[00:00] Brad

Welcome to The Flip Side.

I'm Brad Rogoff, Global Head of Research, and joining me today is Tom O'Malley, semiconductor and semi cap equipment analyst here for us in the US. 

Thank you for being here, Tom. 

[00:12] Tom 
Great to be here, Brad. Thank you for having me.

[00:13] Brad

Before we dive in, just a reminder to everyone. Don't forget to hit subscribe, so you never miss an episode of The Flip Side.

So, Tom, I mean, I could really could have had you on here any of these episodes to talk about AI, but the reason I have you here today is because you and Ross Sandler recently put out what I call a rather controversial note talking about the peak for AI CapEx spending. So, your framework has AI CapEx peaking around 2028 at roughly $1 trillion, driven by recursive self-improvement, or RSI, which I had to use AI to look up exactly what that was, and lab level demand. So that's close to $300 billion above current consensus in 2028, which was already being questioned by the market. 

[00:56] Tom 

That's right. But it's not so hard to get there if you take a different approach and focus on demand for compute from AI companies, more so than just using hyperscale estimates.

[01:04] Brad

I'm all for different approaches, and look, all of these AI CapEx numbers are incredibly high. But my concern is, despite your trillion peak in 2028 being grounded in demand, it might underestimate what I’ll call real world constraints. So, when you frame the growth opportunity, what makes you confident that this demand is real?

[01:25] Tom 

Look, we're still in the early days of AI, and the upside here can be enormous – our model suggests $230 billion above street for CapEx in 2027, and close to $300 billion above calendar year ‘28 estimates. The Street has been consistently underestimating the CapEx forecast since the AI started in late 2022 and today is no different. 

In the past, we followed a similar bottoms-up analysis to most on the Street, making assumptions on the models and GPU utilization, but it was messy. So instead, we think it makes sense to use the AI labs OpEx and translate this into hyperscaler CapEx, since their facilities are what support the labs. 

[01:59] Brad

Hyperscalers are clearly willing to spend, but don't we have to assume some financial discipline from them? I mean, I could argue you're just extrapolating AI lab ambition and calling it a forecast. 

[02:09] Tom 

It's a valid question, but these aren't just lab ambitions. The demand is real, and it's starting to show up in the data, causing an arms race for the best solution. We're just turning the corner of  Agentic AI, as you can see on the billboards outside, which requires better models and massive ramp in inference spend – and you're starting to see that with Claud Code, OpenClaw and others.

On the training side, model builders are gearing up for recursive self-improvement, where models can rewire their architecture, improve training methods, and continuously update to provide better products. We think this dynamic will continue to drive significant spend, with the peak in 2028 to support that RSI. 

[02:41] Brad

So, a peak in 2028 could certainly occur, but it's not as meaningful as you know, even if it's a really large number. If we get a drop off, that feels like a cliff right after that. How are you thinking about the ability to sustain this kind of CapEx? 

[02:55] Tom 

Cliffs are scary words, and I'm not saying this is a collapse. It's more of a stabilization. With CapEx down about 5% year over year in calendar year 2029.

[03:04] Brad

All right. So, I want to come back to this RSI concept in a bit, but before we do that, let's spend a little bit of time on the AI labs themselves. So, the ramp and scale of these investments it's unprecedented, right. And given OpenAI and Anthropic are private companies - pretty limited information, you, as a research analyst, got to do a lot of digging. What's the potential for their forecasts to oversell their ambitions?

[03:27] Tom 

So, for OpenAI, the forecasts were recently revised dramatically higher in a five-month period from September ‘25 to February ‘26, and that points to real demand. And to support continued growth,  OpenAI and and Anthropic may look to tap the public equity markets for capital. Nvidia's CEO himself, Jensen, has talked about wishing these companies were public so that everyone could see the ramp signaling in their financials.

[03:48] Brad

I hear that and like to base everything in numbers, but it's also worth keeping incentives in mind. Just like Nvidia, these private AI labs have incentives to make big promises. Something we've heard from Big Tech by chance before. Let's think about the metaverse, an obvious example that never came to pass. So, I'm reticent to treat these lab forecasts as binding signals.

[04:10] Tom 

Those are some tough examples. I would push back, though, that this isn't just companies talking about their book. You're seeing these forecasts that are reflected in real spending decisions by hyperscalers, suppliers and even governments. That kind of broad, cross-industry commitment lends real credence to the build out. 

[04:26] Brad

Still, you're asking hyperscalers to underwrite not just improvements in models but participate in an arms race to fund agent adoption before we even see evidence of the killer AI application. I get the argument we're already seeing efficiencies in spend and some cost savings and businesses. To your point, you know, we try and make sure we have those cost savings too, but for $1 trillion in CapEx, you need a much bigger leap in AI ROI. 

