EP 148: Is the AI boom cracking? Inside July’s sell-off and beyond

Welcome to Beyond the Benchmark, the EFG podcast with Moz Afzal.

Moz Afzal:

Hi everyone. So today we have a very, I'll say deeply requested or highly requested podcast on the current situation in semiconductors, technology and everything that is AI. So to help me or us to navigate this journey, I have Henry Walters. Henry, welcome.

Henry Walters:

Hi Moz. Good to be back.

Moz Afzal:

And Henry is an equity analyst here at EFG to talk about and covers the technology sector. So he will be helping us to navigate everything that's AI and semiconductors and hyperscalers and so on and so forth. So Henry, should we just start straight away, just a quick summary of exactly what's been happening in 2026?

Henry Walters:

Yeah, it's been a busy year and an eventful year. It's gone through multiple stages of euphoria and brief moments of panic. But I'd probably summarise 2026 overall as a year where enterprise AI in terms of agentic AI adoption really hit an inflexion at the start of the year with the release of Anthropic's Opus 4.5 model and Claude Code. And the coding applications of AI really inflected at the start of the year. And you could see that in the numbers reported from Anthropic in terms of the revenue generation. They ended December at about nine billion revenues and in May hit about 45 billion annualised. And then these are all kind of private companies that aren't formally disclosed financials, but reportedly hitting 60 billion in June. And that really drove, that was one of the key drivers for AI is monetizing really, really well. The kind of AI CapEx trade that's fueled markets for the last few years really powered on once people saw, okay, it's generating real revenues.

But at the same time, the other kind of effect this year was all the kind of economics of the AI was really kind of being concentrated in the frontier labs. So really kind of anthropic at the charge and then OpenAI once they pivoted. And so that's caused some interesting developments more recently in terms of who's capturing the value. But that kind of takes you up to the end of June. And then July, if I look at the performance of semi-index year to date to the end of June was up kind of 82%. July hit a bit of a wobble to put it mildly with a few stocks down 40, 60%, that kind of range within one month.

So really kind of a big momentum unwind in terms of, it was very crowded positioning within July. But what I'd say is the fundamental data points that we look at and everyone's tracking in terms of how well AI demand is doing, you look at the cost of computes, GPU availability, tokens generated.

Is there a real concern that AI has hit a stall point? None of that was really happening. I think it was more just around market positioning, concentration, a lot of leverage, particularly concentrated in some Korean traders made the press in terms of single stock levered ETFs. And so there were certain things happening under the hood that caused this momentum unwind. The only fundamental kind of negative data point that you could point to in July during the selloff was rising rates, credit spreads widening for a lot of the hyperscalers, people investing in infrastructure. And that is a real potential kind of concern, something worth monitoring. I think it's kind of moderated since as we speak now. But other than that, yeah, there wasn't much else to point to in terms of negative signs for July.

Moz Afzal:

Yeah. Let's double click on those negative aspects because I think sometimes it is quite confusing and let's try and put it into very simple terms. So overall, and in fact, I've just updated the numbers, but overall global tech earnings for 2026 are in the region of around 70% up on last year. And this is already up sort of 30% plus from the previous year. They are very weighted towards the hardware side and semiconductors and anything that powers AI infrastructure. So the earnings side looks robust, looks strong. All the points you make are very good. But double clicking into the air pocket of July, one can say two things. First, real yield started to climb. 30-year US Treasuries hit a range high as well as bonds and gilts and so on and so forth. So that certainly had an impact. And the second impact was huge amount of issuance that then fed into spread widening.

So very large issuance from Meta is probably the one that was the most high profile of the names, but you had other smaller neo clouds and so on and so forth also issuing a lot of debt over relatively quite some period as well, which we shouldn't forget. So we've seen this sort of spread widening. And maybe to summarise, all of this meant that we are spending or companies are spending huge amount on AI, ultimately over the next few years, trillions of dollars. Part will be coming from operational cash flow. A lot of it hopefully, particularly from the hyperscalers. But there's also a lot that is debt fueled. And so in an environment where interest rates go up, that debt fuel slows the spending. So I think it's important to kind of put it in very simple terms exactly what you've just said. And I think that is the more realistic catalyst for the selloff.

