In partnership with

If you're underwriting a borrower and you can only use one number to do so, a pretty good one to choose is their net worth, ideally calculated by estimating their assets plus the net present value of their future income (less that part that comes from running down the value of those assets). For some corporate borrowers, that asset base is sizable, tangible and has a readily-determined value, but for most of them, the majority of their worth is the present value of future cash flows. Some companies can borrow at lower rates than implied by their credit ratings for this reason—the ratings are more backwards-looking, as is proper protocol for a lender concerned about downside, but if cash flows are expected to grow over time, then a given debt keeps getting easier to service or refinance.

Life would be easier if all distributions of company outcomes were the same, but, of course, they're not. Some businesses mean-revert, like cyclicals. Some grow for a while and then level off. Some grow for a while because they're repeatedly taking risks and getting lucky—not just investment businesses, but any company whose growth model is to exploit pricing power faster than they grow it; a dialysis business or a local monopoly has a return profile that looks like the house's returns from roulette: lots of small wins, and the occasional big loss when either regulators or competitors get serious.¹

There's a profile that fits a few different trillion-dollar companies right now, which is that they're losing money, growing fast, and offsetting increases in gross margin by adding more operating expenses. The big AI labs fit this description perfectly, and for companies that predate the AI boom but are leaning into it, a growing share of their market value is represented by this bet. If there were one profit-maximizing company doing this, they might eventually reach the point where they decided that the risk-adjusted return on training another model was not worth it, and then happily collect profits from an AI monopoly. But there are several of them, which means that at any given time, whoever has the best model also has a model that's rapidly depreciating and needs to be replaced by an even better one.

An equity investor can look at this situation and say that it's just a matter of bet sizing: a company that can go to zero has a fair value greater than zero if there's some chance that it doesn't, and the higher the odds of a wipeout, the lower the amount of risk to take on that specific name. A lender needs to be more cautious, because they can definitely find themselves in a situation where they lend the company money for five years, its valuation doubles annually for the first two, and then it goes to zero and they're wiped out.

But this doesn't make lending impossible. What it does, instead, is to move loans around in the supply chain and the capital stack. The ideal collateral is some long-term commitment by a company that has some core business that produces stable cash flows, but that also either uses or resells AI. In this case, a lender to the neocloud counterparty to that contract is basically underwriting a mix of how well the neocloud can execute, the residual value of its chips after the contract is done, and maybe a few basis points to cover the risk that the ultimate customer defaults.² (Big Tech borrows cheaply, but not at a price implying zero credit risk.)

As the labs mature, and as equity investors demand not just a hypothetical future path to immense profits but some signs that they are actually heading in that direction, the labs themselves will get more creditworthy. Strategically, OpenAI and Anthropic want to be bad credits, in the sense that they can raise more money from equity investors with a triple-or-nothing than they can from credit investors by getting free cash flow positive. But the market isn't infinitely patient, and as the funding labs need rises, they tap out the sources of capital looking for crazy risk and have to tap into the sources that are looking to add another large-cap growth name to their portfolio.

But by that time, the other problem will be that lenders are all stuck making the same bet. It's a problem they tend to have: whenever there's a capital-intensive buildout that's funded partly by credit, it makes credit more and more of a bet on that sector. Sometimes, this is isolated—there was a point in the 2010s when "distressed credit" basically meant "frackers"—but railroads at the turn of the 20th century were about three quarters of the non-financial corporate bond market, and utilities were 38% by 1929. More recently, telcos were 40% of high-yield bond issuance in 2000. This mix of dates and industries is the kind of thing that makes people nervous, though it's worth noting that the prior years would also have had similarly scary record-breaking numbers, and that it was a career-limiting move just to sell because of that. But it also illustrates that industries tend to use a lot of credit when they're growing, and to need less once they're big and well-established. A growing industry uses credit because a mixed capital structure is more attractive in the aggregate than pure equity, and more feasible than pure debt. But when a mature industry levers up, it's because they're converting higher confidence in future returns into cash they can use to buy back stock.