[04:53] Tom 

I think that's a totally fair point. And these assumptions, there's some level of product improvement and market adoption that you need to bake it in. But with that being said, you already have seen a massive spike with the chatbots. For the next leg, you can take a look at China and see the popularity of OpenClaw and the feedback from the industry is that we're rounding the corner on something big. This is one of the ways that hyperscalers justify the spend today to support the labs.

[05:15] Brad

So, I would argue that some of these innovations are already reflected in their spending plans. But perhaps more importantly, if they wanted to spend on this scale, how are they going to fund these builds when some of the markets, especially some of the credit markets that they've used in the recent past, maybe aren't as open as they have been?

[05:35] Tom 

From my conversations, funding these builds is the most common pushback we hear from investors on the Nvidia bull case. In our model, CapEx can take up 85% to 90% of hyperscaler operating cash flow between 2026 to 2028. 

[05:49] Brad

85 to 90%. I mean, that's a huge commitment. It also doesn't leave much room for error.

[05:53] Tom 

Upfront it is. We would argue that as agentic adoption drives real ROI this year, that can soften up some of the concerns on hyperscaler spending and justify the higher CapEx.

Outside of hyperscalers. This year at GTC, Nvidia's GPU technology conference, we learned that 40% of demand comes from outside hyperscalers, in sovereigns and in enterprises. We should see further diversification and strength outside of the largest spenders as AI ramps.

[06:18] Brad

But we've also seen some high profile stumbles recently. And I get this happens with something that’s a new technology, but OpenAI walked back an MOU with Nvidia for 10 GW and the Stargate data center project in Texas that's had some turbulence, too. So, to me, that highlights that regardless of their desires, in practice we may need to take a haircut to the AI lab estimates. 

[06:41] Tom 

Definitely some setbacks, but our view is that these are normal hiccups when investing in large scale datacenter build outs. These projects required large scale coordination across construction, infrastructure, energy, and compute. While you could take a discount to the underlying numbers, I would argue we already are conservative and maybe even a little overly conservative in our framework. 

[06:59] Brad

Wait a minute. That's saying a lot, because we talked about earlier; you guys are above consensus with your estimates. So, walk me through how that could be the case that you're conservative.

[07:11] Tom 

Sure, so three key considerations here.

First, our framework assumes new inference demand is entirely served by old training chips starting in 2027, However, you will see incremental spend for inference specific silicon from Nvidia for example, the recent Groq LPU announced at GTC and others as in hyperscale ASICS, for example.

Second, we assume training peaks in 2028, but this can be revised higher as models grow and become more intense for agentic use cases.

[07:36] Brad

Okay, so maybe everyone wants the new and shiny chips. Based on some of the financing of these projects, I think there is a real hope that these chips have some life to them. Your second point, I guess let's just call that the AI productivity bull case. So, what's your third piece? 

[07:51] Tom

So, third, our model only uses Blackwell chips, the current generation from Nvidia, rather than future Nvidia generations, which may undercount total CapEx due to ASPs. If companies follow the typical upgrade cycle, spending likely runs higher than our numbers. 

These factors, among others, may cancel each other to some degree, but we think more items swing the pendulum to the right rather than to the left. 

[08:10] Brad

What about custom silicon efforts from Nvidia's competitors? So custom chips are generally lower cost, lower power, and can be optimized for model architecture. Wouldn’t these chips lower overall investment needs?

[08:23] Tom

The short answer is yeah; more custom chips would require lower spending levels. I would just point out, it's difficult to tailor a chip to a specific workload, as these chips are intended to work on one specific vertical and are designed years in advance, and they constantly change in terms of workloads that they're required to be built for. So, we have only seen two instances of these chips that have been built and actually scale in high volume production in massive scale. 

Flexibility is very important when designing AI hardware, and we just don't know where the ball is moving over the next 12 to 18 months.

[08:54] Brad

Okay, you're the expert. I get your skeptical but just play along with me for a minute here. If we're using your framework, would custom chips have a larger impact in training or an inference?

[09:04] Tom

It would impact both. I know medium answer here, but inference is definitely a bigger opportunity for these chips longer term. In inferencing, we would argue we see more customized designs, and hyperscalers have a good understanding of the workload they want to address and can save on cost. We touched on it earlier – inferencing demand is also undercounted in this framework, so the CapEx numbers can be a bit biased higher here.

[09:24] Brad

So, let's shift gears; model architectures are constantly changing and evolving. You talked about some of the major pushbacks that you guys have gotten. The one I keep hearing is, could we have another DeepSeek moment where compute can be done cheaper, and we wouldn't need as many chips to get to the same result. Your operating framework, it just doesn't seem to account for any chance of thi

[09:45] Tom

This relates to the idea of Jevon’s Paradox. So, when technological improvements make a resource more efficient and cheaper, it leads to increased usage, not less usage. So, we saw this with DeepSeek. The breakthrough led to further adoption of AI resources and spending has only trended higher since the DeepSeek moment. I, for one, have received more access to AI tools here at Barclays. 

