Comment by riknos314
8 hours ago
Deepseek prices their models based on a 10 month hardware payback period. They report ~$38 revenue/GPU-hour on V4.1-Flash clusters.
Inference is clearly profitable. The question is whether it's profitable enough to justify the training research costs.
Are those statements audited or fact checked? Inference profitability is great but how much of that profit gets eaten up paying back training costs, R&D, and operating expenses.
Their models and inference frameworks are open, you can run them yourselves to check how many tokens per day you can generate, and then validate that against their API prices.
Not sure how they can get more transparent than that.
No, and it’s very rare for Chinese companies to allow independent audits. I would not trust DeepSeek financial disclosures. However from industry data we can see that inference is often profitable. It’s the training which is very expensive. Which is not a big problem for Chinese companies since to date much of their training is just distilling frontier models.
At best, this tells you v4.1 flash is profitable, but can they really get $38 per gpu hour? Sounds like a lot.
From their own inference framework docs: "A single cluster currently serves 155 trillion tokens per day, with overall revenue of US$38 per GPU-hour (at valley pricing)."
https://github.com/deepseek-ai/DeepInfer/blob/main/doc/infer...
So this implies they actually hit about double that during peak pricing.
If they keep it running at near constant usage, you'd only need 80-120 million tokens an hour to hit 38$ of revenue (at off peak pricing, half as much for peak pricing).
V4.1-Flash is probably quite smaller and much more resource effective than the frontier models. Add to that much cheaper energy in China.
Clearly potentially profitable. Not clearly profitable.
How much did it cost to train Deepseek v4.1 flash?
Great question. Nobody knows.
The cool thing about models designed for cheaper inference is that post-training (which often requires inference as a part of the feedback loop) becomes cheaper as well, so almost certainly cheaper than models that are less optimized.