And plus we had highly leveraged investors who got margin called and cut off that led to that exceptional volatility. And unfortunately that volatility probably will stay for a little bit, I suspect, for a few more months once people have kind of calmed down and kind of moved on. So I think that -

Henry Walters:

Yeah, you've obviously got for many of the big spenders, free cash flow going to zero or negative for Meta and Google. And then you also at the start of July had the reported news that Meta was looking to rent out some of their compute.

Moz Afzal:

Which surprised people.

Henry Walters:

Yeah. When you're raising CapEx and you've been raising CapEx and all of a sudden you're saying, "Oh, I'm going to rent some of it out to someone else." And the market's been believing, is that a sign of demand is weak? Are you overspending? Are we in an oversupply?

Moz Afzal:

I think the point that you've made or something you've made to me is that the Meta situation, let's cover that situation. I think it's quite interesting because you're right. People say, "Oh, you've overspent, you've got too much infrastructure and actually you need to rent it out." But there's a reason for that, right? Because spot price for a token or for that infrastructure was way higher than the long-term contract. Explain that dynamic.

Henry Walters:

Yeah, of course. So when you build these AI clusters, you can either use it as kind of a cloud and rent it out or you can use it for internal workloads. So Meta's obviously one of the largest digital advertising businesses in the world. However, this year as the enterprise AI adoption has taken off, that's kind of left Meta without that growth engine. So if we talk through how, if you're a cloud, if you're a neo-cloud or a hyperscaler or co-location data centre, how do you monetize? You typically rent out your compute with some kind of large marquee customer or a group of them or usually kind of a five-year contract. Why is it a five-year contract? Because otherwise no one's going to lend to you. How are you going to fund it if you don't have kind of backing and some real credible customer who's going to rent it off you? So you rent it on this longer term contract. They've been priced around, call it $15 billion per year per gigawatt.

What's happened since is the value of that cluster, that AI infrastructure has gone up. So the spot markets are much, much higher than that $15 billion per year. So SpaceX came in with some short-term contracts. They announced one with Anthropic, one with Google, and these might be 90-day contracts where they're talking about an earning school this week doing six month contracts. And instead of at $15 billion per gigawatt, it can be 30 or 50 at kind of peak pricing. And so if you're Meta, you're looking at that, you're thinking, oh, if I can monetize this at a much higher rate just by renting some of it out and I can use that to fuel my CapEx. It's a very attractive option. And Musk said on the call this week, if you can do $50 billion revenues, which we'll see if that sustains, it's a one-year payback on your CapEx.

Whereas if you're doing the 15 billion, it's still not a bad return. Andrew Jassy, CEO of Amazon, was talking about less than three year payback for these AM structures. You still get a good ROIC on that investment. But at the moment, because demand's been so strong, there's a real premium on what people are willing to pay for that compute.

Moz Afzal:

And I think that's kind of critical because the underlying strength here is the demand. So demand is literally off the charts and supply is not coming on fast enough to be able to satisfy that demand and hence you're creating these road distorted markets in the short term. So let's quickly go through the results that we've had so far.

Henry Walters:

Yeah, I'd say Meta, there's a lack of clarity around how they're going to monetize. They talk about selling enterprise services, becoming more of a cloud company, but that requires a very different kind of go-to-market distribution function. And yet they're still investing heavily. I think the other narrative that the market always gets fixated on is who's at the frontier in terms of model development. So the market's always a bit skittish around model leadership. But in aggregate, in terms of the cloud results, so across Google, Amazon, Microsoft, you're seeing very, very strong acceleration. I think as a group growing over 40% year over year and doing well over 300 billion of annualised revenues across the three of them, these are massive businesses that are accelerating. And the operating cashflow also supporting that and inflecting upwards, which is a positive sign in terms of how much debt they need to raise if operating cashflow can keep up with their CapEx ambitions.

