First, the shortage of computing power has been urgently solved
A major consensus in the artificial intelligence industry is that computing power is as important to AI as lithium mines are to smart cars. As a real resource and scarce product in the AI industry, AI computing power has been fully verified under this industry trend. As an important supplier of AI infrastructure, NVIDIA's explosive performance and increased demand fully prove this.
With the recent surge in AI, the industry's demand for GPUs is strong, and with the surge in demand, GPUs have fallen into a supply shortage, and many AI founders expect the shortage to continue at least until next year. And even if you pay to grab the production capacity, you may face the problem of waiting for more than 6 months. And as demand increases, Musk expects demand to increase by as much as 3-5 times.
Therefore, the shortage of computing power has been urgently solved.
Second, NVIDIA DGX Cloud is expected to land in China, and the strongest computing power undertaker opens up GPT 100 billion space
1. At the NVIDIA GTC conference, CEO Jensen Huang announced the DGX Cloud plan, which will open supercomputers for the development of artificial intelligence technologies such as ChatGPT to commercial customers, and almost any enterprise can use cloud leasing to use these powerful but expensive equipment; In the future, building a large model is no longer exclusive to giants, and computing power benefits the whole society in the form of cloud services!
2. Under the current international relations, the leasing model can bypass the embargo restrictions, and the future space is huge: the United States embargoes NVIDIA high-end GPUs, and the corresponding exports need to be licensed by the US government, the core reason is that it is worried about military use. NVIDIA plans to open computing power in the form of cloud services, and the tenant form ensures that the relevant computing power is in the hands of cloud service providers, without worrying about the downstream use direction, the most beneficial is China-related enterprises, can obtain mutual trust, and the future computing power bottleneck is completely opened!
Third, who needs computing power leasing?
Similar to the IDC and cloud computing markets in 19-20 years, small and medium-sized enterprises that generally lease computing power are small and medium-sized enterprises with data elasticity and certain difficulties in production environment, investment capacity and hardware level.
We take 175B parameters and 300B tokens training set size as reference; DGX A100 (5petaFLOPS), 45% training effectiveness; 30 days of training is the training standard for medium-sized enterprises. (There are a large number of research reports and calculations, not expanded)
We can simply get such a model that requires about 55 DGX A100 servers, according to the market price of 200,000 US dollars, then only considering the server procurement cost needs 11 million US dollars (78 million yuan), which is obviously a huge investment for a small and medium-sized AI company, not considering factors such as computer room construction, operation and maintenance, and replacement impairment.
4. How much does the lease cost?
Assuming that domestic pricing is consistent with CoreWeave's, a GPU on an A100 80GB NVLINK costs $2.21/hour.
The single cost of training a model for one month is:
2.21*24*30*8*55≈ 700,000 USD (5 million RMB)
So the core question is: how many times a year does a small and medium-sized business need to train a large model? If the industry in which the model is located is relatively stable and has rich historical data accumulation, then after the first training + first update, theoretically no major changes are required to enter the inference stage, then this only requires about 10 million yuan, which is 1/7 of self-built, and does not need to bear factors such as computer room construction, operation and maintenance, and replacement impairment.
For rental operators:
Using a single server as a unit of calculation, assuming that its cost structure is similar to the cost structure of IDC, the server procurement cost accounts for 65%, a single server is 200,000 US dollars, and the maintenance time of 1 year is 15 days, then the first year of revenue is:
2.21*24*8*350 ≈ USD 148,500 (RMB 1.05 million)
1. Considering the performance improvement factor, the cost per unit of computing power decreases by 20% per year, then the annual revenue of A100 decreases by 20%, and the revenue decreases by 30% due to performance reasons after 4 years;
2. The server is calculated according to 20% of the residual value of 4 years, and the land and other equipment are calculated according to 10-year depreciation without residual value, which is unified as the average depreciation method, but the revenue life cycle is 8;
3. According to the parameters of DGX A100, the system power consumption is 6.5KWh, which is calculated according to the cost of KWh of 0.65 yuan for convenience, and the PUE is 1.1;
4. The three rates refer to Runze/Dataport, etc., regardless of financial costs, and are calculated at 8%;
5. The discount rate is calculated according to 10%.
The final calculation yielded a net cash return of US$105,200 (750,000 yuan) over the 8-year life cycle of the unit server, and the pre-tax IRR was roughly 35%.
