Turns out there was a lot of low-hanging fruit in terms of inference optimization that hadn't been plucked yet.
> A year or more ago, I read that both Anthropic and OpenAI were losing money on every single request even for their paid subscribers
Where did you hear that? It doesn't match my mental model of how this has played out.
> Turns out there was a lot of low-hanging fruit in terms of inference optimization that hadn't been plucked yet.
That does not mean the frontier labs are pricing their APIs to cover their costs yet.
It can both be true that it has gotten cheaper for them to provide inference and that they still are subsidizing inference costs.
In fact, I'd argue that's way more likely given that has been precisely the goto strategy for highly-competitive startups for awhile now. Price low to pump adoption and dominate the market, worry about raising prices for financial sustainability later, burn through investor money until then.
What no one outside of these frontier labs knows right now is how big the gap is between current pricing and eventual pricing.
[1] https://epochai.substack.com/p/can-ai-companies-become-profi...
chasing down a few sources in that article leads to articles like this at the root of claims[1], which is entirely based on information "according to a person with knowledge of the company’s financials", which doesn't exactly fill me with confidence.
[1] https://www.theinformation.com/articles/openai-getting-effic...
They are for sure subsidising costs on all you can prompt packages (20-100-200$ /mo). They do that for data gathering mostly, and at a smaller degree for user retention.
> evidence at all that Anthropic or OpenAI is able to make money on inference yet.
You can infer that from what 3rd party inference providers are charging. The largest open models atm are dsv3 (~650B params) and kimi2.5 (1.2T params). They are being served at 2-2.5-3$ /Mtok. That's sonnet / gpt-mini / gemini3-flash price range. You can make some educates guesses that they get some leeway for model size at the 10-15$/ Mtok prices for their top tier models. So if they are inside some sane model sizes, they are likely making money off of token based APIs.
The interesting number is usually input tokens, not output, because there's much more of the former in any long-running session (like say coding agents) since all outputs become inputs for the next iteration, and you also have tool calls adding a lot of additional input tokens etc.
It doesn't change your conclusion much though. Kimi K2.5 has almost the same input token pricing as Gemini 3 Flash.
so my unused tokens compensate for the few heavy users
Anthropic planning an IPO this year is a broad meta-indicator that internally they believe they'll be able to reach break-even sometime next year on delivering a competitive model. Of course, their belief could turn out to be wrong but it doesn't make much sense to do an IPO if you don't think you're close. Assuming you have a choice with other options to raise private capital (which still seems true), it would be better to defer an IPO until you expect quarterly numbers to reach break-even or at least close to it.
Despite the willingness of private investment to fund hugely negative AI spend, the recently growing twitchiness of public markets around AI ecosystem stocks indicates they're already worried prices have exceeded near-term value. It doesn't seem like they're in a mood to fund oceans of dotcom-like red ink for long.
VC firms, even ones the size of Softbank, also literally just don't have enough capital to fund the planned next-generation gigawatt-scale data centers.
The evidence is in third party inference costs for open source models.
are we sure this is not a fancy way of saying quantization?
The same is happening in AI research now.
Just curious, which formats and how they compare, storage wise?
Also, are you sure it's not just moving the goalposts to CPU usage? Frequently more powerful compression algorithms can't be used because they use lots of processing power, so frequently the biggest gains over 20 years are just... hardware advancements.
And if you've worked with pytorch models a lot, having custom fused kernels can be huge. For instance, look at the kind of gains to be had when FlashAttention came out.
This isn't just quantization, it's actually just better optimization.
Even when it comes to quantization, Blackwell has far better quantization primitives and new floating point types that support row or layer-wise scaling that can quantize with far less quality reduction.
There is also a ton of work in the past year on sub-quadratic attention for new models that gets rid of a huge bottleneck, but like quantization can be a tradeoff, and a lot of progress has been made there on moving the Pareto frontier as well.
It's almost like when you're spending hundreds of billions on capex for GPUs, you can afford to hire engineers to make them perform better without just nerfing the models with more quantization.
