Welcome to Pathways, a Macquarie Asset Management podcast where we provide fresh perspectives and insights for institutional investors and consultants about real assets, private markets, and macroeconomics.
Daniel McCormack
AI is, without fail, the number one question we get from investors right now. And most of the commentary out there is either too excited or too dismissive. We wanted to go further than the headlines - to dig into what impacts we think it’s going to have specifically on infrastructure. So we published a paper, called What AI Can't Build, and it's the foundation for section one of this podcast. Our global infrastructure strategist, Aizhan Meldebek, wrote it. And right out of the gate, she chose to start with the jobs question. Specifically, a question about where AI is actually having an impact in the labour market, versus where it could have an impact, even if not yet observed. The answer, it turns out, tells you something important about infrastructure.
Aizhan Meldebek
The narrative usually around AI and jobs tends to be quite negative, which can impact consumption growth. However, if you look at the historical evidence from prior technology transitions - whether it's the rise of IT or general electrification - the net employment effect has typically been positive. This time around, there will be job displacements and some jobs destruction, currently estimated at about 8% of total employment over the next five years. But there will also be job creation, estimated at 14%, which means the overall net effect on employment is about 6% positive. And the net employment typically supports consumer spending and discretionary travel demand.
Daniel McCormack
Thanks very much, Aizhan. I'm a bit of a YouTube and podcast junkie, and I often hear tech billionaires talking about unemployment rates of 20%, 25%, 30% as a result of AI. Like you, I'm very sceptical of that. The only times through economic history where we've had those kinds of unemployment rates have been when you've had a massive negative demand shock - usually caused by some kind of financial crisis, a banking crisis, a currency crisis, something like that. It's never happened when we've had a big positive supply shock. It's never happened when we've had a new technology come forth, because as you very rightly point out, new technologies create jobs as well as destroy jobs. It's that fantastic creative destructive process that is at the heart of the entire capitalist system. So yes, I'm with you on that - I don't think we're going to see the kind of unemployment rates that some people are worried about.
That employment picture matters for infrastructure investors. But Aizhan's paper also does something I found genuinely useful, which is to go sector by sector and ask: where does AI create demand, and where does it threaten it? The paper maps this out in what she calls a heatmap - ranking each sector we invest in by its exposure to AI across two dimensions: the impact on demand, and the impact on operational efficiency. I asked her to walk us through the key findings.
Stepping outside the digital infrastructure space and to infrastructure more broadly - which sectors do you think are the biggest beneficiaries here, and what are the benefits?
Aizhan Meldebek
Thank you, Daniel. So I think data centres are obviously the most direct and obvious beneficiaries, but I would also highlight renewables and power grids as also compelling. When we talk about renewables specifically, hyper-scalers are increasingly the largest buyers of clean energy globally. For example, in the Americas region, the top 10 data centre developers accounted for more than half of clean power procurement by corporate PPAs in 2025. We also see a lot of grid capex required over the next decade - in the US, the projections point to about $1 trillion US dollars, which is roughly $650 billion in distribution and the rest in transmission. In Europe, this number is even larger, at around $1.2 trillion US dollars over the same period. So these are very large investments required in the space.
Fibre I would rank as number four, but of course it's benefiting from the AI-driven traffic growth between and within data centres, and it's creating a step change in demand for high bandwidth, low latency connectivity. In the US, fibre miles needed are projected to more than double, from around 160 million fibre miles in 2024 to almost 400 million miles by 2029.
Daniel McCormack
AI cannot replace a pylon, a runway, or a water pipe. But it absolutely can make them more valuable - or it can expose the ones that are falling behind. Nowhere is that more true than in transport, where the picture is genuinely different depending on which asset you're looking at.
Aizhan Meldebek
I think there are also other sectors that are not directly benefiting from AI and are more nuanced. Transport is a really good example of why we can't just make a blanket call on AI's impact on infrastructure - you really have to go sector by sector because the dynamics are genuinely different depending on what kind of asset you are looking at.
