That's all that's going on with compute agreements for Anthropic and OpenAI. In Anthropic's case, their risk is that revenues are insufficient to cover the minimum payments...which are never more than $12bn/yr as-of June 30. They did $16bn of revenue in the first half. So.

Show Post

Either way, next year's costs aren't expected to be paid out of this year's cash...they're matched to next year's revenue. Flows follow each other! The stock (cash) of prior flows is there (in part) to hedge against mismatches.

Show Post

Costs defined under long-term contract (so for instance, committing to spend X on widgets from a key supplier for Y years to come) are a trade-off of risks. You're guaranteeing supply (or at least some fraction of the total) in exchange for the risk you might buy too much.

Show Post

Let's imagine a business that has one cost (labor) negotiated via long-term contracts with a union. Should the business match that contract's size to 1) the cash it has right now or 2) the revenue it expects to take in over the life of the contract Answer is *obviously* 2, right?

Show Post

There seems to be a widespread perception that the foundation labs' compute commitments have to be covered by either cash they have on hand right now. It's come up so many times now, I'm just going to do a thread explaining why this is wildly inaccurate.

Show Post