Is DeepSeek's 545% profit margin a nuclear bomb for computing power?

In the past week at the open-source week, DeepSeek's "five consecutive explosions" left the market in awe. Just when the outside world thought this feast was about to come to an end, DeepSeek presented an even more shocking "Easter egg" — a cost-profit margin of up to 545%, with the theoretical daily profit of the V3/R1 inference system reaching 3.46 million RMB.

While the market marvels at this extraordinarily high "profit," it is also more concerned with several questions: How should we interpret this 545% profit margin? Is it a nuclear bomb for computing power? What does this mean for the cloud industry chain? What does this mean for peers in large models? What does this mean for the ecosystem? The main viewpoints of several tech bloggers are as follows:

The 545% profit margin is still a theoretical return; the profit margin of the DeepSeek R1 model is about 85%. If calculated based on the pricing of V3, the profit margin would drop to around 70%. Even so, this figure is still very impressive.

For the computing power industry chain, DeepSeek's case proves that even under relatively limited hardware conditions (using H800), extreme infra optimization can achieve very high computing power utilization and performance.

However, there are still many disagreements about whether DeepSeek's innovation will reduce the demand for computing power. Well-known investor Duan Yongping agrees with NVIDIA CEO Jensen Huang's previous view that the demand for computing power will continue to grow. However, some foreign tech bloggers claim that DeepSeek has already "knocked out" NVIDIA, suggesting that based on DeepSeek's current ultra-high utilization of computing power, the global AI demand is not that high.

In addition, DeepSeek's case demonstrates that the similarities between AI cloud computing and traditional cloud computing are becoming more apparent. AI cloud computing will also face challenges such as "idle rates during low peaks" and "stability during high peaks."

DeepSeek's open-source and technology disclosure have set a new benchmark for the entire industry. Peers may face greater competitive pressure, and a new round of price wars is on the way.

For the industry ecosystem, DeepSeek's open-source technology and output will attract the industry to build B2B and B2C businesses on its foundation, forming a complete upstream and downstream industry chain.

1. How to interpret this profit margin?

First, it is important to clarify that the 545% profit margin officially announced by DeepSeek is based on a "theoretical" calculation under specific conditions, assuming that all tokens are calculated according to the pricing of the R1 model, and does not consider the lower pricing of V3, the proportion of free services, and night discounts, among other factors. In fact, according to DeepSeek's official statement, their actual profit margin is far less exaggerated.

According to [tech blogger 180K's](https://mp.weixin.qq.com/s?__biz=MzkwODQzMjcwOA==&mid=2247485748&idx=1&sn=00d2b37e0a7fc28f120ade56294e163a&chksm=c1efc15e7661eb2ddb620faf5c5efe885514010195f0df6dba3daa7c161ecb7ebe1f13fe4404&mpshare=1&scene Interpretation of the link, **** The profit margin of DeepSeek's R1 model is approximately 85%, while if priced according to V3, the profit margin would drop to around 70%. Even so, this figure is still very impressive.

180K indicates that this can be understood more deeply by comparing it to Anthropic's profit margin. According to TD Cowen's breakdown, Anthropic's profit margin for 2024 is expected to be 61%. If calculated according to DeepSeek's standards and considering AWS's cloud computing profit margin (assumed to be 25%-40%), Anthropic's profit margin could reach 74%. In extreme cases, if we assume AWS's profit margin is 50%, Anthropic's profit margin could even reach 85%, comparable to DeepSeek's R1 model.

This indicates that, although OpenAI and Anthropic may not be as extreme in cost control as DeepSeek, they can achieve similar high profit margins due to higher pricing and more generous customers (at least for now). It is important to note that OpenAI is often reported to be "losing money" because, during financing, investors typically focus on financial accounting profits and losses rather than the theoretical costs from the perspective of large model leasing. Costs for model training, data licensing, personnel, and promotional expenses are usually included.

II. Is it a nuclear bomb for computing power?

The case of DeepSeek proves that even under relatively limited hardware conditions (using H800), extreme infrastructure optimization can achieve very high computing power utilization and performance, which has a huge impact on the entire computing power industry chain:

First, 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 the simple accumulation of computing power.

Moreover, the upper limit of domestic chips is expected to improve. If the H800 can achieve such results, then through infrastructure optimization, the performance ceiling of domestic chips may be further enhanced.

**Additionally, [tech blogger Information Equality](https://mp.weixin.qq.com/s?__biz=MzkyMTU4OTE2OA==&mid=2247490414&idx=1&sn=75c5319803093bdd78225d92af8adf92&chksm=c0e31b33bd200775bb743fb378bae6e9d51e399d3cf29b771ef76a4c2edd3ecb96ed947aa512&mpshare=1&scene=23&srcid=0301JXE6lSYKnZyrU7WLfrt9&sharer_shareinfo=6da7310563c9cbeee15b532ed6f3f9e2&sharer_shareinfo_first=ec8cc7e96fbe03139927b df1828909fe#rd) believes that the "Jevons Paradox" continues to be effective. **The improvement in computing efficiency will not reduce the demand for computing power; rather, it will stimulate the emergence of more application scenarios, driving the continuous growth of computing demand. Just as Barclays predicted last June, by 2026, the industry's capital expenditure will be sufficient to support "12,000+ ChatGPT-level applications."