Starting with airports: historically, the core demand driver has really been GDP growth and middle income consumption growth, and global air travel has grown at about 1.8 times GDP growth. So airports carry a structural multiplier on economic activity. To the extent that AI delivers on its productivity promise and lifts long-run economic growth, airports are actually well positioned to benefit from that.
Beyond the GDP growth story, there is also a change in the mix of travel. There is a genuine question about business travel: if more meetings happen over video and AI makes remote collaboration more efficient, does business travel soften? Potentially yes, but what we're also seeing is that leisure travel growth could compensate, because AI-enabled flexible working is likely to expand discretionary leisure time - extending, to some extent, the shift we already saw post-COVID, where reduced business travel was more than offset by increased leisure travel across many airport assets.
Moving to toll roads, there is a similar story in terms of GDP linkage, particularly for roads with heavy freight and logistics exposure where volumes are driven by e-commerce and supply chain patterns, which AI can actually potentially accelerate. We could also see more long-term upside from autonomous vehicles unlocking new demand from population segments that currently don't drive - but we're really speaking long term. The risk for toll roads is more the exposure to urban commuter roads, where AI-enabled remote and flexible working reduces how frequently people commute. Roads with commuter traffic exposure need to be more carefully stress tested.
For container ports, the picture is broadly positive in the near-term because AI can help improve berth planning, crane scheduling, and equipment utilisation - real operational gains that can increase throughput. But there is potential risk on the longer-term horizon. If AI and robotics reduce the cost of domestic manufacturing sufficiently, you could see some reshoring of lower value goods production, which would reduce container import demand. This is more of a medium to longer term consideration, as any reshoring would likely prioritise higher value manufactured goods first, which have less of an impact on container volumes.
So while it's a bit of a mixed picture for demand impact, there are still very strong operational efficiency gains for transport. For example, AI-enabled energy monitoring has delivered annual energy savings of up to 20% for airports that have implemented it, and AI-driven toll management has increased throughput by 40% compared to traditional toll processing systems. The operational upside is significant and consistent across all infrastructure sectors.
Daniel McCormack
The sector most exposed to AI disruption in our infrastructure universe isn't software-adjacent. It might be the car park. Aizhan's paper makes the case that office-centric parking assets face a genuine structural headwind - autonomous vehicles and flexible working pulling in the same direction, over time. Airport and retail-adjacent car parks are a different story. But it's a useful reminder that even within a single sector, the asset-level detail matters enormously. The paper also covers healthcare and waste management - both net positive, for different reasons - and we'll link to the full paper in the show notes.
But here's the thing. Everything Aizhan has just described assumes the AI buildout continues more or less as expected. And a question we hear a lot that I put to her recently - what happens to infrastructure if the AI investment cycle disappoints?
And Aizhan, from your perspective - coming with your broader lens on infrastructure as a whole - what do you think would be the impact on the asset class if we were in a bubble and if it was to burst?
Aizhan Meldebek
From the research perspective, we have looked into the IEA's data centre power demand scenario analysis, and even in the most cautious scenario - what they call the headwind scenario, where AI monetisation disappoints and investment pulls back - data centre power demand still grows at a CAGR of roughly 5% to 2035. So it slows market growth, but it doesn't reverse. And there is an important reason for that: AI is not the only demand driver for data centres. There is cloud migration, e-commerce, digital government, fintech workloads - all of which are expected to grow independently of how generative AI evolves. From that perspective, the case for utilities and power infrastructure remains intact even if there were a bubble and it burst. Data centres actually represent less than 10% of the power demand growth to 2030, so the overall utility capex cycle should still be driven strongly by industrial electrification, space cooling, and the general energy transition - well beyond data centres.
Daniel McCormack
So even in the downside scenario, the infrastructure case holds. The growth slows but it doesn't reverse. And the reason, as Aizhan puts it, is that AI was never the only demand driver. The structural case remains intact.