Moz Afzal:

And I think that's the critical function here is as long as your operating cash flows are moving higher and are strong, market will forgive you for spending basically. If your main operating business is not doing as well and you're spending and you have to go to the debt markets or even equity markets to raise capital, that's probably a more negative signal. And that sort of inflexion point is going to be the key thing I guess to watch.

Henry Walters:

Yeah. Yeah, I agree.

Moz Afzal:

Okay. So let's talk about the semiconductor companies. How they look so far?

Henry Walters:

Yeah, they move very differently. There's a lot of different sub-themes, I'd say, within AI infrastructure as a whole and different flavours of the month appear and disappear. And what you've had is, I'd say, in aggregate during the year, we've obviously had the memory shortages and memory pricing. Memory's been a massive theme this year. It's a much more commoditized market, but pricing is very susceptible to any kind of supply demand imbalance. And foreseeable memory shortages has driven up prices of DRAM and NAND well, well above 35X versus what it was prior. So memory's seen a huge resurgence, but tends to be much more volatile. And then you have -

Moz Afzal:

And that's why you have these sort of crazy situations in memory where the price earnings multiples of these companies are low single digits. But the reality is they're not necessarily priced off price earnings multiples. They're usually price to book or other measures that you'd use to value in those types of companies. But I think it's a tricky one because you just don't know when the peak's going to be.

Henry Walters:

Yeah, that's why everyone's watching price of compute, that's the main signal or how well is AI monetizing is really a signal for demand. And yeah, so far it's remained very, very strong. And the infrastructure budgets are growing confidence for continually rising.

Moz Afzal:

So overall, not much concern in terms of semiconductor and the food chain of semiconductor. Results looked actually, I guess, okay, but it's the frothy part of the market. And that's when you've seen these share prices move up a couple of hundreds of percent, in some cases, a thousand percent in some cases in a relatively short space of time. That's kind of where the froth is sitting in there and hence the volatility is just also is off the charts.

Henry Walters:

Yeah. And it's just been a very narrow market. I very much focus on tech, but you know the other parts of the market far better than me. And a lot of managers are chasing what's momentum and what's been working just to try to keep up. So that kind of whole positioning effect and crowding into some of these names amplifies everything on the upside, but similarly on the downside.

Moz Afzal:

So let's move on to talk about the language models, how Claude is doing open source versus closed source.

Henry Walters:

Sure. So I think again, repeating myself, but the challenge is these companies aren't public. So you rely on analyst estimates and company sporadic disclosures on what the financials are of an anthropic or an OpenAI. But if I go off what's reported for Anthropic, as you mentioned, the premium kind of frontier models. So that being Anthropic suite of models and OpenAI as well, they've captured the lion's share of the economics from AI software tokens, the applications built on top of that. And if you look at the reported financials, OpenAI, Anthropic might be at 60 billion revenues. And the kind of part people debate the most is what margin do they earn on that? Which I've seen from 60% gross margins up to 80% plus, which aligns with what we were saying earlier about if they're willing to spend not just $15 billion a gigawatt, but 30 or maybe higher, they must be monetizing it very, very well at 15 billion.

And what you've had is collectively a lot of the tech industry or the major players are getting or have been quite nervous that a lot of the power and the value capture is concentrated in these frontier labs. And so open weight models and what they are are models where you can freely download them. They're released for free. Anyone can download them. They're still massive models, so you still need to run them on servers, but you can host it yourself and you don't need to pay a big chunk of margin just for the privilege of using that model, which you'd pay to Anthropic OpenAI. And typically they lag in capabilities. They're not quite as powerful as the best models. What was last year's frontier is this year's commodity has been the case for the last few years. Usually there's kind of a six-month lag, give or take, between what is at the very frontier and what's at the edge.

