CapEx Crunch: Are AI Infrastructure Bets Creating the Next Tech Bubble?

AI spending is exploding, but are the returns keeping pace? Mary MacNamara sits down with James Yendrey of InvestMentor^SM to explore hyperscaler CapEx, rising debt loads, shrinking free cash flow, GPU utilization concerns, and whether today’s AI infrastructure boom resembles the telecom buildout of the 1990s or the setup for a costly correction.

Summary

The following is a summary of a live audio recording and may contain errors in spelling or grammar. Although IBKR has edited for clarity no material changes have been made.

Mary MacNamara  

Hello everybody, and welcome to Cents of Security. Based on the current landscape, Q2 2026 earnings showing hyperscalers pushing combined 2026 CapEx guidance to roughly approximately 77% YOY, and Alphabet’s first ever negative free cash flow, a growing reliance on debt financing and persistent debates over GPU utilization versus real returns.

We thought it would be good to review CapEx. So, we’re here with James Yendrey of InvestMentor^SM to break this all down. Hi James, how are you?

James Yendrey

I’m doing well, Mary. Thanks for having me. How are you?

Mary MacNamara  

I’m doing great. Thanks for asking. Okay, so the first question, let’s just get this out of the way. Please simply define what is CapEx and why it’s important.

James Yendrey

So, CapEx is the money spent building long-term assets. Think like data centers chips, servers. These things aren’t necessarily expensed immediately. They’re depreciated over years. And it’s a company’s bet on the future demand.

Mary MacNamara  

So, Goldman Sachs says hyperscalers will finance more than a third of their 2027 CapEx with debt, and we just saw Alphabet post its first negative free cash flow since going public in 2004. Are we watching the AI build-out quietly transition from a self-funded cash flow story to a leveraged bet?

And what does that do to the risk profile if demand disappoints?

James Yendrey

Some impact there. So yeah, so yes, Alphabet did post its post negative free cash flow for the first time in, in 22 years and hyperscalers are basically on pace for roughly 33% in 2026 for CapEx, with debt rising towards 35% in 2027 against an estimated 1.14 trillion. And so, this is something that’s a bit different that you would assume to see from hyperscalers because cash-funded CapEx has a self-correcting function within it. So, if demand disappoint so if demand disappoints it, it is able to adjust accordingly. But if you have this over-levered debt, you have a fixed interest and principal payments that don’t necessarily care whether the GPU generates the revenue in the time horizon that was initially anticipated. So, it’s a shift from an equity-like risk to a leveraged risk.

Mary MacNamara  

So, is this strategic overbuilding for a demand curve everyone believes is coming, or are we building the AI equivalent of empty office towers?

James Yendrey

Yeah. I would say the data’s pretty, pretty stark in this regard. Average GPU utilization across, say, about 23,000 clusters sits at just 5%. So, the strategic overbuilding has a precedent. Think of the early ’90s when telecom laid fiber well ahead of its actual use case The issue in that same example, those fiber cables they don’t lose value. GPUs do lose value because you have a new one coming out every 12 to 18 months with better and faster features. And so, a cluster of 5% utilization, it’s not really waiting patiently. It’s a wasted asset. So that means the bull case needs to arrive faster and not eventually.

Mary MacNamara  

So, the CapEx to sales ratios are hitting levels analysts call seemingly untenable, north of 80% for some infrastructure-heavy players. At what point does the market stop rewarding growth at any cost and start punishing capital intensity the way it punished Meta in 2022? We don’t know, but we can speculate a little bit, right?

James Yendrey

Sure. That’s a good one. CapEx to revenue has moved from 10% to 15% this is like pre-AI, to about 30% now. So that is 2022 stop drop. The market had stopped funding CapEx once it lost faith in the payoff story. So right now, it’s not really about a specific ratio, it’s more about the credibility, right?

Investors had tolerated AWS’ build-out because the returns started to show up, right? Now they’re pushing these vague narratives and, a-and shrinking free cash flow, hence, buybacks, slowdowns across AI CapEx leaders. So, AI’s stock or excuse me, Alphabet stock reaction to its own CapEx raise suggests that we may already be nearing an influx point

Mary MacNamara  

So, is the real investable opportunity possibly in the picks and shovels layer rather than the AI platforms writing the checks?

James Yendrey

Chip prices rising for the first time in, in two decades reflect that. But it’s not inherently risk-free, right? It’s just a different type of risk. So, think suppliers have no demand independent of these hyperscalers right now. And so, the lack of diversification on the balance sheets to absorb that slowdown is inherently a risk in itself, and I guess these companies would trade risk for that concentration risk.

Mary MacNamara  

So Evercore and BofA now project combined big tech CapEx clearing one trillion by 2027, and some forecasts put annual hyperscaler spending at 2.5% to 3% of US GDP. Is this the new normal capital intensity of the AI era, like telecom building in the ’90s, or are we setting up for the most expensive correction in tech history? Scary!

James Yendrey

Do you want the honest answer? The honest answer is we really don’t know yet, right?

You can have a case for either side, right? That says the new base case for AI, right? The infrastructure could genuinely be foundational, like the electrification.

You have closed loop system or data centers now, so you’re advancing industry outside of it, right? Google’s Cloud I think it’s 460 billion backlogs suggests that there is real demand that exists. But on the other side, you have the correction case, right? Every super cycle like telecom or housing that we saw over the last couple of decades saw a real demand arrive, but only after a brutal repricing kind of wiped out the over-leveraged players.

So negative free cash flow, rising debt, and single utilizations look back at like a land grab as opposed to a more disciplined player in it. So, in my opinion, I don’t necessarily think it’s binary. The technology may justify higher capital intensity, higher capital and long-term investments, but it’s the pace of the financing and the debt funded. The debt funding with a low utilization rate it looks unsustainable, right? So, it’s Amazon versus Pets.com, right? The, the thesis was right, but investors still lost fortunes betting on the wrong balance sheet and timeline.

Mary MacNamara  

Okay, so let’s talk about this revenue coverage problem, which is pure play AI companies like OpenAI and Anthropic combined are generating maybe 35 billion in annual revenue against roughly 700 billion in hyperscaler infrastructure spend, call it 5% coverage. How long can that gap persist before either enterprise adoption has to dramatically accelerate or someone blinks on spending?

James Yendrey

So that number has actually moved quite a lot, between Anthropic and OpenAI, they now run combined at about $75-$76 billion, pushing the coverage closer to 10% rather than 5%, right? The core problem is concentration risk, right? So much of that revenue comes with massive losses.

OpenAI reported about $21 billion against a $13 billion revenue back in ’25. And then growing revenue doesn’t mean the economics work on it, right? The, the gap can persist while hyperscalers are also betting against their own cloud businesses. Eventually adoption has to scale or capital markets have to force discipline through higher borrowing costs.

Mary MacNamara  

All right. James Yendrey of InvestMentor, thank you so much for shedding light on CapEx. We really appreciate it, and thanks for joining us today!

James Yendrey

Yeah. Thanks, Mary. Great being here.