GuShiio.com Article source: First Finance and Economics
Image source: Generated by AI
DeepSeek brings immediate returns to cloud computing vendors.
“Basically, the cloud computing power inventories of major cloud vendors, including operators, have been cleared.” A researcher from a large model research institution told reporters. This organization reached this conclusion after surveying domestic cloud vendors.
On January 20, after the release of the DeepSeek-R1 model, various cloud computing manufacturers launched their full-blood or distilled versions of models. It is open source and relatively efficient, which has promoted the further increase in the penetration of artificial intelligence in China and promoted cloud computing vendors to increase investment in infrastructure.
“All the cloud computing power has been eaten up.”
DeepSeek is now the focus of competition among cloud computing vendors.
“DeepSeek is the best touchstone. Which cloud has good hosting service performance, high throughput, and low latency can be known directly.” Tan Tan, president of Volcano Engine, said in his circle of friends.
The models released by DeepSeek include multiple versions, including full-blood versions with 671B parameters and small models with 1.5B, 7B, and 70B parameters. Different parameters require different computing resources, and different terminals are deployed. For example, the 1.5B model can be deployed on mobile phones worth several thousand yuan, while the 70B model can be deployed on workstations worth tens of thousands of yuan, and the 671B full-blood model is more cost-effective when deployed in the cloud.
Volcano Engine is a cloud and AI service provider owned by ByteDance. In addition to supporting its own large model of bean buns, it also launched multiple series models such as the Dark Side of the Moon and Wisdom Spectrum. Just after the Spring Festival, the Volcano Engine launched a size DeepSeek model. The DeepSeek model is open source and easy to use. Enterprise users who want to use the full-scale model in the cloud need to purchase computing power or storage services from cloud computing vendors.
Tan Dai’s statement was a message to corporate customers and also to other cloud computing vendors.
The competition among cloud computing vendors around DeepSeek started during the Spring Festival. In order to launch DeepSeek as soon as possible, many engineering and technical personnel spent a non-stop Spring Festival holiday. Competition among cloud computing vendors is already fierce, and price wars have been fought round after round. Therefore, Tencent Cloud has concerns when accessing Alibaba’s big model. Conversely, Alibaba Cloud will not wholeheartedly accept Tencent’s mixed model, and Baidu Intelligent Cloud will not favor Tencent or Ali’s big model.
The open source DeepSeek seems harmless to humans and animals, but is beneficial to cloud computing vendors, so it has gained wider acceptance. China’s technology companies have released hundreds of large models, but no product is as penetrating as DeepSeek and can jump lightly on every cloud.
DeepSeek has also driven the wave of open source.
Baidu was once a staunch supporter of the closed-source approach, but it also announced after the year that it would fully open source its Wenxin Model 4.5 series on June 30. Baidu Smart Cloud has also connected to the DeepSeek model and attracted developers to use it at a lower price.
On the 2025 Step Star Ecology Open Day, Step Star, one of the six big model tigers, announced that it will open source its graphic video model in March this year. Previously, Step Star has released 11 multimodal large models, two of which are open source products.
“Step has planned a roadmap for realizing AGI since its inception. Specifically, there are three stages: simulating the world, exploring the world and summarizing the world.” Jiang Daxin, founder and CEO of Step Star, said: “Step work is carried out along these three stages. Over the past year, we have released a series of models, forming a range of hundreds of billions of parameters to trillions of parameters, from language models to multimodal models, from multimodal understanding to multimodal generation, from imitation learning to reinforcement learning. The comprehensive model matrix covers multiple key nodes in this roadmap.”
Once a model is open source, it has its own vitality, and the open source model ecosystem can stimulate developers ‘enthusiasm for using it.
China’s big model industry is undergoing structural changes. Some companies such as One and One Things have abandoned the costly research and development of general big models and turned to vertical industrial models with lower input costs and faster commercial results.
After the DeepSeek-R1 model ignited market enthusiasm, the effectiveness of the “Jevons Paradox” was verified. DeepSeek points to lower computing power requirements, but the improvement in the efficiency of computing power use has enabled inference models to quickly penetrate into thousands of industries, thus driving the overall growth of computing power requirements.
“Models are updated too fast. The best models must be on the cloud.” Tan said in an interview at the end of 2024.
OPPO phones have been connected to the large model of Step Star, and Xiaohongshu has been connected to DeepSeek to enhance its search capabilities. Volcano Engine said that it has assisted 60 financial institutions including Guosen Securities, Ping An Securities, and E Fund Fund in accessing the DeepSeek model.