But with the massive growth we saw in how frontier models and the value they generated at the start of this year, it's not surprising that some of the open models are now hitting the point we were at the start of the year and they're really useful. And so as you mentioned, there were some big releases out of China from Alibaba with Aquen models, Kimmy K3, ZAIs, GLM 5.2, which were all very, very capable models and can be used as an alternative to using these expensive closed end models. And that benefits that who does that benefit? Who does that hurt if the world switches towards closed models? Is it bearish kind of AI infrastructure as a whole? It's clearly kind of a negative. If you're a frontier model maker and you want to charge that premium margin, it's more competition, which net of everything, all else equal, that's going to hurt your business model.

But if you're a consumer of tokens or if you're an inferencing kind of host, you host inference, you're an inference provider as the hyperscalers are, that can benefit you a lot more because you don't need to share as much of the profits with an anthropic or an OpenAI. So there was this big kind of open weight letter that went out with most of the VC industry, Nvidia backing, a lot of hyperscalers in support of open weights models. And I think it's mostly around who's capturing the value really, but it's not necessarily negative for the AI CapEx and the infrastructure. A token's a token. It doesn't necessarily matter which one you serve. If you're dividing the pie, who's capturing that value?

Moz Afzal:

Yeah, I think that's going to be the debate. The reality is for most, and the cost per token for a frontier lab token is much more expensive than a open weight one. And obviously when it comes to budgets, CFOs are certainly putting their foot down on individuals and departments even to say, hang on a second, don't use this one, use this one because it's 70% cheaper, which is what's driving through. It'll still use the same AI compute, if you like, infrastructure, but it will be just a lot cheaper.

Henry Walters:

Yeah. And I think that they'll be used in tandem. I don't think it's a either or. I think they'll continue to grow together.

Moz Afzal:

So I think that feels like the direction's going. I suspect the reason why investors are focused so much on the AI infrastructure semis and so on and so forth, it's an easier decision to make than it is who's going to be the winner in AI. That's a really hard, really difficult decision for investors to make.

Henry Walters:

Yeah, in terms of who's going to build the applications to capitalise on this. There's a lot of competition. You've got the incumbent kind of SAS vendors who've all pivoted to this and they definitely benefit from cheaper LLMs and open weight models. But then you've got the labs don't just want to sell tokens, they want to sell products and applications on top of that. So they're kind of meeting in the middle. And then additionally, you have every hyperscaler selling additional services and custom models for every enterprise.

Moz Afzal:

I think what's interesting is, again, trying to draw analogies here for our listeners is the, I call it the Netscape and Yahoo argument back in the early mid - 90s. And that's a really hardcore. Sitting in late '90s or mid '90s, no one had an idea who would be winning. And at the time everyone thought that Netscape would be the winner.

Henry Walters:

Yeah. And things shift very, very quickly. So if you're in the lead now, it's not a given that you retain that.

Moz Afzal:

Exactly. Exactly. Yeah. And I think that's also, let's maybe sort of finish off in terms of the risks that this all brings about. Because the risks are actually quite a lot. First of all, just technology risk. We just don't know. And anyone who pretends to know is probably lying.

Henry Walters:

Yeah. And even you don't know what the next model's going to be capable of. Just the technology, it's still very nascent. If you talk to anyone at the AI labs and they believe that the digital deity and super intelligence is coming, it's a bit hard to get your head around what does that mean in terms of products?

Moz Afzal:

Yeah, exactly. So I think demand is there, but let's talk about some of the negatives that are developing other than the technology risk. We talked about credit spreads and the rising cost of capital.

That obviously it won't kill AI demand because it's there and it'll stay forever probably, but it could slow it down, which is the challenge or certain companies are not able to keep up with the cost of capital and just fold. That's I'd call the macro risk. Again, we need to watch out for very carefully. The other risks are geopolitical. We obviously have a China-US AI race, and that's clear.

