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.
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.
> Deepseek also came with heavily subsidized plans, at least at first.
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%.
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
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.
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.
OpenCode is building their own inference and they've independently stated that DeepSeek's old (cheaper) pricing is achievable without subsidies.