Comment by vineyardmike
16 hours ago
I'll bite. I suspect that these plans aren't as intensely subsidized as people assume. I believe that API usage is probably also not subsidized at all. First, yes, subs are probably subsided, but I bet a significant % of users are profitable to serve, especially the "chat" users who don't use dev tools and have short context window conversations. Yes, I think the subs also exist as a driver to get lock-in and market share. Claude Code, for example, is very good and I stopped using their competition when they released their superior product.
That said, I assume that (1) their long-term goal is to create cheaper-to-serve models that fit within their pricing targets, and use the (temporarily) subsidized subscriptions to find the features and costs that best serve the market. Maybe even while capturing more margin on the API in comparison (eg keep API prices high while lowering cost to serve a token). I've largely stopped using Opus, and sometimes even chose to use Haiku, because the cheaper models are fast and usually serves my needs. It's very possible to work all-day and barely hit the usage limits with Haiku on the $20/mo option. Long term, that could be profitable outright.
And (2) subscriptions with lower SLOs than API calls have the potential to provide "infill" usage for high fixed-cost GPUs as an alternative to idling, similar to their batch APIs. I'd believe that overnight usage limits could/should be higher than during California work-hours. I assume most big providers have pre-paid fixed cost servers, so pumping more tokens through an otherwise idle GPU is "free". They can also do a lot more cost-optimization behind the scenes, such as prompt caching, to reduce the cost of tokens.
> First, yes, subs are probably subsided, but I bet a significant % of users are profitable to serve, especially the "chat" users who don't use dev tools and have short context window conversations.
Why would WebChat users need a subscription? It's free; I've even pasted tarballs of entire repos in there, and haven't hit limits!
>
More limited features, like lack of model selection, more restricted use of “thinking” models.
>>> a significant % of users are profitable to serve, especially the "chat" users who don't use dev tools and have short context window conversations.
> More limited features, like lack of model selection, more restricted use of “thinking” models.
Yeah, but... do the "chat" users actually care about any of that? Would they even notice a difference?
My point is that, if all you're doing is chat, there's no value in any of the subscription models - for chat the free webapps are more than sufficient, so even someone spending the whole day chatting about something isn't going to hit any limits.
2 replies →
> I'll bite. I suspect that these plans aren't as intensely subsidized as people assume. I believe that API usage is probably also not subsidized at all. First, yes, subs are probably subsided, but I bet a significant % of users are profitable to serve, especially the "chat" users who don't use dev tools and have short context window conversations. Yes, I think the subs also exist as a driver to get lock-in and market share. Claude Code, for example, is very good and I stopped using their competition when they released their superior product.
I somewhat agree, somewhat disagree with this. I think API based is not subsidised. If you do some basic napkin math they should have enough room there to serve the models below cost if the models aren't insanely large (you can compare with 3rd party openrouter offerings and have an idea of what $/Mtok you can serve per model size. e.g. Haiku level models can be ~700B tokens and still be profitably served)
I think 20-200$ all-you-can-prompt are likely subsidised. If you track token usage (there are many 3rd party tools that do this) you can get 4-5x the API usage out of them (it used to be even higher before they added weekly limits. People were seeing 10-20x usage). Now I think that's a bit tough to make the napkin math work out. I've compared sessions served over API with sessions from subscriptions, and you get much more usage out of them, even with 5h / weekly limits. Strictly for coding, I think they're subsidising them.
I somewhat disagree that they're doing it for market share / user lock-in. I think signals and usage trends are much more valuable for them. While there might be user retention for "casual" users (i.e. web) I think the power users in coding will move as soon as the competition has a better product. So at the end of the day having data to improve models and have the "best" model in a niche is more productive than retaining users with an inferior product. That is an assumption tho, and there isn't much math you can do to figure that out from the outside.
One thing to remember is that not all users are going to max out their plan.
This is more likely to occur on $20 plans though, especially since those are often necessary to unlock the more useful features (e.g. deep research) so people might be paying for that even if they don't actually use the tokens.
OTOH someone who's paying $200 will likely want to squeeze the most out of their subscription for that amount of money. So I wouldn't be surprised if it turns out that $20 users are subsidizing the $200 ones.
As a company with little other (any?) revenue you have to include all costs though. Data centers, power, hardware, salaries, marketing, etc. Not just training models and serving requests.
I don’t see how it’s not subsidized substantially considering how much money they’re burning right now (I only base that on their rounds though).
That's not really how people discuss subsidies and finances though. Yea I guess a not-profitable company means that every operation is technically a "subsidy", but again, that's not really what those words mean.
Anthropic (as the ever-chosen example) has explicitly stated they've made more money than they've spent on training when they sell/serve a particular model in the past. They said that the reason they're negative is the next model costs more than the "profit" they've made on the previous one. This wasn't strict financial disclosures, but I'd presume this means that their data center costs (eg. power, hardware, etc) are baked into that, but probably not company-wide costs like marketing.
They do have several sources of revenue, all tied to their models: APIs, Subscriptions, and model licensing. Their licensing and APIs most likely have a positive margin -> the money they make to serve the n+1 customer is more than the cost to serve that customer, on a per-financial-transaction basis. It's speculated that they lose money per-customer to serve the subscriptions, and they eat that cost... for various potential reasons.
It is when you discuss financial health of a company, at least that’s what I picked up after doing fintech and loans, it’s the bottom line that matters, or the projected outcome of the same. What point is there making money in area A when area B costs more. If you can stop doing B without affecting A that’s usually what happens, but it’s not always possible.
Saying ”we’re positive except the foundation of the company (training models) isn’t” is a tell tale sign.
And I’m sure Anthropic is doing what most others are doing, heavily massaging numbers to make them look good for VC rounds.