[10:04] Brad

You're welcome!

[10:06] Tom

Well, thank you. My point here is that we are in the early stages of an adoption curve, and model breakthroughs may help instead of hurt. We are hearing from the industry that we are in a compute shortage so any resource that effectively makes compute more efficient or increases resource utilization, still likely leads to needing more compute. 

There is also an untapped market with humanoid robotics, and our FICC thematic investing research team has actually written about it. All these robots need to be trained on large quantities of data, and sometimes synthetic data. You may see this become a larger portion of the mix as we move through the end of the decade. 

[10:37] Brad

Let's shift to whether the build that's actually even feasible. So, you have to consider power constraints, permitting challenges and labor shortages, which present real risks. Hyperscalers may pay their share, but in the US, it takes about two years to develop a data center, five plus years to source and commission a new large gas power plant, and over ten years to permit and develop new transmission lines.

[11:02] Tom

We agree that power is a gating factor, and it's our biggest area of concern even when looking at semiconductor shortages across components and memory and lasers and even manufacturing. This is top of mind for the industry, and we are seeing investments to either make derivative products consume less power or for data centers to use power more efficiently. 

In late March, Nvidia put out a press release about working with leading AI energy companies to make AI factories more flexible grid assets, so you are seeing real efforts to conserve or increase efficiency of power outside of those to build new power lines.

[11:34] Brad

All right. So, it sounds like power is a legitimate bottleneck today. But let's do a little bit of math here. The AI wave really took off in, let’s call it, 2022, so transmission lines started then wouldn't be ready until next decade. I'd also assume most of the demand explosion happened more recently.

[11:55] Tom

The industry is telling you that they can do this build. They are conscious of the constraining factors. Just to ground this in numbers, we expect 19GW of new supply can be added for US data centers from the ISOs in calendar year ‘27, which doesn't include behind the meter applications or global supply additions. Our work with Ross suggests that the industry needs 13GW of new capacity in 2027 and 21GW of new capacity in 2028. So power is tight, but it's not prohibitive. We have confidence that we can get to the $1T in 2028 CapEx in this power environment.

[12:27] Brad

Even if we hit your numbers, Tom, I think there are still other hurdles. So, we're hearing more about how to view AI development when it raises regional utility bills and raises concerns on water use and then there's just general environmental risk.

[12:41] Tom

I agree with you and this something that scares me a lot, and we definitely watch out for and in the industry is definitely considering this as well. Hyperscalers are investing in solutions to get behind the meter power that won't stress the grid. To your earlier point, this will take time to develop, but there are a wide range of other measures companies are looking at to get access to power, which includes small modular reactors or even sending chips to space to use solar. Some of these are more out of the box ideas, I understand, but again, power is not taking the back seat here.

[13:11] Brad

Coming back to the point on RSI, as I promised, while the technology's interesting, we have no guarantee that it will come to market or work the way AI researchers are expecting. So even for cash rich hyperscalers betting $1 trillion in CapEx on it, it could lead to a repeat of prior tech overbuild. 

[13:30] Tom

The difference with this cycle is the quality of the spender is much higher. We're seeing hundreds of billions in investment from well-established companies, and they can finance the investment with operating cash flow instead of debt like in prior cycles.

[13:41] Brad

That's fair. But unlike OpenAI and Anthropic, the hyperscalers are public companies that still operate under shareholder constraints that limit how far they can go and really invest to support this cycle. New technology adoption is generally slow at the enterprise level, and we may see spending plateau or diminish before seeing proper ROI. How do we know when spending has gone too far?

[14:07] Tom

I think this view is already reflected in large cap Semis valuations. So Nvidia, for example, is trading at less than 14x calendar year ‘27 numbers and is significantly below the five-year median at 30 times. So, on one hand, we understand that this is the case. We understand the challenges, but the stock reflects this. And on the other hand, fundamentals and revenue estimates continue to improve.

The mistake is treating AI like a normal tech cycle, where we are seeing an arms race to have the best model and the best platform for AI adoption. The industry is shouting that they are compute constrained, and hyperscalers will need to fund this capacity. 

[14:40] Brad

So, we're still early in the days of the AI cycle, and there's no question that AI CapEx is accelerating – the opportunity: huge. 

But whether that translates into a $1 trillion peak by 2028 still hinges on some big questions, and those are power availability, infrastructure timelines and the ROI required to justify that level of spend. That's where my skepticism lies, and why I still think some conservatism is warranted.

[15:06] Tom

I am sure we will be revisiting these estimates plenty more times, and we'll watch closely to see how it plays out.

[15:13] Brad

Let's keep working on our own recursive self-improvement while you do those revisions. Thank you for joining me today, Tom. To learn more, clients of the investment bank can explore on our equity page and read more of Tom's views there. 

Thanks for listening and don't forget to subscribe so you never miss an episode. See you again on The Flip Side.