That's the macro view. But I wanted to pressure-test it with someone who's doing this on the ground - looking at actual deals, actual power connections, actual sites. Anton Moldan is our Head of US digital infrastructure, and he's been building data centres for this company for the better part of a decade. I sat down with him to find out what the market looks like from where he's standing."
Let's say all the spending on AI is a bit of a bubble and it falls off a cliff. What's the impact on digital assets as a result? You've talked about this before - these guys have got 15-year leases with all the hyperscalers - so the operational performance is fairly resilient to this thing slowing down. Is that right?
Anton Moldan
Look, I would say our general response to this is that structurally we're not in a bubble. This is a structural change and a required investment - AI has moved quickly from a novelty to a necessity, and it's nearly existential to some of these very large customers. But whenever there's what you might call a gold rush, there are microbubbles. And where we see that is a lot of people doing speculative capital - which is getting larger and larger - for land make-ready, utility down payments, deposits for long lead items. We probably get a dozen of these pre-lease capital deals a week at $200 million. And I think a few of those will be very successful and a lot of those will be left in the red.
Daniel McCormack
What are the sort of challenges that you're seeing? Is it the NIMBY stuff?
Anton Moldan
Obviously NIMBY is making all this much more complicated. You're seeing moratoriums on large builds as people want to better understand the impact on utility networks and consumers. When things become harder - specifically in the risk of overbuild - we actually kind of like those somewhat complicated barriers because it leaves the people that can navigate those complexities well positioned, but it is causing natural barriers, including capital barriers, which I think is healthy.
The last observation is that we're suddenly seeing single-site campuses become multi-gigawatt campuses. The concentration of capital on single-site campuses I think will also cause a bit of strain on capital markets, as well as raise questions about the ultimate value of those things - specifically some that may not have a tangible path to utility connection. There is a bit of an existential question on the terminal value of those.
Daniel McCormack
And that's because you just worry that they can't get connection - that's the main thing?
Anton Moldan
It's that when you've got these single sites that are $40, $50, $60 billion of value, it's just an unprecedented level. And we're all financial investors, and 101 of financial investing is diversification. When you're pulling away diversity and you've got concentration of an operation at a single site, concentration to a single power source, concentration to a single customer, a single regulator, single geography, single labour market - it just causes a lot of capital concentration. People are chasing the growth today, but I don't think we fully understand what that type of asset base looks like.
And specifically, in a time when capacity is so tight, there is a race - nearly everyone is breaking all their rules in their own books, specifically around power solutions. While they all want a path to utility connectivity, they are sometimes taking behind-the-metre solutions or even permanent behind-the-metre solutions, nearly out of desperation. And specifically if those ones are not in core or tier-one adjacent markets, the residual value of those if that tenant disappears is questionable.
Daniel McCormack
Sorry - your first point was about pre-leased capital. Are you talking about data centres that have got leases signed up to them, that those projects come to you to finance? Walk me through that again.
Anton Moldan
Yeah - the amount of capital required pre-lease is expanding because of utility down payments, utility letters of credit, the timetable for delivery, the further stretching of supply chain, the need to put deposits down on long lead items. What we're seeing now is single sites adding to the hundreds of millions of dollars in pre-lease capital requirements, which in one way is actually a healthy requirement from the utilities.
What was happening historically is one customer goes to a market and says they need 1 gigawatt of capacity, then 10 providers go to the utility and say they've got a gigawatt customer. The utility is thinking it has 10 gigawatts of demand when it really is only 1 gigawatt from the underlying customer. Everyone was getting into these queues and the queues were overly inflated. So what the utility is saying is: we need deposits down to maintain more thoughtfulness about what this actually looks like.
It does mean that some small, undercapitalised players are struggling, and we'll see a lot of assets come back to market. But it also means the capital one needs to spend for pre-leasing - the speculative capital - is accelerating and concentrating. And that is where I expect we'll start to see some meaningful capital losses.
Daniel McCormack
That's just adding risk to the whole equation. What in your view is key to successful deployment of capital in the space at the moment?