5. Sorting out the concept stock of computing power leasing benefit
1. Hongbo shares: has signed a 1280P computing power order with Baichuan Intelligence
InBev Digital currently signs a computing power leasing service contract with Wuji Intelligent (Beijing) Technology Co., Ltd., and InBev Digital will provide high-performance GPU computing power rental services to Wuji Intelligent within 12 months after the signing of the agreement, with no more than 256 servers delivered in the first phase, with a total computing power of not less than 1280P. InBev Mathematics generated operating income and profit in early May.
2. Litong Electronics: Increase capital by 450 million yuan and operate AI computing power related projects
The company intends to cooperate with Shanghai Century Longteng Data Technology Co., Ltd. (hereinafter referred to as "Century Longteng") and sign a cooperation framework agreement, by Century Longteng set up a project company "Century Litong" in Shanghai Songjiang, with a registered capital of 50 million yuan, the main business is to provide AI computing power leasing services, the AI computing power leasing business of the project company is all deployed in Tencent Yangtze River Delta Artificial Intelligence Supercomputing Center in Songjiang District, Shanghai and Shenzhen Yifeng Technology Bright 5G Big Data Center in Guangming District, Shenzhen. And Century Longteng or a third party recognized by both parties to provide a full set of IDC hosting and operation and maintenance management services.
3. Ningbo Construction Engineering: computing power leasing + equity participation in CEICC Cloud Data Storage Technology (Beijing) Co., Ltd. + data center + innovative optical-magnetic hybrid storage architecture + edge computing
The Company and its wholly-owned subsidiaries invested in CEICC Data Storage Technology (Beijing) Co., Ltd., whose main business is data center operation, and directly and indirectly held 32.07% of its equity. At present, CEICC mainly provides data center service business combining software and hardware, and has passed Alibaba's certification, signed service contracts and provided cabinet leasing services.
4. Zhongrun Resources: The main development direction of the wholly-owned subsidiary is computing power leasing
Xide Shende Cloud Computing Technology Co., Ltd. (hereinafter referred to as Xide Shende Cloud Company or the Guaranteed Party) is a key investment project company in Liangshan Yi Autonomous Prefecture, Sichuan Province, with big data center construction, server operation and hosting services, computing power leasing and sale as the main development direction, mainly undertaking the server hosting business and computing power sale business of Xide County People's Government and large state-owned enterprises and central enterprises.
5. Runjian shares: Wuxiang Cloud Valley cloud data center includes data storage and computing power leasing
The company takes the communication operator IDC and computing power network management and maintenance requirements as the starting point, has completed the whole chain layout of IDC service business, has a number of digital SAAS platforms, and strong technical implementation and operation and maintenance teams, can provide IDC with one-stop services for the whole life cycle, including IDC consulting and design, construction, operation and maintenance, energy-saving transformation, operation, etc., as well as cloud consulting, cloud integration, cloud management, cloud operation and maintenance and other value-added services, and gradually form basic telecom operators as core customers, financial institutions, Government and enterprise institutions provide professional and standardized IDC and computing power comprehensive services for major customers.
6. Qingyun Technology: the first hybrid cloud listed company in China
The company's QingCloud EHPC is the company's elastic high-performance computing products, including CPU, GPU and other computing power services, in terms of GPU cloud server, the company can provide customers with NVIDIA A100 GPU cloud host, in order to meet the needs of high-performance computing, a single host can be configured with 8 GPU cards, support large-scale computing tasks, while building a GPU resource pool, the comprehensive utilization of GPU to 3~10 times.
7. Nanling Technology: layout computing power service
As of December 31, 2022, the company has built 67 business nodes around the world, involving 36 cities in 23 provinces, municipalities directly under the central government and autonomous regions in China, and 4 regions abroad, of which 6 edge computing nodes have been deployed, involving domestic and overseas regions. The company can provide enterprise-level CPU and GPU computing power services.