Is that for per token costs or in these bundled subscriptions companies are selling?

For example, when playing around with claude code using a per token paid API key, it was going to cost ~$50aud a day with pretty general usage.

But their subscription plan is less than that per month. Them lowering their limits suggests that this wasn't proving profitable at the current limits.

Is this like saying a gym runs at 40%+ margin because 80% of users don't really use it heavily or forget they even had a subscription? Would be interested to see the breakdown of that number.

TL;DR: their subscriptions have an extra built-in margin closer to 70%, because the entry price, target audience and clever choice of rate limiting period, all keep utilization low.

----

In this case I'd imagine it's more of an assumption that almost all such subscriptions will have less than 33% utilization, and excepting few outliers, even the heaviest users won't exceed 60-70% utilization on average.

"33% or less" is, of course, people using a CC subscription only at work. Take an idealized case - using CC only during regular working hours, no overtime: then, even if you use it to the limit all the time, you only use it for 1⁄3 of the day (8h), and 5 days a week - the expected utilization in this scenario is 8/24 × 5/7 = 24%. So you're paying a $200 subscription, but actually getting at most $50 of usage out of it.

Now, throw in a rate limit that refreshes in periods of 5 hours - a value I believe was carefully chosen to achieve this very effect - and the maximum utilization possible (maxing out limit, waiting for refresh, and maxing out again, in under 8 hours wall-clock time), is still 10 hours equivalent, so 10/24 × 5/7 = 30%. If you just plan to use CC to the first limit and then take meetings for the rest of the day, your max utilization drops to 15%.

Of course people do overtime, use same subscription for personal stuff after work and on weekends, or just run a business, etc. -- but they also need to eat and sleep, so interactively, you'd still expect them to stay below 75% (83% if counting in 5-hour blocks) total utilization.

Sharing subscriptions doesn't affect these calculations much - two people maxing out a single subscription is, from the provider side, strictly not greater than two subscriptions with 50% utilization. The math will stop working once a significant fraction of users figure out non-interactive workflows, that run CC 24/7. We'll get there eventually, but we're not there yet. Until then, Anthropic is happy we're all paying $200/month but only getting $50 or less of service out of it.

That's how nearly every subscription service works, yes. Some fraction has a subscription and doesn't use it, another large chunk only uses a fraction of their usage limits, and a tiny fraction uses the service to it's full potential. Almost no subscription would be profitable if every customer used it to its full potential

Is that to the $20/mth plan or the 137?

They'd be still at massive losses. You can spend your monthly subscription price in a day.

Exactly. The enormous margin is why companies like OpenAI and Anthropic are known for being so immensely profitable. Just money printing machines compared to the amount of cash they burn

Cost of revenue should include R&D and amortization. Pointing to EBTIDA is a very old trick.

Definitely doesn't sound true unless you have the receipts.

I thought it was understood all the large vendors were losing money bigly on inference and will have to pull the rug eventually.