And this process is somewhat self-fulfilling. Credit managers no doubt get very nervous when every new stress-test of their portfolio attributes higher losses to an AI bust, especially if they designed those stress-tests some time in 2022 after losing money lending to software buyouts. So they'll make it relatively more expensive to grow pure AI businesses, but also relatively cheaper to grow the complements—if you're nervous that too much capital is pouring into an industry, then lending to that industry's customers is an even better idea; an AI bust means cheap inference for AI users.

This boom has produced plenty of n-of-1 financial structures, and probably will produce more. And it's incredibly rare for a massive investment boom coupled with novel financial structures not to lead to a disaster somewhere along the line. That's compatible with the services being useful—the Internet was a much more useful tool in 2002, when Internet valuations had been wiped out, than it was in 1998, when they were soaring. But markets are also somewhat self-correcting, in that the financeable parts of the AI ecosystem are the ones that generate recurring revenue with positive contribution margins, whereas the more speculative bits tend to be financed by equity investors who know what risks they're taking.

We've covered capital-intensive buildouts over the years in The Diff, with the current one getting plenty of attention. Some related pieces:

• A few weeks after the release of ChatGPT, we covered the question of whether it's an asset-light layer on other businesses, or a capital-intensive and potentially cyclical business. One of those definitely didn't happen, but the cycle question is still up in the air.

• Meta (disclosure: long) is a good example of a company that tamped down aggressive spending plans in the face of investor skepticism ($).

• More on neoclouds slicing up cash flows to find the leverable piece.

• Even if you're lending to one specific AI company, if the loan is backed by AI infrastructure, it's backed by industrywide demand ($).

Blu Dot surpasses 2,000% ROAS with self-serve CTV ads

Home furniture brand Blu Dot blew up on CTV with help from Roku Ads Manager. Here’s how:

After a test campaign reached 211,000 households and achieved 1,010% ROAS, the brand went all in to promote its annual sales event. It removed age and income constraints to expand reach and shifted budget to custom audiences and retargeting, where intent was strongest.

The results speak for themselves. As Blu Dot increased their investment by 10x, ROAS jumped to 2,308% and more page-view conversions surpassed 50,000.

“For CTV campaigns, Roku has been a top performer,” said Claire Folkestad, Paid Media Strategist, Blu Dot. “Comping to our other platforms, we have seen really strong ROAS… and highly efficient CPMs, lower than any other CTV partner we've worked with.”

Using Roku Ads Manager, the campaign moved from a pilot to a permanent performance engine for the brand.

Learn More

Join the discussion!

1  If this piece had been written a few decades earlier, the obvious example of a monopolist with roulette economics would be a local newspaper in a one-paper town, because it was a good place for local companies to advertise and the only place for classified ads. These companies create economic potential energy for their future competitors by creating a wide range of prices that can support a decent business and offer massive savings relative to the incumbent. When a company grows earnings through price hikes over long periods without this kind of drawdown, it implies that they've found a way to ensure that consumer surplus growth outpaces the increase in their take rate. But this works just as well with falling prices, as with chips (historically), batteries, or, more recently, AI.

2  And even neocloud execution is becoming somewhat derisked for lenders. For example, in CoreWeave’s—dislosure: long a little, for a little bit—newest delayed draw term loan structure (DDTL 5.5), the funds are not disbursed until 1) a powered shell has been secured and literally turned on (and a power purchase agreement has been signed for the length of the customer contract), 2) the hardware (GPUs, networking, etc.) have arrived, either at the powered shell, or at a nearby location, and 3) legal title to the hardware has transferred, or is set to transfer concurrently with funding. Of course, there’s still execution risk in that CoreWeave has to successfully install the hardware, but that’s easier to underwrite than hardware or power procurement risk.