In the past week’s open source week, DeepSeek’s “five consecutive bombs in five days” made the market applauded.
And just when the outside world thought that this feast was about to come to an end, DeepSeek came up with an even more shocking “easter egg” – the cost profit margin was as high as 545%, and the theoretical daily profit of the V3/R1 inference system reached 3.46 million yuan.
While the market sighs again and again for this ultra-high “windfall profit”, it is also more concerned about these questions: how to interpret the profit margin of 545%? Is it a nuclear bomb on computing power? What does this mean for the cloud industry chain? What does this mean for their big model peers? What does this mean for ecology? The main points of the digital tech bloggers are as follows:
The profit margin of 545% is still a theoretical return, and the profit margin of the DeepSeek R1 model is about 85%. And if you go with the V3 pricing, the profit margin will drop to about 70%. Even so, the number is still very impressive.
For the computing power industry chain, DeepSeek’s case proves that even in the case of relatively limited hardware conditions (using H800), extremely high computing power utilization and performance can be achieved through extreme infra optimization.
However, there is still a lot of disagreement about whether DeepSeek’s innovation will reduce the need for computing power.
Duan Yongping, a well-known investor, agrees with the view of former Nvidia CEO Jensen Huang that the demand for computing power will continue to grow. However, some foreign technology bloggers said that DeepSeek has “knocked down” Nvidia, and according to DeepSeek’s current ultra-high utilization rate of computing power, there is not so much demand for AI in the world.
In addition, DeepSeek’s case proves that the similarity between AI cloud computing and traditional cloud computing is more obvious. AI cloud computing will also face the challenges of “idle rate during low peak periods” and “stability during peak periods”.
DeepSeek’s open source and technical disclosures have set a new benchmark for the entire industry. Peers may face greater competitive pressure, and a new round of price war is on the way.
For the industry ecosystem, DeepSeek will attract the industry to build to B and C services on its basis through open source technology and output, forming a complete upstream and downstream of the industry.
01
How to interpret this profit margin?
First of all, it should be clear that the 545% profit margin officially announced by DeepSeek is based on a “theoretical” calculation under certain conditions, that is, it is assumed that all tokens are calculated according to the R1 model, and do not consider factors such as the lower pricing of V3, the proportion of free services, and nighttime discounts. In fact, according to DeepSeek’s official statement, their true profit margins are far less exaggerated.
According to tech blogger 180K, the profit margin of the DeepSeek R1 model is about 85%, and if it is priced according to V3, the profit margin will drop to about 70%. Even so, the number is still very impressive.
According to 180K, this can be understood more deeply by comparing Anthropic’s profit margins. According to TD Cowen’s teardown, Anthropic’s profit margin for 2024 is expected to be 61%. If you follow DeepSeek’s caliber and consider AWS’s cloud computing margin (let’s say 25%-40%), Anthropic’s profit margin can reach 74%. In extreme cases, if AWS’s profit margin is assumed to be 50%, Anthropic’s profit margin can even reach 85%, which is comparable to DeepSeek’s R1 model.
This illustrates that while OpenAI and Anthropic may not be as cost-effective as DeepSeek, they can achieve similarly high profit margins with higher pricing and more generous customers (at least for now). It should be noted that OpenAI is often reported as “loss-making”, because when raising funds, investors usually focus on the profit and loss of financial accounting, rather than the theoretical cost from the perspective of large model leasing, and operating expenses such as model training costs, data authorization fees, personnel and publicity are usually also included.
02
Is it a nuclear bomb on computing power?
DeepSeek’s case proves that even in the case of relatively limited hardware conditions (using H800), extremely high computing power utilization and performance can be achieved through extreme infra optimization, which has a huge impact on the entire computing power industry chain:
First of all, tech blogger 180K believes that the importance of “effective computing power” will be highlighted. The industry will pay more attention to “effective computing power” (computing power x computing power utilization), rather than just computing power accumulation.
And the upper limit of domestic chips is expected to increase. If H800 can run such an effect, then through infra optimization, the performance ceiling of domestic chips may be further improved.
In addition, tech bloggers believe that the “Jevons paradox” continues to be in effect. The improvement of computing power efficiency will not reduce the demand for computing power, but will stimulate the emergence of more application scenarios and promote the continuous growth of computing power demand. As Barclays predicted in June last year, by 2026, the industry’s capital expenditure will be enough to support “12,000+ ChatGPT-level applications”.
Moreover, the logic of computing power demand may be questioned in the short term. Some companies, especially CIOs or CFOs of major overseas companies, may face pressure from investors and bosses to explain why their ROI is much lower than that of DeepSeek.