Anton Moldan
At the moment it's very much a rising tide, and there's a lot of buoyancy in the market, but that level of supply-demand imbalance won't structurally be there for a prolonged period of time. Also, our customers will over time rebalance to insourcing. So we're very much focused on sprinting in this time to capture value in the short window where that insatiable demand is there - but really leaning forward on execution to make sure that when we get through this period, we are seen as one of the preferred partners of choice to the hyperscalers because we did what we said and we delivered on time.
I think we're going to see a lot of challenges where there's an ability to differentiate in this environment. That really means working with experienced management teams, thoughtfulness to underwriting, creating appropriate tone and transparency in discussions with customers. But ultimately you need product to sell - which means you've got to de-risk supply chain and de-risk powered land.
Daniel McCormack
OK, so you sound a bit more concerned than a few months ago?
Anton Moldan
It's the same concern we probably had for 18 months. We don't expect capital risk - we only deploy capital once the lease is signed, we come in below cost base, so our returns aren't predicated on multiples staying elevated over a long period of time. I think we're just seeing more and more of that pressure cooker environment based on that pre-leasing capital need.
Daniel McCormack
What's the latest with power connections? Power's not keeping up, I take it - that's still a problem?
Anton Moldan
Look, there is still excess power and power availability in the markets - it's just going to more non-traditional markets where you've got excess power. We've seen that in Alexandria, Alabama, Oklahoma, North Dakota, and certain markets in Pennsylvania that haven't been traditional data centre markets. The biggest solution now is how does one go to the utility and bring both load and off-take - you're actually providing that solution, and depending on the market you have the ability to effectively jump the queue if you're bringing load. Rather than just bringing power behind the metre, how do you bring more connectivity and load to the utility?
Daniel McCormack
I'd now like to talk about this difference between training compute on the one hand and inference on the other. Is it as simple as saying that in the US it's all about training and compute, and in Europe it's all about inference?
Anton Moldan
No, I think there's a lot more nuance. There has been significant investment in training in the US, and that really comes back to what it takes to get AI off the ground - the commercialisation aspect is inference, which requires significant investing in training first. What we do spend time on is thinking about how training and inference are located and which markets they're going to.
One clear observation we've made is that the topology of AI infrastructure is going to be structurally different - specifically in the US - from how cloud availability zones were built. Cloud availability zones were built around the density of population centres and network centres. AI is really being built in large clusters that allow for significant parallel compute, and where the availability of those large clusters hasn't been in those traditional tier-one markets, it's really shifted the build to more frontier markets.
There was a misconception that these markets were only for training and that inference had to be at the edge in tier-one markets, driven by latency requirements. What we're actually seeing is that the biggest adoption curve on inference will come from enterprise wallet share. Only a small subset of use cases will need real-time, low-latency applications - like a call centre emulating a real-time conversation. But the biggest use cases - data synthesis, data summary, content creation, document creation - a lot of this doesn't need real-time application. One, two, three seconds is really fine in response time. So the topology of even inference now exists in these frontier markets.
And when we talk to customers themselves, they're actually planning the deployment of inference in these markets and they like the co-location of inference and training, or the flexibility of facilities to do either as they manage their overall workloads - because inference and training is ultimately a reinforcement loop. Training continues to happen every day, and how inference is being used is actually a key component of that training.
Ultimately, the topology of AI compute is going to be very different to cloud. There will still be some inference that needs to exist in more traditional edge and cloud markets, but the overweight of deployments and builds will be these newer frontier markets, where there can be large-scale campuses with big clusters of GPUs doing both work products.
Daniel McCormack
And, and how is it being used in digital, you know, the moving systems and..?
Anton Moldan
The training versus inference question is a big topic, and one we get asked a lot by our investors. I think it's one where we've had a slightly different view to the market - the market has misunderstood and mispriced how training and inference is deployed and the topology of that deployment.
You can access most markets from anywhere in the US in under 100 milliseconds round trip, so the movement of data has a limited impact on the overall compute lifecycle.
