Comment by skerit
19 hours ago
> The major providers are unprofitable after you add in the training and research costs and other overheads
The very short lives of all these models also factor in.
I think the funny thing is the more inference, the more tokens a specific version of a model spits out, actually does make it more profitable.
But somehow a lot of people truly think that each token generated increases the cost of OpenAI/Anthropic/... and shout bloody murder about "subsidies"
And I'm certain there are people out there that actually think the API prices of OpenAI and Anthropic somehow equal their electricity costs.
Nobody meets expectations of a 2 trillion dollar IPO selling a fungible commodity like tokens for old models.
And that's the rub - it doesn't matter if inference is profitable, the US labs have promised the whole kitchen sink, and taken serious investment on that pitch. They won't make it back in a cut-throat market for inference. They have to continue pitching the idea that only their models are worth using, for quality/safety/performance/etc reasons.
Which means continuing to develop and release new models. At huge expense.
Side note - while deepseek seems to be doing fine, I'll note that the emphasis on cost effectiveness is very much not a thing for the US labs right now... and it was for non-us labs. I'd be quite interested in knowing if anthropic is pulling in the same returns, or what the margin on their subscription plans actually is.
---
I'm already seeing this directly at work - lots of devs still happily use claude for development, but the products we're making aren't. They'll use kimi, glm, gpt-oss, etc... basically - if cloudflare is hosting it, we're exploring it (mainly for contractual/security reasons): https://developers.cloudflare.com/workers-ai/models/.
Because we are price sensitive, and we already have to do all the work of creating very robust evaluation frameworks for our products, so we might as well choose the cheapest model that meets our standards.
whiiiiich is why we're seeing things like Haiku and gpt-luna priced down near the open models at ~.10/mm in, ~.50/mm out. They've been forced to cut prices to keep commercial customers on their cheaper models instead of moving away.
This will continue. If inference is profitable and easy, prices will continue to erode. Go browse openrouter, plenty of providers happy to undercut the competition. Bigger shops with auditing requirements will avoid them, but smaller ones won't.
You know, it's not just cost-effectiveness, it's also iteration time and speed. DeepSeek V4.1 Flash is astoundingly fun to use because it's so quick. Tens of times faster than the standard Frontier models. With, in my experience, practically the same results. Because most of the trouble is not the intelligence of the model. It's the feedback from the world. The quicker you can get feedback from the world, the better your process is, irrespective of raw intelligence metrics of the AI you're using.
I 100% agree and am a big fan of DS V4.1 Flash in general for this reason, but I'll add my own caveat: Sol 6.1 isn't far behind on completion time for tasks because it's so damn efficient. I regularly get ~300 tokens/sec with DeepSeek Flash, but it spends the majority of that time thinking and executing slow commands one by one, whereas Sol thinks less and is much smarter about the commands it runs, resulting in far fewer turns to get the same thing done. For highly agentic tasks, DS might only be ~25-50% faster, not 500% as the TPS may suggest.
(But to be clear: 25-50% faster is still a big win in my book.)
1 reply →
> Which means continuing to develop and release new models. At huge expense.
I guess my question is when do we say a model is "good enough"? For my use cases Opus 5.5 feels as if it is exactly at that point. And at that point of course one of the main factors I will evaluate against is price. So far I am happy with my (heavily subsidized) subscription, but if the time comes, I'll happily switch the inference API any of my agents are calling. So far I have no big lock in on any providers.
define "we".
Because again - I think lots of businesses are already taking a good look at models well off the "cutting edge" because they're good enough for what that business is trying to do. They're already there.
If you're doing categorization, tagging, form-filling, customer support routing, etc... you don't need anywhere close to cutting edge expensive models.
Hell - Jev being so popular is basically exactly that group going "what's the cheapest way we can make this decision".
Meanwhile, there are plenty of spaces where you still want real model improvement. But a lot of traditional business was basically process and paper pushing. You don't need a Phd level model for that, you need an intern level model.
---
But that's the trap for the large providers - intern level models are cheap and easily available with current open weights. It's going to be a very tight market, and margins will go down to essentially the floor (especially once we exit the current constrained-compute market).
Somewhat ironically - the technology itself makes switching costs basically zero. It's not just that there's no big lock in, AI is like an anti-moat. It actively fills in and paves over whatever moat you might have for price conscious customers.
1 reply →
When it solves your problem. I don’t think it makes sense to globally qualify a model as “good enough” any more than it does with programming languages. Python is good enough for a web app, but not for an OS kernel.
To your point, I think we’re past the hump on the Pareto distribution. The models are good enough for most things, and the rate they’re improving at the things they’re bad at is dropping. The odds a new model is good at something the old one was bad it is okay, the odds it’s something you care about is dropping. It makes a lot of sense to discriminate on price rather than features in that environment.
"640K ought to be enough for anybody"
1 reply →
The Chinese labs are just arms of the Chinese government. They have no obligation to be profitable or please (private) investors. If they weaken/collapse American models, then that is the ROI to the lab owners.
This is a pretty strange take on an article explicitly pointing out the profitability of chinese models.
Further, even if this is your take (and I won't touch the geopolitics here, maybe you're right, maybe you're not regarding ulterior motives) - go use non chinese open models that are also good.
Laguna, Nemotron, Mistral, Arcee, Llama, etc...
9 replies →