Comment by Aurornis | Hacker News Reader

Comment by Aurornis | Hacker News Reader

Parent Full thread

Deepseek also came with heavily subsidized plans, at least at first.

This is why there are so many comments confused that you spent that much money on Deepseek. Those who got on one of the discounted plans could use very large numbers of tokens for trivial prices.

There is also a strange double standard for accounting for open models. People will look at OpenAI or Anthropic and say that we need to consider all of their training costs and employee compensation when thinking about the cost to serve their models, but when the models are released as open weights those costs are ignored. So in that way, the Deepseek models are heavily subsidized as well, with the possibility of them being served by a different company that paid nothing to develop them.

> Quality was ok, seems slightly above Luna quality perhaps?

I agree that it’s about in line with what you can get from Luna or Haiku, but I give the edge to Luna and Haiku when it comes to tasks that require world knowledge. It feels like their training sets were just cleaner.

Luna and Haiku are also close to free with a subscription plan, and they’re even very cheap at API rates.

So the amazing thing about Deepseek Flash is that you can almost kind of get that level of performance from an open weight model. It’s not as amazing when you start comparing it for how we really use smaller models from frontier labs on subscription plans.

> So in that way, the Deepseek models are heavily subsidized as well, with the possibility of them being served by a different company that paid nothing to develop them.

You could then of course argument that oAI/ant/G's models were also heavily subsidized by using (scraping) the bulk of humanity's global knowledge for free while paying nothing for it and trying to privatize it.

Deepseek never subsidized their plans. At some point they dropped pricing, then after a while they raised them, but that's because they had capacity issues, not because of dropping subsidies. They don't want your money, they'd rather have fewer users.

OpenCode is building their own inference and they've independently stated that DeepSeek's old (cheaper) pricing is achievable without subsidies.

> So in that way, the Deepseek models are heavily subsidized as well, with the possibility of them being served by a different company that paid nothing to develop them.

Has anyone seen the neo-cloud profit margin on deepseek? I wonder if the deepseek served API prices account for the training cost? Because they don't / have not raised the money to fund their future operation / training- and they depend on that cashflow for now? That would suggest a really high profit margin on inference only.

    > Deepseek also came with heavily subsidized plans, at least at first.

Do you have a source for that?

I got a source straight from the hoses mouth, which claims that DeepSeek-R1 was highly profitable:

    > If all tokens were billed at DeepSeek-R1’s pricing (*), the total daily revenue would be $562,027, with a cost profit margin of 545%.

https://github.com/deepseek-ai/open-infra-index/blob/main/20...

With DeepSeek-V4.1-Flash, the memory requirements for the KV cache have been reduced by over 50x and FLOPS by over 5x, so it is even cheaper: https://arxiv.org/pdf/2609.19969

Deepseek Flash plans were very cheap when they first launched. They raised prices https://finance.yahoo.com/technology/ai/articles/deepseek-ra...

Your link is about a different model.

The talk about subsidies is confusing because for OpenAI and Anthropic people usually include their salaries, training costs, and everything else. When the topic switches to open weight models we pretend the models appeared out of the ether at zero cost, and the only cost is running the servers.

Right, but I'm not sure that was because they were subsidizing it, or because they ran into capacity issues and needed to offload some users.

That "If" is carrying a lot of weight.

That accounting exists because DeepSeek doesn't need to exist for me to use 4.1 Flash. If Anthropic goes under, Claude probably just ceases to exist.

That accounting doesn't make sense when looking at future models, but if you're evaluating the current state it makes sense to ignore training costs for companies that don't pay them.

You can also get subscriptions for the open models, which are typically much cheaper. It's not crazy popular because the open model crowd switches often, but they do exist if you want a firehose of tokens.

Reply on news.ycombinator.com