Henry Walters:

But yeah, there are two different kind of technological spheres. We talked about China in terms of the context of some of the open models, but they are clearly building their own self-dependency in terms of chips and infrastructure and are aggressively trying to catch up. And then the world, US and China still remain heavily interdependent upon one another. So any kind of rising tensions there can easily put a spanner in the works in terms of the supply chains required to build all this. The other kind of policy risk, or I'd say regulatory risk you see around AI is just the mass broad public negative sentiment towards AI. There's been a lot of surveys done and it's wildly unpopular amongst the -

Moz Afzal:

Well, I think the analogy, and I think unfortunately the AI leaders haven't been particularly good at communication. In fact, they've been awful at communication. And I think the way to put it is if you tell a consumer that, by the way, building a data centre to increase your electricity costs, your water cost bills are all going to go up, your electricity bill's going to go, and I'm going to take your job away. It's not exactly something people are going to vote for. And I think that risk is clearly increasing with moratoriums now already in New York, Texas even, starting to see some negatives.

Henry Walters:

Yeah, no, clearly. Yeah. Whether it's electricity prices or just as you mentioned, the heads of these companies are telling you it's going to end work as we know it. That's fair enough. It causes a bit of uncertainty across the labour market. And the final risk, I'd say, we talked about technological risk, but it's just adoption typically isn't smooth. There is a chance we see a slowdown in terms of enterprise AI adoption or willingness to spend in ever increasing AI budgets. So far this year, it's only surprise to the upside, but you can never rule that out as with any kind of technological adoption. It's rarely, rarely such a smooth straight line upwards. Yeah.

Moz Afzal:

I guess the last risk would be cyber risk. And we've seen a lot of challenges where OpenAI software has sort of gone amuck and cyber attacked different platforms, hugging face in this case, or the one that's gained most of the attention. What are your thoughts around that? And does that then force governments to really slow the adoption down because they just don't know what they can control?

Henry Walters:

Yeah, it's a great question. And the cyber risk is clearly very, very real. And the leading AI labs is again, the closed source. I think Anthropic has Project Glasswing where their frontier models, they'll only test with trusted partners, including the government. And yeah, they clearly are very, very worried about cyber risk. You had the OpenAI incidents with Hugging Face. And then you had Anthropic, I think Anthropic had another incident relatively recently where your models can just kind of go rogue and very, very diligently perform their task, but maybe in ways you didn't really intend. And yeah, it's an open question about how you solve that because it may not be a problem if it's a closed model that you can just turn off at one provider. But if it's an open model, which there's zero reason to believe they won't be capable of these kind of threats, then it's much more easily used by bad actors around the world. So the best defence people have come up with is to use more AI to defend against AI.

Moz Afzal:

Henry, thanks for taking us through all of that. I think we covered a lot of ground. I suspect we'll probably find many of us are going to have to re-listen, should I say, or reread even this podcast. So thank you very much. And given how fast moving this is, no doubt, we'll have you again on very soon as well. So thank you very much for taking that. I appreciate it.

Henry Walters:

Thanks for having me.

Moz Afzal:

Thanks. So with that, we'll wrap up. Thank you very much for listening. If you've got any questions or any thoughts, please reach out to us. In the meantime, have a great day.

This podcast is provided for general information only and assumes a certain level of knowledge of financial markets. It is provided for informational purposes only and should not be considered as an offer, investment recommendation, or solicitation to deal in any of the investments or products mentioned herein and does not constitute investment research. The views in this podcast are those of the contributors at the time of publication and do not necessarily reflect those of EFG International or New Capital. The companies discussed in this podcast have been selected for illustrative purposes only or to demonstrate our investment management style and not as an investment recommendation or indication of their future performance. The value of investments and the income from them can go down as well as up, and investors may get back less than the amount invested. Past performance is not a guide to future returns, return projections, or estimates, and provides no guarantee of future results.