DeepSeek hadn’t only taken the world by storm by launching its R1 model which performed admirably on several benchmarks, but it was also substantially cheaper than the competition. And DeepSeek has now said that in spite of being extremely cheap, it can still be extremely profitable.
DeepSeek has claimed a cost profit margin of 545 percent in a post on X. “Statistics of DeepSeek’s Online Service: 73.7k/14.8k input/output tokens per second per H800 node, Cost profit margin 545%,” it said, while also detailing how the company had technically optimized its infrastructure to achieve the number.
The tweet led to several responses from the AI community. Some people wondered why OpenAI was losing money, while DeepSeek had a 545 percent profit margin. “Holy shit, DeepSeek is able to get 73.7k tokens/s input and 14.8k tokens/s output throughput per H800 node! Their profit margin is 545%, while OpenAI is bleeding money despite charging so much?? that’s how awesome their inference stack is,” wrote X user Yuchen Jin. “DeepSeek categorially proves we don’t need all those GPUs..DeepSeek is almost 300x more efficient (than OpenAI),” wrote Bindu Reddy.
But there’s more than meets the eye with DeepSeek’s headline 545 percent cost profit margin number. DeepSeek elaborated its 545 percent number in a blogpost.
Within the 24-hour statistical period (UTC+8 02/27/2025 12:00 PM to 02/28/2025 12:00 PM), V3 and R1:
- Total input tokens: 608B, of which 342B tokens (56.3%) hit the on-disk KV cache.
- Total output tokens: 168B. The average output speed was 20–22 tokens per second, and the average kvcache length per output token was 4,989 tokens.
- Each H800 node delivers an average throughput of ~73.7k tokens/s input (including cache hits) during prefilling or ~14.8k tokens/s output during decoding.
The above statistics include all user requests from web, APP, and API. 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%
How DeepSeek has arrived at the 545 percent figure is a bit unusual. Profit margins are usually below 100 percent, but DeepSeek calculates the costs of running the models, and compares it with the revenue it generates. By spending $100, DeepSeek generates revenue worth $645, which makes for a 545 percent cost profit margin. In more traditional terms, DeepSeek is earning margins of around 84 percent.
But crucially, DeepSeek isn’t even earning all the revenue it uses in its calculation. It says so in its blogpost;
However, our actual revenue is substantially lower for the following reasons:
- DeepSeek-V3’s pricing is significantly lower than R1,
- Only a subset of services are monetized (web and APP access remain free),
- Nighttime discounts are automatically applied during off-peak hours.
So DeepSeek’s revenue is theoretical, not its actual revenue. Also, the number is likely overestimated — it assumes that all the inference is on its R1 model, while in reality much of it is on the cheaper V3 model. Also, DeepSeek assumes it’s making money from services it’s giving away for free, like its free web usage, which overestimates its cost profit margins.
But even if the number is inflated, DeepSeek has made some impressive optimizations in running its models. The company has managed use NVIDIA’s older H800 chips, written software to optimize them, and managed to achieve results that more than hold their own against US companies like OpenAI and Anthropic. And even more impressively, it’s been sharing its methods and results for all to see, unlike US labs. DeepSeek’s 545 percent cost profit margin might not tell the entire story, but it doesn’t take away from the fact that it’s created an exceptionally powerful model — and is still looking to improve and optimize it to yield even better results.