“The release of the inference model has led to a large number of players pouring in, consuming all the cloud computing power.” The above-mentioned researcher told reporters.
Compared with leading companies such as Alibaba Cloud and Tencent Cloud, cloud computing companies at the waist and below are in urgent need of such opportunities to change the decline in revenue.
Ucede’s revenue in 2021 will reach 2.9 billion yuan, but its revenue will decline continuously in the following two years, reaching only 1.5 billion yuan in 2023. In the mid-year report of 2024, Ucede’s revenue also fell to 730 million yuan, and a loss of 100 million yuan occurred. After revenue of 429 million yuan in 2020, Fastcloud’s revenue is also sluggish, with revenue of 330 million yuan in 2023.
In this round of revaluation of China’s technology assets, the market values of Ucede and Fastcloud have doubled since the beginning of 2025, but now they also need to double their real performance.
New investment cycle?
DeepSeek is stimulating a round of investment in cloud and AI infrastructure among China’s major technology companies.
Alibaba Group CEO Wu Yongming announced on February 24 that more than 380 billion yuan will be invested in cloud and AI hardware infrastructure in the next three years. This investment exceeds Alibaba Group’s total investment over the past ten years.
On February 25, Baidu announced that it would acquire Huanju Group’s domestic video entertainment live broadcast business YY Live for approximately US$2.1 billion. After the new agreement was reached, approximately US$1.6 billion previously deposited in the escrow account has been fully returned to Baidu. Baidu said it will invest US$1.6 billion in cloud and AI infrastructure construction.
In January this year, the first phase of Mobile Cloud’s Suzhou Fenhu Intelligent Computing Center was put into operation, with a total investment of 2.68 billion yuan in the first phase. Mobile Cloud’s Suzhou project has a total planned investment of 6 billion yuan. When completed, it can carry 24,000 cabinets and an installed capacity of 500,000 servers. It can finally reach 10EFLOPS.
“The computing power reserve of the volcano engine must be very sufficient, which is also an important reason why we are confident of undertaking industry services on a large scale, at low prices and with high throughput.” Tan Taizu told reporters in an interview at the end of 2024.
Compared with companies such as Alibaba Cloud and Baidu Cloud, Volcano Engine is a latecomer to the cloud computing market. “If you want to do faster, you must be first in new technological changes. What are the new technological changes? It’s actually AI and big models, so our goal in this area is first, and we are moving in this direction.” Tan Dai said.
Domestic computing power is expected to take this opportunity to embed deeper into the artificial intelligence ecosystem.
“The explosion of DeepSeek has a strong demand for our computing power.” Mobile Cloud expressed to First Finance and Economics. Mobile Cloud will make some fine-tuning and distillation of DeepSeek’s model to meet the needs of enterprises in specific industries. Mobile cloud provides a fully localized deployment environment. This can especially impress some central state-owned enterprises, whether they are deployed locally or in the cloud.
According to Mobile Cloud, the computing power it provides largely meets the needs of state-owned enterprises. In February 2025, Mobile Cloud assisted PetroChina China in completing the training, promotion, adaptation and privatization deployment of DeepSeek’s full-stack localization, which is inseparable from the support of Shengteng’s computing power.
Nvidia’s GPUs meet more than 90% of the needs of cloud computing companies, but the emergence of DeepSeek has made domestic computing power possible. Muxi, Moore Thread or Huawei Shengteng GPUs are being accepted by more and more cloud computing platforms or terminal companies.
China chip companies have also seized this opportunity to take advantage of the momentum.
Shortly after the DeepSeek series of models were launched, Kunlun Core completed the full version adaptation, and announced that it had lit up its new generation product P8 million card cluster on the day of the launch of the Year of the Snake; on February 20, Kunlun Core officially announced that it would implement a full-blooded version of stand-alone deployment on the P800.
Since DeepSeek launched “Open Source Week”, it has successively opened up four codebases. Moore Thread has now completed the adaptation of DeepSeek open source libraries FlashMLA and DeepGEMM.
The United States has imposed restrictions on China’s imports of high-end chips, which has forced China companies to achieve technological breakthroughs. At present, Nvidia GPUs have advantages in computing power performance and ecological maturity. However, with the advancement of domestic chip research and development, the improvement of algorithm capabilities, and the industry’s emphasis on supply chain security, there is hope to gradually reduce the risk of single dependence on Nvidia in China.
“This process requires time and technology accumulation, but it is an irreversible trend. Future computing power bases are more likely to take on the form of ‘diverse coexistence’ rather than the absolute dominance of a certain manufacturer.” Moore thread previously told reporters.
