?

(24:59):
Yeah, that's fair. So I talked about the long semiconductors trade and how that waterfall down into the different areas. So what's really interesting around there, when you look at it, you see all these new ETFs being created and they're specialising in really particular parts of the market and there are companies there that these providers put in there. That means they're becoming really popular with more passive investors. So it's not the really active guys like your hedge funds or your actively managed funds. When you start issuing ETFs, that means you're trying to get it into the pension. I just think long and hard about who is selling what to whom in that scenario. And for me personally, this feels like a bit of a start of a distribution phase in the market. So where you've got perhaps a slightly more proactive money exiting, that's one element to take into account.

(25:42):
On the geopolitics side, the second point here is around credit, and this has made the Financial Times front page a lot in the last few weeks. And what you've seen here is concerns around inflation risk. These have been generated by events in the Gulf, the geopolitical events, and what we've seen here is peaks in the 30-year yield, and more recently, the 10-year went over 5%, which hasn't been over for a long period of time. Now, what that means is the cost of building out these projects, these really important data centre and the really long-term projects becomes incredibly more expensive. And just to give you a bit of an idea about how much more expensive it gets, on a 30-year power station, for example, every hundred basis points that you add to the cost of doing that, you have to increase the price of power charged by 12%.

(26:28):
That's incredible, isn't it? So the point here is that geopolitics doesn't dampen the demand for energy, but it certainly makes it more expensive. And I think that's what we've seen more recently.

(26:37):
Walter, let's look at the hardware supply chain because behind these data centres sits another choke point. We've touched on it a little bit across the whole episode. Taiwan's TSMC still produces over 90% of the world's most advanced logic chips, and ASML holds a monopoly on EUV, extreme ultraviolet lithography equipment. The US is desperately trying to catch up, but in specific areas like advanced fab capacity under the guise of national security. Can you map out the geopolitical story behind chip manufacturing and supply and what this means further up the pipeline?

(27:15):
So chips are essentially the bread and butter of data centres. Most of them are tiny. They have little switches that make essentially the calculations that allow data centres to actually work. Generally speaking, chips can be considered one of the most complex creations of humankind. And because of that complexity, we see that the supply chain for creating these chips from the moment of ideation until the moment that they actually landed those data centres is highly decentralised, highly monopolised, requiring different levels of expertise and specialisations across the world. So if we were to, for example, map out the production cycle of chips, you would start usually with the design. And with the design, you could compare it almost like to drawing up the map of a city. And the companies that draw the map of the cities need to produce the software that creates that map. That's usually a couple of US companies, Synopsys being one of them.

(28:15):
Then you would actually have the firm that makes the design or draws the whole map, which will be Nvidia, which has become one of the most valuable companies on the planet. Now, once you have the whole design of the chip, then you need to start etching with the EUV lithography into the silicons. That is only one company on the planet that dominates the market by a hundred percent. And that company is, as you just mentioned, ASML in the Netherlands, but even that company relies then on another firm, which is Zeiss in Germany for its mirrors. So it's almost like a choke point within the choke point. Once you have done the printing, then you actually need to fabricate them and package them. And that's what brings us to Taiwan, which as you just said, controls over 90% of all the most advanced AI chips. Now, as all of the listeners here know, Taiwan just sits in front of China and China makes a historical claim over Taiwan.

(29:12):
That has brought up all sorts of geopolitical intricacies and geopolitical competitions. In addition to that, there's all sorts of contractual agreements between those firms that make everything even more complex. But if we look at how the various actors and countries have responded to that, the US already for a few years put initially or tried initially to prevent China to get access to the most advanced AI chips. And notably, what they tried doing is that they wanted to incentivize TSMC to move some of its fabs into the US. The problem then is that TSMC or Taiwan more generally feels that if they move their fabs into the US, then they lose what is known as their silicon shield, which is essentially the main incentive for the United States to protect them from a potential Chinese invasion, which meant that TSMC now essentially has put in a rule that only after three years they would actually shift over some of those chips productions so the US would basically get not the most cutting edge ones produced at home.

(30:15):
China in the meantime, this year in particular, has been able to develop not an EUV lithograph, but something that is essentially a more primitive version if we want. So they have been able to some of their own inventions to some of their own domestic innovations to essentially produce a lower quality, poorer supply chain of chips. And realistically what this means is that we're likely going to see a chip supply chain that is not just one major global one, but as there is more and more geopolitical competition, what I advise my clients, whether it's either governments or investors, is that we can see almost the supply chain being split into on the one hand, one is that is more advanced, that is dominated primarily from the US and some of the Western forces. And then on the other hand, one where you have China at the middle of that.

