Brad Rogoff: Welcome to episode 76 of The Flip Side. I'm Brad Rogoff, Global Head of Research here at Barclays. Today, our conversation focuses on a debate that's captivated markets and investors, the outlook for AI as an investment theme as capital expenditures rise to unprecedented levels. And you can't be an equity strategist and not have a view on AI. Luckily, our Head of US Equity Strategy, Venu Krishna, has a pretty strong one. Venu, thanks for joining us. Let's get into it.

Venu: Thank you for having me, Brad. As you alluded to, AI is perhaps the defining investment theme of the current cycle. The potential opportunity is immense if AI proves to be as transformative as PCs, the internet, mobile and cloud computing were in years past. It is this opportunity, the fear of missing it, that has kept investment dollars flowing towards US AI innovators. AI has become the biggest, some may say only, secular growth narrative within the US risk asset landscape. And I think it will remain central to the upside case for US equities as we look into 2026.

Brad: The opportunity may be real, Venu, but I'd argue that the market's optimism may be running ahead of itself. I mean, almost $400 billion in hyperscaler CapEx announced for this year, projected to increase 30% annually for the next several years. Nvidia thinking we'll need $4 trillion in total AI infrastructure spending by 2030. These are massive numbers. All this while big tech and semiconductor stocks are trading at what multiple?

Venu: About 29 times forward earnings.

Brad: 29 times. So we should be taking a hard look at just about anything at that price at that multiple, right. And these sky-high valuations combined with massive CapEx numbers and an unclear path to really return on investment, that has many folks drawing comparisons to the most famous example of speculative tech frenzy outrunning the intrinsic value of companies. Obviously, I'm speaking of the dot-com bubble.

Venu: Brad, valuations are certainly full right now. I won't deny that. But we are not in a bubble territory quite yet. Comparisons to dot-com are everywhere, but I would first argue that 29 times forward earnings real earnings, I might add, are a far cry from the 70 times multiples we saw for the tech sector back in 1999, when a lot of internet companies were making zero profits. Second, hyperscalers today are on much stronger fundamental footing than the big tech infrastructure builders were 25 years ago. Take, for example, CapEx to sales. Hyperscaler CapEx has been climbing steadily since 2023 as the AI arms race picked up speed, and it is now 25% of sales. Significant, but still well below 40% we saw with telcos in the early 2000s.

Brad: I know the valuations aren't as extreme. And look, frankly, the next crisis never repeats the last one. But we're still talking about massive numbers, 25% of sales versus 40% of sales. It almost feels like we're splitting hairs a little bit. And also, if CapEx is going to grow at 30% annually for the next few years, that CapEx to sales ratio is not staying at 25% for long. It's probably increasing from there since revenue growth for the hyperscalers is kind of more like 10 to 15%.

Venu: That is a fair point, Brad. But CapEx to sales is only one of the metrics we consider. Hyperscalers are also not nearly as levered as your average dot-com era telco was, with net debt a fraction of EBITDA rather than a multiple of it. In addition to carrying less financial leverage, hyperscalers are generating more operating leverage, expenses are increasing, but revenues are keeping pace. At no point during the dot-com bubble did the big telcos show the kind of margin expansion and cost discipline that we have seen from modern day hyperscalers? I truly don't think that it is comparable.

Brad: But back when we were partying in 1999, there were plenty of people arguing this time is different. So fine, everyone who listens to this knows I'm a credit guy by background. Leverage is lower. I got to give you that. And these hyperscalers are generating a ton of cash. But neither of those things change the fact that the scale of investment is just unprecedented. Mark Zuckerberg acknowledged there's a risk of misspending a couple hundred billion dollars. Sam Altman has said that AI stocks are overvalued. It doesn't have to be on the scale of the dot-com bubble, but if demand doesn't keep up, couldn't we see a massive adjustment in valuations and CapEx?

Venu: Look, Brad, I am optimistic about the outlook for demand. Over the most recent earnings season, already one in ten US companies discussed AI in the context of efficiency gains or productivity improvements from 20 to 40% reduction in software development time, to 30% improvement in marketing campaign ROI, to 50 to 80% reduction in select workloads. AI service provider revenues are moving up and to the right on a steep slope. So I think we are in a good spot for demand and adoption to continue scaling from here, especially as advanced reasoning models and agents begin to show their capabilities. But at the end of the day, Brad, given the kind of numbers we're talking about, I do think it's prudent to consider how all this investment could go wrong.

Brad: Okay, there we go. So with my cynicism here, I'm happy to dig into that a bit. One possibility is that future AI infrastructure requirements are simply being overestimated given the direction that the technology is going.

