“I call it earnings before bad things. So if I cut out all of the things that are costing me a lot of money and then I calculate EBBT — not EBIT, but EBBT, earnings before bad things — my cash flow is tremendous.”
Better Offline · Tech Strategy & Big Tech · September 2026
Kedrosky and Ed Zitron were arguing about whether AI inference is actually profitable for the hyperscalers, against claims that pure inference margins are very high. His point is that those margins only look good once you strip out frontier-model training spend. His test for the argument: show him a hyperscaler that has stopped spending on frontier models.
Transcript
rise, but on that order of things. That's right.
And that's what I'm saying. Which people will say is, oh,
that's really straightforward. You know, Gavin Baker's out there telling you that, right? That the margins on these things in terms of pure inference are so high. But of course, the problem is that this is the classic phenomenon of I call it like earnings before bad things. So if I cut out all of the things that are costing me a lot of money and then I calculate EBBT, not EBIT, but EBBT, earnings before bad things, my cash flow is tremendous. Well, then fine. Show me the hyperscalers that have stopped spending on frontier models and I'll buy the argument that the economics now make sense purely in terms of commodity inference. But if you go to the world of saying, okay, fine, industrial inference is actually a relatively high cash flow business. Let's just grant them that. Then the question becomes, well, who's going to win that? Well, if it's just industrial inference, so then meaning that I'm just doing token completion sequences as low cost as possible and at the highest scale possible, that's really just an energy problem. So it's just who can throw the most energy at this. China's throwing on an annual basis roughly three times, adding three times as much capacity per year as the United States is and is already ahead of the U.S. in terms of the amount of energy it's thrown at this. So it's fairly straightforward. You're just looking at solar panels, but it happens to be tokens.
So, but that's the thing. Even with standing up inference, because is inference profitable? It's like the Wario is a libertarian conversation online when it comes to this stuff. It's like, and no one can really, I think it's unprofitable. But the reason I do is because inference isn't just, oh, I turned on the inference machine. It's you buy an allocation of GPUs and you need a certain level of saturation and a certain amount of customers to make it viable. And if you miss that demand calculation, it's horribly unprofitable.
Oh, absolutely. Which a classic scale problem, right? That either I'm, it's kind of like running an airline, right? Either I'm a very, very high fixed cost business. So either I'm fantastically profitable or I'm aft, right? And And if you and very few people are able to run it at a scale requisite to be able to generate those kinds of cash flows from covering the high fixed costs. And so this is the deep problem. And then if you take it the next step and say, if that is the nature of what's going on, and tokens are this globally fungible commodity. And so that's the market's moving towards industrial inference. It's fairly straightforward to see who the winners are going to be. And it ain't going to be OpenAI and Anthropic.
Yeah. It just, it feels so inherently doomed. But now, you know what? You said Anthropic. And before we got on, you were mentioning you'd heard some strange things. And I am now curious about how strange those are.