Leverage All the Way Down
In previous posts I’ve alluded to the market’s tendency to latch onto a trend and turn it into a craze, with the end result of a crash, sometimes contained (the .com boom + bust), sometimes not (2008). AI is now in the craze stage, but there are still a lot of high-profile players who want to dive in head-first. I have to admit I’m surprised at how bullish Jon Gray of Blackstone is: in his rebuttal to Ken Griffin’s claim that AI is almost certainly a bubble, Gray says AI demand is ‘almost parabolic,’ pointing to Ode, Blackstone’s AI consulting collaboration with Anthropic, as evidence. But this is consulting, and their clients themselves could be corporations that are caught up in the frenzy, which is not necessarily representative of the macro economy. And with the scale of Blackstone’s investments in AI infrastructure, Gray could perhaps be talking up his own book.
As we’ve seen with Accenture , when the token bills come in, the cost-benefit analysis becomes more objective, and then the punch bowl gets taken away. This isn’t some peripheral narrative: for the frontier AI labs like Anthropic and OpenAI the tokens are the revenue generator. If clients cut back on their token usage, Anthropic and OpenAI need to either lower their prices, or find markets where they can charge maximum prices. The growth in Chinese and open-weight models risks commoditizing token prices, forcing Anthropic and OpenAI to be price takers. The frontier labs are the bedrock of the AI industrial complex: if their revenue declines, then AI-related revenues of the neoclouds, hyperscalers and chip manufacturers decline as well, and the expectations of dramatic revenue growth disappear. You can check out a detailed website named The AI Loops outlining the interdependence of the players. Moving down the AI complex stack, if revenue isn’t going to outstrip the hyperscalers’ and neoclouds’ exorbitant capital expenditures and leasing costs, then those earnings models go out the window, and any predictions of share price growth are revised downwards. Furthermore, the inherent revenue-expenditure mismatch is likely to continue, with data centre construction costs remaining high, and leases to remain costly (no construction without a leasing commitment), as the value of any expected revenues drops.
As we have seen in the markets, these sentiments are surfacing, with companies like CoreWeave and Oracle getting kicked around in the bond market, as repayment of their massive debt loads become more doubtful. But bond investors aren’t the only players with concerns about leverage: global equity leveraged ETF assets are up 56% since the end of 2025. See the IMF’s analysis here, highlighting the growth of leveraged ETFs and leveraged single-stock ETFs.
So what happens if market expectations drop for these companies, and share prices aren’t parabolic? What if interest rates continue to rise and cost of funding these equity positions becomes untenable. Most likely, investors will have to sell. And this starts the less-than-virtuous circle: selling depresses prices, making leveraged positions more expensive to hold, and drawing fewer new buyers into the ETFs, further depressing prices for AI-related equities. Companies such as Google and Meta have huge cash inflows to absorb their capex increases and not see their share prices decline precipitously, but the objective of that capex is to drive future revenue and profitability. If the probability of those future revenues and profits being realized decreases, then the boom driving the markets will go bust.