When we speak to customers directly, we're typically hearing that they're planning 60 to 80% of inference to be deployed in frontier markets in the US. A lot of people have been nervous about those markets. We've been leaning into them because we ultimately think the topology of cloud availability zones and AI availability zones are going to be quite structurally different.
Daniel McCormack
Have we missed anything from your perspective? Is there anything else?
Anton Moldan
Yes - I think we generally think of the continuum of AI in three phases. The first phase is just embedding tools like CoPilot and Claude to improve responses, data synthesis, and data creation, and using AI to review decision-making. For example, routing software or staffing call centres - you can look at call volume, missed calls, and staffing and get a response very quickly suggesting changes. Things that help you immediately with just the implementation of AI.
Phase two is where we see the highest friction points in manual or repetitive work. We use that to create efficiencies in repeat tasks and automation.
Phase three is much more the agentic phase, which really requires a full review of workflows - not just automating what we do, but changing the way we do things. Now you've got agents that can work 24 hours a day; you don't have that limitation of staffing and resources.
I would say we are very much finding ways to add value in bucket one and bucket two, and probably doing more long-lead planning for bucket three. The first two buckets could be as simple as improving consumer experience through better chatbots, or reviewing decision-making in routing - anywhere there's a more complex environment or a lot of unstructured data. You may have, for example, a churn reason code in analysis that's very hard to distil if it's non-structured data, where AI can be very helpful. So I think we're very much still at phase one and phase two.
Daniel McCormack
Phase one and phase two. Which, if you think back to where we started - to that gap between what AI is theoretically capable of and what's actually been observed in the labour market - is exactly where we are. The observed reality is phases one and two. Phase three, the agentic phase, the full reinvention of how work gets done - that's still largely theoretical. And the infrastructure to support it, when it arrives, is being built right now. In frontier markets across the US. At a scale and a pace that would have seemed implausible five years ago. Whether the technology catches up with the buildout - that's the question. And it's one we'll be watching closely.
Thanks for listening. You can find a link to Aizhan's paper, What AI Can't Build, in the show notes. And if you'd like to discuss any of the themes we've covered today, please don't hesitate to reach out to your Macquarie Asset Management Relationship Manager.
Disclosure
This recording is intended as general information for financial professionals and qualified investors only. This is not intended for use with the general public.
The views expressed in this podcast represent those of the speaker and are subject to change. Nothing presented should be construed as a recommendation to purchase or sell any security or follow any investment technique or strategy and does not constitute advice, an advertisement, an invitation, a confirmation, an offer, or solicitation to engage in any investment activity or an offer of any banking or financial service.
Investing involves risk, including the possible loss of principal. All examples herein are for illustrative purposes only, and there can be no assurance that any particular investment objective will be realized or any investment strategy seeking to achieve such objective will be successful.
Past performance is not a reliable indication of future performance. Before acting on any information, consider the appropriateness of it with regard to your particular objectives, financial situation and needs, and seek advice.
No representation or warranty expressed or implied is made as to the accuracy or completeness of the information, opinions, and conclusions presented. This recording may include forward looking statements. No representation is made or will be made that any forward-looking statements will be achieved or will prove to be correct, or that any assumptions on which statements may be based are reasonable.
In preparing this recording, reliance has been placed without independent verification on the accuracy and completeness of all information available from external sources.
Macquarie Asset Management, MAM, is the asset management division of Macquarie Group. Macquarie Group refers to Macquarie Group Limited and its subsidiaries and affiliates worldwide.
Any Macquarie Group entity noted in this podcast is not an authorized deposit taking institution for the purposes of the Banking Act 1959. The obligations of these Macquarie Group entities do not represent deposits or other liabilities of Macquarie Bank.
Macquarie Bank does not guarantee or otherwise provide assurance in respect of the obligations of these other Macquarie Group entities.
In addition, if this podcast relates to an investment, the investor is subject to investment risk, including possible delays in repayment and loss of income and principal invested, and none of Macquarie Bank or any other Macquarie Group entity guarantees any particular rate of return on or the performance of the investment, nor do they guarantee repayment of capital in respect of the investment.