Venu: That is a real possibility. Brad. Initial training of a base large language model has historically been the most resource intensive phase of AI development, and is a big reason why demand for data centers has greatly outpaced supply thus far in the AI arms race. There are some signs that this demand mix is shifting, and initial training is being replaced with fine tuning models for specific tasks, and of course, simply running the models to generate output. What's referred to as inference. Both are faster and cheaper to run in real time compared to initial training, which could theoretically flip the current supply demand imbalance and reduce the need for further CapEx. I think this is a scenario that is drawing so many comparisons to the dot-com bubble. Folks are waiting to see if AI will have its dark fiber moment when exponential efficiencies lead to critical underutilization of infrastructure assets.

Brad: Well, I do share some of the concerns about the folks that you just mentioned. The last time we discussed AI on this podcast was following the DeepSeek moment, so to speak, at the beginning of the year. If we go back to your dot-com analogy, ROI was eventually realized on the network build out from 25 years ago, but too late to benefit the companies that broke ground for it. And we're talking about assets with a much longer useful life than the ones that are mission critical for AI. When a newer and faster Nvidia GPU comes out seemingly every year, and it's a must buy for companies trying to stay ahead in the AI arms race, it seems like supply versus demand is not the only mismatch we have to look out for here. It's also expected versus actual depreciation timelines. How many stranded assets could we be looking at here?

Venu: It's a valid concern. But this is where I think there are some silver linings in the demand mix shifting. As we talked about earlier, AI adoption is already underway and I think it has plenty of room to run, especially in enterprise. How this shows up is inference, which may be faster and cheaper than initial training. But as small models and agents are deployed for specific tasks, that market will continue to grow. This should support demand and useful life of chips that may not be at the absolute cutting edge, especially as software optimization leads to models becoming more efficient since inference does not necessarily require the performance or efficiency of a next generation chip.

Brad: So let's take that efficiency point and pivot to a potentially immediate threat to data center investment, which is power. Next generation chips need to become more efficient because data centers are already creating massive strain on the aging US power grid. So our thematic research team has done some great work on this. They call it their Powering AI Series, and we've discussed it in past Flip Side episodes. But the gist of it is that data centers could account for about 12% of total US electricity demand within three years, which is three times what it was in 2023. Electricity pricing surging and the DOE is forecasting widespread power outages by 2030 if data center demand keeps growing and planned capacity retirements are not replaced. This doesn't sound sustainable to me.

Venu: It's a real risk, Brad, and one I think is acknowledged by AI infrastructure players. We are seeing data centers increasingly turn to off grid power as a stopgap or replacement solution as power supply becomes de-bottlenecked. For instance, building small scale, independent microgrids that can power a data center autonomously using natural gas generators, fuel cells, batteries, or a mix of sources. But I think in the medium to long term, we will need more private and public coordination to improve US power generation. Otherwise, there is a material risk that data center investment will be forced to slow because there is just not enough power to supply future installations.

Brad: All right, so what I'm hearing here is we might have too many chips and we could run out of power to operate them. Here I was thinking that I was going to be taking my kids to play laser tag in empty data centers in a few years, and maybe that isn't even possible. Anything else that might shake your confidence?

Venu: Sure. Funding requirements are something I'm keeping an eye on, especially with some of these circular financing concerns that are cropping up. As we talked about earlier, hyperscalers are mostly funding their CapEx with operating cash flow. But this doesn't mean that debt issuance is not somewhere in the equation, especially on the private side. We have seen several multi-million-dollar private deals announced in the AI space this year, and there's almost $5 trillion of value locked up in VC funded companies, a big chunk of which is AI related. As data center investment continues to grow, we have to assess whether the next leg of funding is on as strong a fundamental footing as the previous one.

Brad: The private credit side, certainly something I can opine on. I'm less worried about the strength of those lenders and more what they might do if one of the assets they lend against is not as attractive as they originally believed. Then they might start to question the LTVs of the whole portfolio, and that's where I think you could see some issues there. So clearly, though, you're more bullish than me in your base case, but we've just talked through some potential risks. What's the real bear case here though, and does data center CapEx grind to a halt?

Venu: No. Absolutely not. Remember that while AI is growing quickly as a share of total data center demand over the last few years, the majority, let's say 90% as of 2024, is still from traditional public and enterprise cloud workloads. Cloud is a maturing tech ecosystem, though showing some signs of decelerating growth is still expanding at double digit pace annually. So I think a reasonable bear case, should the AI narrative stumble, is for data center CapEx to decline 20% in total over the next two years. There is some precedent for this scale of CapEx pullback. We saw that with Amazon and Meta in 2023, Amazon reined in its warehouse buildout and Meta pivoted away from metaverse projects.

