In Today’s Note:

• We split the AI landscape into three groups: infrastructure builders, compute scalers, and AI application developers
• The infrastructure builders have been remarkably profitable this year, funded by unprecedented capital spending from the compute scalers
• The key source of vulnerability in the AI trade is the concerning revenue generation from application developers, who ultimately need to pay for the spending done at the building and scaling layers
• Our view is that AI valuations are completely detached from reality, even if we assume explosive revenue growth in the sector
• AI exuberance and over-inflated valuations are a key risk to investors who own broad indices that are heavily concentrated in this sector

Many of the biggest winners in this year’s gravity-defying stock market rally have been part of the much-hyped AI trade. Our clients know we think the expectations built into the valuations of these companies are unrealistic and hiding a great deal of downside risk. But even if we ignore those valuations, there’s a much more concerning underlying economic reality to contend with: for the time being, nobody’s really making any money off of AI.

This is a peculiar thing to say on a day after Nvidia posted yet another set of strong quarterly earnings. However, it’s important to delineate between three different camps in the AI world: 1) Builders: companies helping build out datacenter infrastructure (Nvidia, AMD, Broadcom), 2) Hyperscalers: companies building datacenters to provide the cloud compute needed to train AI models (Meta, Google, Amazon, Microsoft), and 3) Developers: companies buying compute from the hyperscalers to train AI models and finally sell the product to end users (OpenAI, Anthropic, Grok). There is plenty of inter-mixing between groups, particularly between the Hyperscalers and Developers.

The Builders group has been extraordinarily profitable so far, funded almost entirely by unprecedented capital spending from the Hyperscalers. Everyone who wants to build out AI training capacity needs to come to a very small group of companies capable of providing the right type of compute and datacenter capacity, so margins in this business are enormous. But it also means that the primary risk to this group is a reduction in spending from the Hyperscalers (who in turn at risk from a lack of profitability from the Developers).

Hyperscalers have also been faring well, but in many ways it’s a mirage, because that capital spending is not yet putting downward pressure on profit. Even though this group is spending hundreds of billions annually, it is not hurting their profit or margins, because the spending is classified as capital expenditures (capex) rather than operating expenses. However, that capex will be amortized as depreciation over time, so the expenses (from an accounting viewpoint) are on their way, even if they’re not here yet. But when the Hyperscalers start seeing tens of billions in depreciation expenses on their income statements, they’re also going to see a huge hit to earnings unless those expenses are more than offset by enormous sales growth. That sales growth needs to come from the Developers buying compute from the Hyperscalers.

The Developers have of course been receiving plenty of attention from the media, and have created products (ChatGPT, Perplexity, Sora) that are being used on a daily basis by hundreds of millions of users. That the Developers are generating massive losses in these early days is to be expected, but what’s most concerning is that their sales figures are so small, despite the massive usage from end users.

If the stratospheric valuations built into the entire AI trade are to ever be validated, the Developers need to start generating more revenue. And when I say more revenue, I mean much, much more revenue – at least 10x current levels, just to start. So while the AI trade appears incredibly profitable right now, the money is being made on more of an “inter-department” sense, with cash flowing from the Hyperscalers to the Builders and from the Developers to the Hyperscalers, but without any meaningful source of external revenue from end users (it’s virtually impossible to pay for $100B+ in annual spending with $30/month subscriptions).

To put this in more concrete terms, imagine a company looking for investors to buy the compute and infrastructure necessary to train a new AI model. We ran some numbers showing that even using aggressive assumptions – interest only loans, a low interest rate, and zero additional operating expenses – the loan wouldn’t come close to being repaid even if revenue quadrupled in 10 years. Now throw in the extremely steep depreciation of the asset itself (chips become obsolete much faster than other assets, even when they’re not working 24/7), and the economics become untenable unless we assume 10x growth or more. Anything is possible, but we would happily pass on this investment from both an equity and credit viewpoint. If you’d like to see a more detailed analysis of our hypothetical investment, send an email to ryan.harder@rbc.com and I’d be happy to include it with additional commentary.

So the main issue is that the entire system eventually relies on Developers becoming immensely profitable, even if this year’s excellent earnings growth makes it seem like everything is fine. For the time being the Builders and Hyperscalers can tell themselves that massive new revenue streams are on the way for Developers, but time is not on their side. After all, as Benjamin Graham said, stock markets in the short run are voting machines, but in the long run they’re weighing machines. As time goes on, the weight of those miniscule revenues from the Developers are going to start entering the calculus for markets. Given the market action we’ve seen in the past two months, it’s clear an increasing number of investors are starting to do the math, and are concerned with what they see, just as we are.