(31:03):
Whether China can catch up and produce potentially entirely new form factors and entirely new innovations to potentially supersede some of these lithographs or some of these fabs, that's a different question. But the chip supply chain is where the AI industry has its choke point and the choke point has several other choke points as well.

(31:23):
Matthew, are we seeing more seasoned investors getting into all scopes of the AI build out? I mean, where are we heading with this chip and hardware volatility and what about the memory industry?

(31:35):
So we talked earlier a little bit about how the concentration has moved down the stack. And the first area we looked at was semiconductor providers like Nvidia. On our platform, it's one of the most traded equities. But what we've seen this year is some incredible moves from memory companies. So Sandisk up 1600%, incredible. Micron up nearly 500%. I mean, that tells you exactly what's going on. However, the question assumes that as an investor or a trader, you've got a choice about whether you're investing in AI. And my point here, I want to just challenge that, the Magnificent Seven, just the hyperscalers by themselves, make up 34% of the S&P 500. So if you are like you and me and have a pension, a sit perhaps where we choose our investments and we want to follow a tracker and you go and buy the S&P 500, you've got 34% of your money directly in those seven big companies.

(32:25):
And then you've got all the other companies that are not even in that number. The second point on that highlights the risk more is around half the market is now passive investors. So that's pension funds, ETFs, and mutual funds. The point here is you don't really have a choice whether you invest in AI. If you've got a pension, you're going to be exposed to it.

(32:43):
Bernard, if access to compute and energy becomes concentrated in just a few wealthy nations or corporate hyperscalers, what does that mean for our global economy? Will there be a much wider gap between the haves and the have nots? Will there be a new multi-tiered global economy where smaller nations are priced out of frontier intelligence infrastructure? As a futurist, you must have an interesting take on this.

(33:09):
Yes. And I don't think, not necessarily. It all depends on geopolitics and other dimensions because in the end, the value from AI is generated by the companies using it. So they don't need the data centres or a hyperscaler in their own country. So is businesses, the governments using this technology to generate value for their customers. Having said that, there is obviously a geopolitical dimension to all of this. So at the moment you touched on that, we have the US and China fighting for dominance here, and the US is probably globally very well positioned for this. They have most of the hyperscalers, they have the most successful AI companies, they have the land available, the energy available, and the politics that support building these data centres. So there's a risk that if I'm an end user business here in the UK and suddenly politics shift and the president said, okay, Anthropic can't make their most powerful AI available to anyone else in the world other than US companies, that is a risk.

(34:18):
So for countries and for businesses, I think it's really important to have that in the back of their mind. They don't need to have the data centres in their own country, but they want to make sure that they have secure access to this. And this is where we are having a big debate at the moment around sovereignty of AI, sovereignty of data centres. So governments are thinking about, okay, I better have a data centre in my own country so I don't become reliant on some of the superscalers or some of the two biggest countries that are fighting for dominance here. And this is tricky. So for me, AI on the one hand side has the potential to be the world's biggest equaliser because it gives everyone access to these superpowers. We now have the potential of having the first unicorn company just driven by one person or a few people making that happen through AI.

(35:17):
So the potential is huge. I think Anthropic recently said that they believe pretty much all of the world's diseases could be solved and we could find cures for this in the next 10 years using AI. So there's huge potential, but we need to make sure we have access to it. And at the moment, I don't see this going in the right direction. I think there's too much competition between China and the US, which is dangerous. And we are seeing already in politics, the US is restricting access to their top models, which makes them extremely competitive and introduces real risks to the rest of the world.

(36:00):
We're coming towards the end of the show. We're going to have a quick fire last two or three questions. The first one, what is the single biggest physical bottleneck, whether that's energy, silicon, land, or geopolitics that could stall the AI economy over the next five years? Walter, let's go to you.

(36:16):
Yeah. I mean, fundamentally the question that I would ask you is AI economy for whom and what do we mean by AI here? Many times when I hear people talking about the technology is almost considered to be almost like a global public good. But if we're talking about AI advancements, the two major bottleneck, in my opinion, on the one hand, and this is going to sound very 2018, is still the data. And that means data that is not just tax-based data that is taken from the web, but also about data in different formats such as for instance, from what we know nowadays as work models that try and capture what it's like to actually see and to have AI embedded more in context. And then the second thing is still the brain power or the talent that ultimately allows us to find new model architectures and new real breakthroughs that will be happening in the next few years.

(37:02):
Yeah, absolutely. Matthew.