Brad: So down 20% sounds pretty negative to me when we consider that the hyperscalers are forecasting 30% growth annually over the next several years, as we mentioned before. So if this actually comes to pass, semiconductor stocks are probably the last place that an equity investor would want to be, but I can't imagine it stops there. AI development is dependent on immense physical data center infrastructure beyond just the chips. We're talking connectors, storage, networks, power and cooling required to support these processors, not to mention the buildings that house them.

Venu: That's right, Brad. The market reaction to DeepSeek back in January offers some clues here. As you discussed in episode 69 of The Flip Side, that was perhaps the closest we've gotten to equity investors pricing in material risk of that dark fiber scenario for AI. While tech stocks were routed, subsectors within energy, industrials and utilities were hit equally hard since they were material beneficiaries of the data center build out via power generation and electrification. I think this illustrates how diverse the AI economy has become, and how much growth has gotten priced into valuations across a widening swath of industries.

Brad: Let's get back to your day job for a second here. So from a strategist's perspective, how does this flow through to US equities? You made a point earlier the valuations aren't telecom bubble extremes, but what kind of downside could we be looking at in terms of earnings or multiples for the S&P 500 if data center CapEx pulls back, and we'll use your example of 20%?

Venu: Interestingly, the headwind for earnings may not be as big as you expect at the index level for two primary reasons. First, while data centers are often the fastest growing segment of the business for companies in this space, current exposure still tends to be quite modest for most industries except for semiconductors. Second, managements appear hesitant to bake aggressive data center growth into guidance, preferring to leave room for upside surprise.

As a result, a 20% step down in CapEx shakes out to mid-single digit percentage downside for S&P earnings next year, in my view. That being said, share prices could be affected more severely considering the growth that is currently priced into valuations. I think S&P PE multiple compresses at least 10% in this scenario, as AI infrastructure beneficiaries, hyperscalers and app developers unwind some of their multiple expansion that was seen in the last two years.

Brad: So mid-single digit downside for EPS and a huge drop in growth expectations. I think that has to put you in an outright bear market for the S&P 500, not just down 10%.

Venu: It's possible, but I think aggregate downside would be offset by rotation into other segments of the market. Turning back to DeepSeek, around 70% of S&P constituents were actually up on the news as bids for value, quality and safe haven stocks offset tech centric losses. Buyer demand shifted towards not only traditionally defensive exposure, but also select cyclicals with favorable outlooks. Hyperscalers could also end up in a better position to return cash to shareholders via buybacks if data center investments is wound on. Remember, by cutting CapEx, free cash flow goes up.

Brad: That is a fair point. But I still think we're overlooking something here as far as the potential downside. When spending on anything scales to the hundreds of billions, we've officially crossed over into the macro arena. Aggregate expenditures on three GDP categories, so those are computer and peripherals, software and data centers contributed about one percentage point to quarterly GDP growth seasonally adjusted in the first half of this year. So by comparison, GDP growth averaged just 1.4% during these quarters. That's more than two thirds of overall activity attributed to data center investments so far this year. I know that concentration within US equities has been a hot topic for some time now, but the same kind of dynamics for the entire US economy makes this an even larger and different ballgame.

Venu: I think you're right, Brad. Data center disinvestment as part of a larger macro dislocation would be a far worse outcome to risk assets than any of the scenarios we've talked about so far. Let's say US macro takes a turn for reasons that have nothing to do with AI. Maybe the labor market deteriorates sharply, or tariff impacts prove worse than expected. A simultaneous reduction in AI CapEx could compound the downside materially given its centrality to the US economy.

Brad: So it sounds like even you would admit that we'd be landing within a bear market territory in this case.

Venu: Absolutely. Growth scare equity valuations like those we saw in April of this year and back in 2022 dropped S&P multiples into the mid to high teens, which would be a big air pocket considering we're almost at 23 times now. Also, intra stock correlations are pretty low right now, and bullish investors are looking for best opportunities in the next leg of the current cycle. These correlations tend to spike during market dislocations, meaning that while high valuation AI winners would be among the worst positioned in a sell off, they are most likely to take the rest of the market with them.

Brad: Look, when I think about how much time I am spending on AI and the potential for productivity enhancement, I understand some of the optimism. But when we were talking about such huge CapEx numbers that are still somewhat speculative, I worry about the chance of a deep sell off, even if the companies spending that money are in much better shape than those that spent the money 25 years ago. I even asked ChatGPT to settle this debate about whether AI was a bubble, and the first word that came back was “yes”.

Regardless, Venu, I still enjoyed the human debate. And thanks for listening to this episode of The Flip Side. If you like what you hear, don't forget to subscribe. And if you're a client of Barclays Investment Bank, you can read more on the AI debate in several recent publications available in Barclays Live, which we've linked to in the show notes. Until next time, see you on The Flip Side.