Their inference margins are already high, the thing bringing their net down isn't training per se, it's the fact that they have to train the next iteration right away due to competition, otherwise an extended inference cycle would've eventually recouped training costs.

The bet is that cost per intelligence will keep going down and demand for intelligence will grow at the same time, if this bet holds, a user could get the same or more done with a cheaper model (running-cost wise) while paying the same subscription and the lab gets improved margins.

Open source is a challenge but their current moat is the fact that they're ahead, and as long as they stay ahead they'll be able to charge a premium.

▲

Betelbuddy 18 minutes ago | parent | next [-]

>> Their inference margins are already high

So they make a profit on each burger they sell you, but however they keep limiting how many they are willing to sell you?

	▲	dofm 14 minutes ago \| parent \| next [-]
		Only in this analogy it also takes years and lots of money to train the people flipping them, and they would like their investors not to notice that.
	▲	binlog 8 minutes ago \| parent \| prev [-]
		How are they limiting it?

▲

dofm 31 minutes ago | parent | prev | next [-]

> it's the fact that they have to train the next iteration right away due to competition

That it would be profitable if only there were no competition is not a really strong starting point for an IPO, though, is it?

(It is, however, the winner-takes-all world view Anthropic still seems to operate with)

	▲	vidarh 19 minutes ago \| parent [-]
		Their competition has the same problem. They're engaged in a giant game of chicken (or giant game of "who runs out of ability to raise more cheap capital first") with respect to overspending on training relative to inference revenue. So, yes, it's high risk - you're basically betting on who will be in the strongest position when one or more of the competitors are forced to slow spending and the rest are able to ease off on theirs as well as a consequence. If you think either that Anthropic is in a strong enough position to come out on top or that we'll end up with a detente like with the big cloud providers where they big providers feel they have a strong enough position to beat off upstarts and collectively basically stop competing other than on the margins because they all know they'll earn more by protecting their margins than by internal competition, then the IPO might be worth putting money on. I think the latter scenario is slightly lower risk, but still risky given how small the moat appears to be vs. the Chinese labs (I say this as I just had Kimi re-design a website in 10 minutes by just giving it the URL to the current one and tell it to replicate everything; they might not be as smart as the frontier models yet, but they're capable enough that they're taking over an increasing proportion of my token budget)

▲

cmiles8 39 minutes ago | parent | prev [-]

The rush towards open weight models defeats this thesis. There’s no prize for being “ahead.” That’s academically interesting but a terrible business plan. 99% of AI use cases don’t need the latest and greatest model, and businesses increasingly understand this.

	▲	ACCount39 26 minutes ago \| parent [-]
		That's like saying "forklifts are a terrible idea, because a forklift is excessive for lifting 99% of the items that I can lift". See the problem? "AI use cases" is a moving target! If you try to define "use cases" by "what AIs today can be used to do", you miss things AIs would do in the future. The more capable and price-performant AIs get, the more there are places where AI has a role. This is a big part of what the frontier buys you. The other big part is that going from a local AI to a frontier model often makes the difference between "this use case works 60% of the time, good tech demo" and "this use case works 99% of the time, good business decision". The difference between "clunky and needs human in the loop" and "just works" is worth paying the API price.