DeepSeek, a Chinese AI startup, has been "open source" recently. Last Saturday (1st), there was a bigger surprise. It fully revealed the DeepSeek-V3/R1 inference system, which not only disclosed the core optimization solution of its inference system, but also disclosed key data such as cost-benefit rates for the first time, causing industry shocks.

DeepSeek released its first article on Zhihu platform last Saturday, announcing the details of model inference cost-profits and disclosed key information on cost-profit rates. If all tokens are calculated based on the pricing of DeepSeek-R1, the theoretical total revenue per day is US$560,000, and the cost-profit rate is 545%. This figure refreshes the profit ceiling in the global AI big model field.

According to the official disclosure of DeepSeek, all services of DeepSeek V3 and R1 use H800 GPU, using consistent accuracy with training, that is, matrix calculation and dispatch transmission adopt consistent FP8 format with consistent training, and core-attention calculation and combine transmission adopt consistent BF16 with consistent training, ensuring the service effect to the greatest extent.

In the statistical cycle of the last 24 hours (12:00 on February 27, 2025 to 12:00 on February 28), if the GPU rental cost is calculated at US$2/hour, the average daily cost is US$87,072, and if all input/output tokens are priced at R1 (input 1 yuan/million token, output 16 yuan/million token), the daily income can reach US$560,027 (about NT$18.65 million), and the cost interest rate is as high as 545%.

After reading the above data, MenloVentures investor Deedy pointed out that the business efficiency of the profit rate exceeding 500% will be a company worth more than 10 billion US dollars in the United States.

Yuan Jinhui, founder of China's silicon-based mobile phone, also expressed his feelings at the first time: "DeepSeek's official disclosure of the cost and benefits of large-scale deployment has once again subverted many people's perceptions."

DeepSeek's high profit rate comes from its innovative inference system design, with three technical pillars: large-scale cross-node expert parallelism (EP), computing communication overlap and load balancing optimization. EP improves throughput and response speed. For model sparsity (only 8/256 experts are started per layer), EP strategy is used to expand the overall batch size to ensure that each expert obtains sufficient computing load, significantly improves GPU utilization, and dynamically adjusts the deployment unit (such as 4 nodes in the Prefill stage and 18 nodes in the Decode stage), and balances resource allocation and task requirements.

In short, EP is like "multi-person collaboration", dispersing the "experts" in the model to multiple GPUs for calculations, greatly improving Batch Size, squeezing the GPU computing power, and at the same time dispersing experts, reducing memory pressure, and responding faster.

DeepSeek further compresses costs at the engineering level, plus day and night resource allocation, fully supports inference services during peak days, and idle nodes at night are transferred for R&D and training, maximizing hardware utilization, and the cache hit rate reaches 56.3%. Reduces duplicate calculations through KVCache hard disk cache. Among the input tokens, 342 billion (56.3%) hit caches directly, greatly reducing computing power consumption.

Some analysts say that the data disclosed by DeepSeek not only verifies the commercial feasibility of its technical route, but also sets a benchmark for efficient profitability for the industry. The cost of model training is only 1% to 5% of similar products. The previously released DeepSeek-V3 model training cost is only 5.576 million US dollars, far lower than that of giants such as OpenAI. In terms of inference pricing advantages, DeepSeek-R1's API pricing is only about one-seventh to half of OpenAI o3-mini, and low-cost strategies accelerate market penetration.

Other analysts pointed out that DeepSeek's "transparent" disclosure not only demonstrates its technical strength and business potential, but also sends a clear signal to the industry, that is, the profit cycle of AI models has shone from ideals into reality, representing a key turning point in AI technology from laboratory to industrialization.

However, DeepSeek officially admitted that there was actually not so much revenue, because V3 was priced lower, and paid services only accounted for a part of the time, and there were discounts at night.

CITIC Securities believes that Deepseek's best practices in reducing model training costs are expected to stimulate technology giants to adopt a more economical way to accelerate the exploration and research of cutting-edge models, and at the same time, it will enable a large number of AI applications to be unlocked and implemented. The increasing effect of scale returns brought by algorithm training, as well as the Jevins paradox corresponding to the reduction of unit computing power costs, all represent that medium and short-term dimensional technology giants continue to make continuous investment in the field of AI computing power, and scale investment will still be a high-deterministic event.

The above is the detailed content of 545%! DeepSeek first disclosed cost profit margin Expert: If it is already a company worth more than 10 billion US dollars in the United States. For more information, please follow other related articles on the PHP Chinese website!