Hacker News Reader

Comment by judge2020 | Hacker News Reader

judge2020

Another part of the discount is that Claude Max 20x is $200 but gives usage equivalent to thousands of dollars worth of API-based token spend.

But also, resellers only need to make an overall profit including kickbacks from the companies purchasing token history for distillation.

View on HN

It's also unclear wherever the subscription price is the real cost, or the API.

I suspect it's closer to the sub price and anthropic is just milking their API users, but that's something you'd only know from the inside

Generally speaking, B2B prices are rarely supply-and-demand priced in the usual sense.

YC has advised startups in the past that it's easier to sell a single $100k customer than 100 $1k customers.

It would also be relatively surprising to learn that i.e. the Chinese providers are OOMs better at inference than OAI/Anthropic (like their prices would imply if they were in a perfectly competitive market).

Chinese prices aren't really OOM cheaper. Deepseek recently did a big price hike too.

The one thing I trust is that Chinese prices aren't overinflated. They're almost certainly closer to the actual cost of inference + training amortization than what Western labs are offering.

DeepSeek's price hike is mostly driven by increased demand, for example. It's not about losses so much as they don't have enough infrastructure and need to reduce demand somehow.

Why do you have this trust? Are they not profit oriented? Or you suspect that Xi's policy push is to undercut American AI?

Are you familiar with Chinas version of capitalism? It is markedly different than “the” western model

I am not, but would be interested to hear. Any articles?

"Overcapacity and low profitability" [1] is how I understand the difference, i.e. different competitive ecosystem from a different policy and intervention style.

https://www.dbresearch.com/PROD/IE-PROD/PDFVIEWER.calias?pdf...

There's no English-language article that can fully capture the system because it's a moving target to begin with, and anything official is vague on purpose (to be fair, much of the legislation in the US is too, except the Constitutional backstop and co-equal judiciary does not exist in China). The official messaging is always targeted and curated with a specific audience in mind and how much truth is contained within varies. The best way to describe it without turning it into an essay is that it's "aspirationally market-driven" system that is not defined by a fixed and consistent ideology but rather the constraints of the capacity of the administrative state to control the parts of the economy that is legible to the state. It's not a free market except for the black market, but the black market is frankly huge if nothing else to compensate for the state asserting not just a monopoly on the use of violence but also the monopolistic determinator of morality, which the people are less willing to buy into, but that disconnection itself creates an opportunity for arbitrage. That is the relationship dynamic that have retained some degree of stability, but everything else is effectively subject to change on a dime without notice. The party will never accept openly that it is capitalistic and in some sense that's not even a lie, but what the state is able to impose and what the people actually do are very different things. As the state excludes 99% of the population writ large from politics, full stop (you are either born into it, or you're not eligible to participate, with some narrow exceptions that occasionally pop up), there's not much of a ground-level appetite for dogma and doctrines. Instead you end up with a pragmatic, descriptive market that exists differently within and without, and just like the socialism label being basically symbolic, same goes for capitalism.

People have been talking about late capitalism in the west for a century. The Chinese state would argue that we cannot be in late capitalism because we're actually in the first stage of socialism, and the people would wonder how we got to late capitalism when they never got to experience early or middle capitalism. But really the only thing we know that definitively does not work is implenting Marxism and Marxism-Leninism and Maoism and its offshoots as doctrine. Whether that means that the default is a prescriptive version of capitalism is, well, open to debate, and probably too simplistic.

There is no "Chinese capitalism" or "Western capitalism" - capitalism is applied thermodynamics, once you remove the limiting factors (both material and political), it's the same everywhere.

Like the 10th Rule of Acquisition says, greed is eternal. Structures and behaviors that flow down from this tend to be the same everywhere.

Once you remove the limiting factors, which is what makes the capitalism be "chinese" or "western".

The problem with this reading is that Western capitalism has been effectively removing the thermodynamics analogy through monopolism, conglomeration, regulatory capture, etc.

I don't think DeepSeek's price increase has to do with increased demand, but rather increased investment. I think they increased their capacity and want to make the money back.

Fair, I hadn't looked recently. It looks like currently kimi is either 1/2 or 1/4 Ant pricing depending on whether you think Opus 5 is usable or not. (Deepseek is still an OOM though)

My pet theory is that the PRC is heavily subsidizing the hosted inferencing costs in order to capture valuable training data, token usage patterns and the occasional leak of valuable data.

Given their past history and current geopolitics, I'd be willing to bet this is more likely than not.

Costs also have to include the amortized training costs.

But the API price is likely simply regular supply and demand, charging as much as the market will pay. Corporations are dropping insane amounts because it's still peanuts for many industries. Software has just been ridiculously cheap before AI. So high prices are still low for companies if it eases some bottlenecks.

The costs only have to include amortized training costs if you are trying to be profitable overall. Having positive unit economics and VC subsidized fixed costs is pretty standard.

What happens when the VCs decide to stop dumping more money onto the fire?

Positive unit economics means that inference as a business continues regardless. The VC's no longer dumping money in means no more training new models.

Not if there is competition who does train newer and better models.

By lighting VC (public soon) money on fire...

So then the entire industry stagnates and most of the companies go bankrupt because they are all selling hype about what models "will" do in the future, not what they do not, and what they do WELL now is very limited.

this would never work if the core business model requires your flagship product to be best-in-class.

> It's also unclear wherever the subscription price is the real cost, or the API.

Define cost: Is it only the inference cost to the provider, or do you also consider training costs as well?

If it's the latter, how would you estimate the number of total tokens that will be sold for the current model (so that we can calculate marginal cost)?

“It's also unclear wherever the subscription price is the real cost, or the API.”

There is no “real cost” other than the cost actually charged.

Open weight SOTA models are not greatly cheaper than Anthropic and providers don't have to cover training capex.

> and providers don't have to cover training capex.

How so?

Providers like Fireworks offer inference, they're not training their own models. Yet training is the most expensive part, inference is easy profit.

And even the ones who trained it can’t afford to offer it at a price covering their training capex, since any good open-source model will see other providers selling inference without any training expenses to cover.

Didn’t see it, but it’s obvious in retrospect. Thanks!

With an open-weights model, anyone with the hardware to run the model can act as a provider. The median provider pays nothing for the model, so they do not contribute to the cost of training it.

Ah! Makes sense. Thanks!

> Another part of the discount is that Claude Max 20x is $200 but gives usage equivalent to thousands of dollars worth of API-based token spend.

this is all it is. it's not complicated.

With Claude Max, you have restricted token consumption per timespan (rolling time window measures token consumption in last 5 hrs, will suspend your account if you are above usage). API doesn't have that limitation unless you run out of credit.

Reply on news.ycombinator.com