Moreover, the logic of computing demand may be questioned in the short term. Some companies, especially the CIOs or CFOs of large overseas firms, may face pressure from investors and bosses to explain why their ROI is far lower than that of DeepSeek.

Well-known investor Duan Yongping also stated on Xueqiu that DeepSeek's experience indeed proves that lower computing power during the model pre-training phase can achieve relatively good training results. He also agrees with Jensen Huang's statement, believing that DeepSeek's innovation will not reduce the demand for computing power.

Previously, Jensen Huang stated in an interview in February that the market's understanding of DeepSeek is completely reversed. He stated that the emergence of R1 does not mean that the market no longer needs computing resources, but rather stimulates the market's pursuit of more efficient AI models, thereby promoting the development of the entire industry.

However, foreign tech blogger Zephyr believes that DeepSeek has already "knocked down" NVIDIA. Moreover, based on DeepSeek's current ultra-high utilization of computing power, it is more than enough to meet the global AI demand.

DeepSeek has already "knocked down" NVIDIA.

The reason I say this is that DeepSeek is currently processing 600 billion tokens daily on 300 H800 nodes (a total of 2,400 H800s) and outputting 150 billion tokens.

If computing power expands 100 times (i.e., 240,000 H800s), it could process 60 trillion tokens daily and output 15 trillion tokens.

But the global AI demand is not that high

3. What does it mean for the cloud industry chain?

The success case of DeepSeek makes the similarities between AI cloud computing and traditional cloud computing more apparent. AI cloud computing will also face challenges such as "idle rate during low peaks" and "stability during high peaks."

Tech blogger 180K believes that the scale effect of cloud computing will be more significant. DeepSeek's practice shows that large-scale clusters and high concurrency utilization can significantly reduce costs. The positive externality of the number of users is more evident; that is, the more users there are, the stronger the ability to smooth fluctuations and the lower the demand for computing power redundancy.

The competitive advantages of cloud vendors may change. Cloud vendors with their own businesses (such as Alibaba, Tencent, Apple, etc.) may have more cost advantages than those without their own businesses, as they can use inference clusters as the foundation for all their businesses, achieving greater scale effects.

Moreover, there is room for improvement in the profitability of cloud computing. DeepSeek's case indicates that in the AI era, through extreme infrastructure optimization, there is potential for further improvement in the profitability of cloud computing.

In addition, the attractiveness of private cloud deployment may decline. Ultra-sparse MoE models may not be suitable for individuals or "half-baked" enterprises, and the cost of small-scale GPU deployment may far exceed that of large companies. This may lead more enterprises to choose public cloud or hybrid cloud models.

Ordinary cloud computing/AI applications need to reserve more space for high-intensity user concurrency. Users have a higher tolerance for DeepSeek's "server busy" message, but this is not the case for other applications. This may lead to a further decline in the profitability of ordinary cloud computing/AI applications.

4. What does it mean for peers in large models?

The open-source and technology disclosure of DeepSeek sets a new benchmark for the entire industry.

Tech blogger Information Equality believes that DeepSeek's case shows that the "bottom line" of inference costs has been significantly lowered, far below previous expectations. Moreover, a new round of price wars may erupt, and peers will face greater pressure to reduce prices to maintain competitiveness.

Furthermore, DeepSeek provides all inference teams with clear optimization paths and goals, increasing subsequent pressure.

In addition, under these circumstances, OpenAI's high-priced subscription model will also face challenges, as the high subscription fee of $200 per month is somewhat awkward.

5. What does it mean for the ecosystem?

DeepSeek's strategy is to focus on foundational models and cutting-edge innovations, attracting the industry to build B2B and B2C businesses on its foundation through open-source technology and output, forming a complete upstream and downstream industry.

[Tech blogger Geek Park states](https://mp.weixin.qq.com/s?__biz=MTMwNDMwODQ0MQ==&mid=2653074804&idx=1&sn=a75fc037dcb5ad97464fc7fedac9a6b5&chksm=7f3fc0f5ba9fda7a7b4a1e8de4a7954866df6c6f723174baee0546cd0b3a756e39cb16651d49&mpshare=1&scene=23&srcid=0302eh4qfFXNz1ZfaY1 The profit margins for ecological partners are increasing. By deploying DeepSeek's services, cloud platforms and upstream and downstream partners can theoretically achieve high returns and profit margins.

Looking ahead to the subsequent ecosystem, the differentiation of model architecture may become a key competitive factor. This is because the architecture of DeepSeek V3/R1 differs significantly from mainstream models, requiring suppliers to adapt, which poses a high development difficulty.

Moreover, DeepSeek's open-source initiative has reduced the difficulty for the community to replicate its inference system, which is beneficial for the prosperity of the ecosystem.

Tech blogger 180K stated that the entire industry may begin to focus on Infra. To some extent, the importance of Infra is increasing, and valuations can also rise.

In summary, DeepSeek's ultra-high profit margins are not only a numerical miracle but also a profound revelation for the entire AI industry. It reveals the enormous potential of infra optimization, drives the transformation of computing power, cloud, large models, and ecosystems, and heralds the arrival of a more efficient, low-cost, and fiercely competitive AI era