Duan Yongping, a well-known investor, also said on Xueqiu that DeepSeek’s experience has indeed proved that lower computing power in the pre-training stage of the model can also achieve better training results. And he agrees with Huang that DeepSeek’s innovations won’t reduce the need for computing power.
Previously, Huang said in an interview in February that he believed the market’s understanding of DeepSeek was completely reversed. He said that the emergence of R1 does not mean that the market no longer needs computing resources, but that it has stimulated the market’s pursuit of more efficient AI models, thereby promoting the development of the entire industry.
But foreign technology blogger Zephyr believes that DeepSeek has “knocked Nvidia down”. And according to DeepSeek’s current ultra-high utilization rate of computing power, it is more than enough to meet the global demand for AI.
DeepSeek has already “knocked Nvidia down”.
I say this because DeepSeek currently processes 600 billion tokens per day on 300 H800 nodes (2400 H800s in total) and outputs 150 billion tokens. If the computing power is increased by a factor of 100 (i.e., 240,000 H800s), 60 trillion tokens can be processed and 15 trillion tokens can be output every day. But the global demand for AI isn’t so high.
03
What does it mean for the cloud industry chain?
The success story of DeepSeek makes the similarity between AI cloud computing and traditional cloud computing more obvious. AI cloud computing will also face the challenges of “idle rate during low peak periods” and “stability during peak periods”.
Technology blogger 180K believes that the scale effect of cloud computing will be more significant. DeepSeek’s practice shows that large-scale clustering and high concurrency utilization can significantly reduce costs. The positive externality of the number of users is more obvious, that is, the more users, the stronger the ability to stabilize fluctuations, and the lower the demand for computing power redundancy.
The competitive advantage of cloud vendors may change. Cloud vendors with their own services (such as Alibaba, Tencent, Apple, etc.) may have a cost advantage over cloud vendors without their own services, because they can use inference clusters as the foundation for all services to achieve greater scale effects.
And there is room for improvement in the profit margins of cloud computing. DeepSeek’s case shows that in the age of AI, cloud computing margins have the potential to be further improved through extreme infra optimization.
In addition, private cloud deployments may become less attractive. Ultra-sparse MoE models may not be suitable for individual or “half-bucket” enterprise deployments, and the cost of small-scale GPU deployments may be much higher than that of large manufacturers. This may lead to more enterprises opting for a public cloud or hybrid cloud model.
Common cloud computing/AI applications need to reserve more space for high-intensity user concurrency. Users have a high tolerance for DeepSeek’s “server busyness”, but not for other applications. This could lead to a further decline in profit margins for general cloud computing/AI applications.
04
What does it mean for large model peers?
DeepSeek’s open source and technical disclosures have set a new benchmark for the entire industry.
Tech blogger Information Equality believes that DeepSeek’s case shows that the “bottom line” of reasoning costs has been greatly lowered and can be much lower than previously expected. And a new round of price wars may break out, and peers will face greater pressure to cut prices to remain competitive.
In addition, DeepSeek provides a clear optimization path and goal for all inference teams, and the follow-up pressure will increase.
In addition, OpenAI’s high-priced subscription model will also face challenges in this case, and the high subscription fee of $200 per month is somewhat embarrassing.
05
What does it mean for ecology?
DeepSeek’s strategy is to focus on basic models and cutting-edge innovations, and attract the industry to build to B and C businesses on its basis through open source technology and output, forming a complete upstream and downstream industry.
Tech blogger Geek Park said that the profit margin of ecological partners has increased. The cloud platform and upstream and downstream can theoretically obtain high revenue and profit margins by deploying DeepSeek’s services.
Looking forward to the subsequent ecosystem, the differentiation of model architecture may become the key to competition. Because the architecture of DeepSeek V3/R1 is quite different from the mainstream model, it requires vendors to adapt, which is difficult to develop.
Moreover, DeepSeek’s open-source initiative reduces the difficulty for the community to reproduce its reasoning system, which is conducive to the prosperity of the ecosystem.
Tech blogger 180K said that the entire industry may start to roll Infra. To some extent, Infra is becoming more important, and valuations can be increasing.
All in all, DeepSeek’s ultra-high profit margin is not only a digital miracle, but also a profound revelation to the entire AI industry. It reveals the huge potential of infra optimization, promotes the transformation of computing power, cloud, large models, and ecosystems, and heralds a more efficient, low-cost, and highly competitive AI era is coming.
Author：朱雪莹 Source：DeepSeek的545%利润率，是对算力的核弹吗？ The copyright belongs to the author. For commercial reprints, please contact the author for authorization. For non-commercial reprints, please indicate the source.
