[reticulates]

Your contention is they're spending it on what, exactly? Leaks put OpenAI's training spend at single-digit billions so that can't be the machines they're building, and they (OpenAI + Anthropic) are famously renting/leasing/borrowing compute through varying-degrees-of-circular deals... so what's the machines they're building?

[twoodfin]

Just as the article supposes: Getting their hands on as much compute as possible to address rapidly growing demand for inference.

[reticulates]

but your contention is they are profitable on inference. Why would they need to raise money to get their hands on compute if they're making money on compute? There's no upfront costs for Anthropic, Anthropic don't own or build compute infrastructure, they just rent access to compute owned by someone else, such as the $15bn/year SpaceX deal they signed recently (which they used to create more demand without increasing revenue).

https://www.anthropic.com/news/higher-limits-spacex

[twoodfin]

The more capital Anthropic has, the more compute they can put their hands on and sell at a massive markup for Claude inference.

Limiting themselves to the profits from the compute they already have when capital markets are dying to give them cash at favorable terms would be foolish.

[reticulates]

You’re missing the point. Anthropic are bringing more compute online by renting it and then they are creating the demand for it by increasing the limits on fixed cost plans. Anthropic are increasing their spend on compute without increasing revenues. I’m not guessing that, it is part of their announcements!

https://www.anthropic.com/news/higher-limits-spacex

“First, we’re doubling Claude Code’s five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans.

Second, we’re removing the peak hours limit reduction on Claude Code for Pro and Max accounts.“

Every one of these plans is a fixed cost. Anthropic doubled their limits without changing the price. Even if inference is wildly profitable and these plans aren’t subsidized, they’ve just cut the profitability in half.

And the non-plan usage that is being paid for directly is paid for monthly. If they can sell $1 of compute as $10 of inference then they have $9 the next month to spend on more compute. Of course the capital markets would want to give them money if that were true but they would have no reason to take it.

Why would a net 30 business need to borrow hundreds of billions over many years? Anthropic currently spends an estimated $5bn/month on compute so at most they need to float $5bn, but if they’re making 90% margins on compute, that $5bn would be paid for… within a couple of days. Where is the hundreds of billions of dollars?

[twoodfin]

My understanding is that Anthropic’s massive revenue surge over the past 9 months has been driven by enterprise consumers paying by the token.

Do you think that’s wrong?

[reticulates]

I think that is correct, and exactly why these purported margins are nonsensical. If Anthropic's reported $50 billion in revenue is majority per-token billing, and tokens have a margin of 80%, that would put Anthropic's profit at $30 billion on per-token usage. Where is that $30 billion going?

And conversely, let's look at the amount Anthropic are spending on compute. Anthropic has just started paying SpaceX $1.25 billion per month for compute. At an 80% profit margin that would mean Anthropic is going to be bringing in $6.25 billion per month... that's more than their current reported revenue.

And that's just one contract for compute. We know that Anthropic also pay Google ~$3 billion per month for compute (based on their committed spend of $200bn over 5 years) which is $36 billion per year. At $36 billion per year on compute with 80% margins that would put revenue at... $180 billion.

Add in their spend with Amazon and Microsoft, Anthropic are spending at least $4 billion per month on compute, or $48 billion per year, all but equal to their revenue. If margins on tokens are 80% and an estimated $37.5 billion of revenue is per-token revenue, that needs just $7.5 billion of compute per year, less than $1 billion per month.

The numbers just don't add up. If margins are 80% and they have $48 billion per year in compute spend, revenue should be over $200 billion.

If the 80% margin made any sense whatsoever, Anthropic would be printing money, yet they're losing money, and have only been profitable for one month based on some financial engineering (pre-commitments billed after the fact to reduce their costs during one month).

My guess is margins are closer to 20% than 80%. That's the only way any of the numbers can make sense.