johnswentworth's Shortform

johnswentworth

This is a special post for quick takes by johnswentworth. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.

Based on their prospectus, I think the chance of an Anthropic insolvency is surprisingly high. I'd put it somewhere between 20% and 50%.

I see three main ways it could happen:

Acceleration of the debt to Broadcom. That debt is approximately equal to Anthropic's current cash pile, so if it accelerates, the current cash pile is gone. And there's multiple different ways that debt could be accelerated - for instance, if Dario gets hit by a bus tomorrow, Broadcom probably has the option to accelerate.
Drop in compute prices. Approximately all of Anthropic's obligations are effectively compute futures: they've locked in future compute purchases at a certain price. If compute prices go down, they're locked in to a little under a trillion dollars of above-market contracts, and will therefore have a lot of trouble paying those contracts while keeping their products competitively priced.
Low demand and/or price for their products, leaving them with a lot of compute obligations and no way to pay for them. The main way this would happen is competition from open-source models.

Note that the latter two are bets which could turn into windfalls if prices go in the other direction. Also note that an IPO will probably bring in a lot more cash, assuming investors aren't scared off by the insolvency risk.

I would guess that OpenAI is doing similar things but even more extreme, so probably all the same considerations apply to OpenAI insolvency.

7Julian Bradshaw5h

It's possible that the reverse is true, and pacing improves financials: if pacing results in increased time between model releases, labs would have a better chance of recouping training costs via selling inference, as models would become obsolete more slowly. I'm not sure if this effect overcomes the "pacing means you can't make models good enough to unlock more demand for inference in time to pay for your compute commitments" effect though. (other considerations: competition from non-frontier/non-pacing labs; if pacing only redirects compute from training to non-remunerative alignment experiments instead of inference (unlikely?), that doesn't help financials; pacing might reduce pressure to compete on inference price)

It's possible that the reverse is true, and pacing improves financials: if pacing results in increased time between model releases, labs would have a better chance of recouping training costs via selling inference, as models would become obsolete more slowly.

Right but they're still paying for the compute commitments, so they spend nothing on training but a ton on compute that was meant for training. I think release timing is zero-sum, like the expected obsolescence of a lab's best model doesn't change with pacing? Except if some labs aren't pacing, then it... (read more)

2jmh9h

In addition to the time frame it would be interesting to know what probabilities are placed on the three drivers for the potential outcome. And I suppose also on the noted bit that two could become windfall returns that we might want to also factor into expectations. But clearly there are risks there.

johnswentworth's Shortform

11

Ω 5

johnswentworth's Shortform

11

Ω 5