The Two Companies Out-Earning Anthropic on Claude Are a Trading Firm and Meta

Anthropic turns a megawatt of computing power into roughly $100 million. Jane Street, buying Anthropic’s models with that same power behind them, turns it into $300 to $500 million. The customer is out-earning the supplier by three to five times on the supplier’s own product.

That figure comes from Dylan Patel, who runs the chip-industry research firm SemiAnalysis, speaking on the Dwarkesh Podcast on August 25. It matters to a marketing budget for a reason that isn’t obvious: Patel’s argument is that buyers of AI, not sellers of it, are where the money lands. If that’s right, your AI spend should be one of the better-returning lines in your budget. Whether it is depends entirely on something his two examples both had and most marketing teams don’t.

The claim

Patel’s framing, from the full episode captions at 18:25:

“Most of the value that these models generate does not get given to OpenAI and Anthropic. Thankfully, so far it is mostly just being given to the users.”

Then, at 20:16, he sorts the stack. The end user “is generating more value than anyone else.” The app layer, meaning the companies building products on top of the models, “has generated very little value.” The model layer went from negative gross margins to strong ones inside about a year.

So the money is not accumulating where the coverage says it is. It sits with whoever is holding the output at the end.

What his two examples have in common

Patel names two winners, and the pairing is the useful part.

The first is Jane Street, a trading firm with an exclusive OpenAI contract for a fast inference mode and a large Anthropic account. At 18:38 he says it’s “generating way, way, way more value out of the tokens they’re paying for than Anthropic is generating in terms of profit, because they get to make money off of the market.”

The second one belongs to us. At 18:55:

“Or take Meta, who at one point was rumored to be as much as 10% of Anthropic’s business. They’re generating way more efficiencies by optimizing their ad algorithms or what have you, getting engagement time 5% longer, all these things. They’re making way more money off of using these models than Anthropic is.”

Patel’s clearest example of a company beating an AI lab at AI economics is an advertising business. That should be encouraging to marketers, and it’s the opposite, once you look at what makes it work.

Jane Street has a market. Every marginal improvement in speed or pattern recognition converts into a trade, priced continuously, with no human step in between. Meta has an auction. A 5% lift in engagement time flows into inventory and pricing automatically, across billions of impressions.

Both had a mechanism that turns model output into revenue without anyone deciding to act on it. The AI didn’t create the machine. It fed one that already existed.

The question this leaves on your desk

Ask what your version of that machine is.

For most 5-to-30-person B2B software companies, the honest answer is that there isn’t one yet. AI output arrives as a draft, a brief, a list, a summary. A person reads it, judges it, edits it and then does something. That human step is where the value gets capped, because the throughput of the whole system is now the throughput of the person reviewing.

Keep using it. Just size the return correctly. Time saved is real, and it’s the smallest of the available prizes: hard to bank, and usually spent on more work rather than showing up anywhere a finance team can see it. We found the same pattern when teams couldn’t connect their AI costs to any return, and again when agents produced more briefs rather than fewer. OfficeChai’s writeup of the same episode reached the same split between end users and the companies selling to them.

The places a marketing team plausibly has a machine are narrower and more specific than the pitch decks suggest. Bid and budget allocation inside an ad account, where the output is a decision the platform executes. Lifecycle messaging, where a model choosing the next message ships it without a meeting. Pricing and packaging tests that run to a number. Those have the Jane Street shape: output converts on its own.

Blog production does not have that shape, however fast it gets.

Who is paying whom here

Two disclosures belong next to these numbers.

Jane Street sponsors the Dwarkesh Podcast, and there’s a Jane Street recruitment ad inside the same episode at 1:06:31. The host raises it himself at 21:46, after Patel uses the firm as an example for the third time:

“This is not an ad. This is not an ad. They’re a sponsor but you don’t have to plug them that hard.”

Read the exchange and it’s clearly a running joke rather than a placement, but the relationship is real and the example is repeated, so it goes here rather than in a footnote. Patel also sells research to the industry he’s describing, which is the standard caveat on any analyst talking about their own beat.

And these are estimates. Revenue per megawatt is Patel’s model, not a disclosed figure from Anthropic or Jane Street. Meta’s share of Anthropic’s business is described in his own words as “rumored.” Nobody has confirmed any of it, and the underlying claim about direction is more defensible than any single number in it.

The counter-case

Patel’s own argument contains the reason it may not hold.

At 1:14:42 he calls buyer-side value capture “the one positive thing here” and then takes it back. If Anthropic can generate hundreds of millions per megawatt using compute internally, he asks, why would it keep selling that compute to a trading firm at a third of the price? His words at 1:15:54: “Why would I let Jane Street make all this money off of these degenerate options traders?”

That’s a forecast of the buyer’s advantage closing. If the labs move up the stack and reserve capacity for their own products, the pricing that makes buyer-side returns look good today is the first thing to go. The economics you’re reading about are a snapshot of an unusual moment, not a stable feature of the market. That pressure is already visible in compute costs pushing into marketing budgets.

What to do with it this quarter

Take your three largest AI line items and mark each one with what its output touches before it becomes money.

If the answer is “a person, who then decides,” you’re buying capacity, and you should price it against a contractor rather than against a growth investment. If the answer is “a system that acts on it,” that’s the line to defend and expand, and it’s the one where a bigger model bill can be justified with a number rather than a feeling.

Most teams will find every item in the first category. That’s worth knowing before the next budget conversation, because it’s the difference between “AI made us faster” and a figure you can put next to the invoice.