(37:03):
So in the US it's electrons. So in 2024, America generated about 4,600 terawatt hours a year of energy. China 10,100. And the growth rate, China knows how to build this capacity. They've been building it out 8% a year, something like that. The constraint in China is silicon. And what we've seen in both these scenarios, I think they're engineering problems and they'll be medium term problems. They won't be problems in the long term, I don't think. We mentioned China becoming more efficient using its chips because it's got this bottleneck with silicon. I think the free market will work out the energy problem. The real issue here is geopolitics. We've mentioned how important these models are. And when they start being used as tools to restrict economic activity and punish countries you don't like, we've seen that.

(37:56):
That's not even a theory anymore that's actually happening and you're getting way more embedded AI in our economies. So I don't feel people in democracies will be prepared to carry the cost if they don't share in the benefits.

(38:08):
And Bernard?

(38:09):
I feel the biggest physical constraint is energy right now. This is something I think we can solve. Another one that is locally constraining is actually access to water. This is something sometimes forgotten. There are some countries that have abundant water resources, but we now get countries like Saudi Arabia trying to compete. They have lots of energy, they have finances available to build these data centres, but they haven't got any water. So I think energy by far the biggest constraint followed by water.

(38:40):
Well, I mentioned at the top of the show the McKinsey $7 trillion figure around data centre spend estimated by 2030. In five years time, will this be seen as the foundation of a modern economic golden age or as one of history's most expensive capital misallocations? Matthew, I'm going to come to you first.

(39:01):
We've actually answered this. Bernard's great point earlier on when we picked up about what happens when these big tech rollouts happen, normally the investors suffer a bit. We saw that in the dot-com boom, trillions wiped off, but then you've got some wonderful fibre laid. I think the key point here is around what they spend it on. So if you are investing it in things that will be there in 30 years, so infrastructure, power generation, I think it will be a boon for everybody. If you're investing it just in microchips, which are obvious in about three years time, it's going to be a struggle for it to be a big economic golden age. Okay, so my view is golden age for the economy, questions for the marginal equity investor, but those two have never been mutually exclusive.

(39:40):
Okay.

(39:40):
Bernard? I think it's both. I think it's really important to understand that the investment we are making is an investment in the AI era, an investment in the future of humanity. But if we believe that every investment will yield some return, that would not be the case. We will make lots of mistakes. There will be assets that would be lost, but overall, absolutely an investment in the future that is going to make the world a better place.

(40:09):
And finally, Walter.

(40:10):
Thank you. Yeah, fundamentally to echo Bernard's point, it's probably a mix of both. I have no doubt there is currently overspending going on. It could happen that there's some changes in the model architecture. It could be that the cost of AI will just simply drop and then we will struggle to, even if we fill up the vacancies in those data centres with sufficient users, that the return on investment that is being made is simply not high enough. At the same time, the cat's kind of out of the bag and innovation has happened. I mean, try telling people, "Yeah, no, sorry, we're just moving backwards folks." That is never going to happen. Fundamentally, ultimately what that means that the question of whether this is a worthwhile investment will be whether the adoptions will be able to produce and yield results that people will consider to be improving their lives or whether at some point there's going to be some backlash coming.

(40:59):
At least for now, I don't see them going down considering the rate of innovation that is happening on the back of that.

(41:05):
Well, it's been a fascinating discussion. Thank you everybody. And with thanks to my guests, firstly, Matthew Wright.

(41:11):
Thank you very much.

(41:12):
And to Bernard Maher.

(41:14):
Thank

(41:14):
You.

(41:14):
And Walter Pascarelli.

(41:15):
Thank you.

(41:19):
The AI revolution isn't just a battle of software algorithms. It is a brutal capital intensive race for physical resources in a world already stretched by geopolitics and resource availability. Then add in the demand for more gigawatts of electricity, the race for advanced semiconductor fabs and specialised data centre infrastructure and the winners and losers of the next five years is anyone's guess. But the physical constraints are real. AI's value and hype can't be realised without power and supply. The demand is forcing nations and companies to rewrite their energy plans, industrial policies, and capital allocations. Amidst market and societal volatility, another word is being rewritten, security, national security, energy security, data, and even supply chain security. Governments, companies, and investors alike are all yearning for that one intangible thing, security. Coming up in our third and final episode, we look at the secure and global movement of people.

(42:32):
We dive into what is driving our future, be it electric or autonomous. We map out the high stakes race to move the world and the power structures building the systems behind it all. I'm Tom Parker and this has been the next Five Podcast. Thanks for listening.