"The order has to wait for half a year"! Large-scale models exploded, GPUs fell into supply shortage, computing power rental demand rose

Author: Fu Jing

**Source: **Financial Association

Image credit: Generated by Unbounded AI tools

"They can't wait for (GPU), otherwise the research will not go forward." Zhang Yazhou, the founder of Shanghai Liuchi Technology Group, revealed to reporters that at the other end of the AI manufacturers who come up with large models to "show their muscles", the training models used by big manufacturers However, the computing power resource GPU has fallen into a shortage of supply. Even if the payment grabs the order, it may have to wait for the first half of the year to get the goods-under this background, a more cost-effective "computing power leasing" model is emerging.

Since June, the A-share computing power leasing sector has continued to be active. Century Huatong (002602.SZ), Litong Electronics, and GCL Nengke have successively announced the "cross-border" layout. IDC manufacturers Runze Technology and Philips have also added "computing The concept of “power leasing”.

A few days ago, reporters from the Financial Associated Press learned from multiple interviews that the current market demand for computing power leasing and market orders for renting computing resources are on the rise, mainly in the fields of artificial intelligence, machine learning, and big data analysis. There is a large room for future growth, but two Many companies in the high-end market are suspected of rubbing the heat.

GPU is in short supply and computing power leasing is on the rise

Recently, when various AI manufacturers have come up with large models to "show their muscles", many observers told reporters from the Financial Associated Press that the rapid development of big data, artificial intelligence and other fields has driven the demand for computing power (especially large models and industry applications, etc.) )Rapid growth. It is understood that GPU demand is strong and has fallen into a supply shortage.

"Unlike consumer graphics cards, many merchants have inventory. Its inventory is very small, and the goods in the entire supply chain do not exceed three or four thousand pieces. After GPT became popular, it was completely locked up by major Internet manufacturers. After March this year, the market There is basically no stock in the market, and some major manufacturers have started to find some goods directly from foreign channels, and now more than 50% of the orders are basically placed by Internet major manufacturers." Zhang Yazhou said in an interview with a reporter from the Financial Associated Press.

Zhang Yazhou told reporters from the Financial Associated Press that major overseas companies, including Microsoft and Intel, are doing their own research and purchasing externally. Part of their pre-orders (GPU orders) may be used for large-scale training, which is not enough for reasoning. Basically, if a manufacturer orders less than 100 million yuan, the channel dealer may not be willing to respond to it, and they will not reply to emails. "The popularity of GPU can be seen from this.

From the price point of view, the prices of various GPUs have increased by nearly 50% since March and April. Zhang Yazhou said, "In April (orders) can probably be scheduled until October, and now it is the first half of next year."

Zhang Hongrun, an expert on the Whale Platform and founder of Jiandian Workshop, told the reporter of the Financial Associated Press, "Many companies and individuals need to use high-performance GPUs for deep learning training and applications, but the cost of purchasing servers to build data centers is very high. , so it is more cost-effective and flexible to choose a cloud computing platform for renting.” Based on this logic, the computing power leasing model of renting computing resources is currently emerging.

Listed companies get together and rush into real demand or hype?

According to the reporter’s analysis, Hongbo shares, Qingyun Technology-U, Dongfang Guoxin, Zhongke Sugon, Youkede-W, Capital Online, Runjian shares, Shunwang Technology, Zhongqingbao and other listed companies including traditional IDC manufacturers They have all aimed at the computing power leasing business, and several "new faces" have poured into the track this month.

On the evening of June 1st, Century Huatong and Litong Electronics made an official announcement: Century Longteng, a subsidiary of Century Huatong, signed a cooperation framework agreement with Litong Electronics to establish Shanghai Century Litong Data Service Co., Ltd. to provide AI computing power leasing services.

On the 8th, GCL Nengke announced that it signed a strategic cooperation framework agreement and an investment framework agreement with Jiuzhou Cloud and Gufan Technology, a subsidiary of Jiuzhou Cloud, respectively. It is reported that Kyushu Cloud has successfully built and operated the Qingtian·Innovation Empowerment Center (Powered by NVIDIA) through Gufan Technology. The center is built on the basis of NVIDIA RTX and includes the NVIDIA Display Experience Center. As a result, Koe indirectly has a relationship with Nvidia."

Jiang Han, a senior researcher at the Pangoal Institute, told the Financial Associated Press that manufacturers related to computing power leasing have met some of the needs, but there are still some gaps. Wu Gaobin, vice president and founding secretary-general of the Metaverse Working Committee of the Chinese People's Association, also said: "Different manufacturers can provide different scales and types of computing power leasing services based on their own technology and equipment advantages, but there are also some situations where the supply is insufficient. .”

However, Zhang Yazhou described the status quo of the industry he observed to reporters from the Financial Associated Press in this way: "The production capacity of large factories cannot keep up, while the 'small factories' Wang Po sells melons and boasts, and many of them are just chasing the heat."

It is worth noting that the computing power leasing business of the aforementioned "cross-border" listed companies has not yet officially started. The reporter called GCL Nengke and learned that "the first project should be launched around August." In addition, a person from the Litong Electronic Securities Department told the reporter of the Financial Associated Press, "We have only signed a preliminary cooperation framework. When will we start? The operation has to wait for the decision of the leadership.” In contrast, Hongbo shares, an AI bull stock that has increased by more than 400% this year, is progressing faster, and its Beijing AI Innovation Empowerment Center project has landed in the first phase, and signed a 1280P order with Baichuan Intelligent .

Zhang Yazhou said, "To judge whether to speculate on a concept depends on whether the manufacturer really has customers and which customers are using it. Many manufacturers have not announced what kind of customers they have. At present, Hongbo shares have clearly stated that they have signed an order with Baichuan Smart. But the card is not enough, already 'overdrawn'."

"In addition, some data center manufacturers have recently added the concept of computing power leasing to themselves, which is also suspected of being hyped. Data centers are old assets, and they must be transformed into smart computing centers if they want to transform." Wu Gaobin also told reporters that computing power is redundant. Redundancy, lack of flexibility, and difficulty in equipment maintenance are the problems faced by the traditional model of purchasing servers to build data centers.

It is expected to become one of the mainstream models of computing power supply

Many industry insiders interviewed by reporters from the Financial Associated Press unanimously stated that the computing power leasing market has a broad space.

Jiang Han said: "With the rapid development of AIGC, the global market share of computing power leasing continues to expand, and it will usher in a period of rapid development in the next few years."

Zhang Hongrun also said, "As the deep learning model becomes more and more complex and large, the demand for computing power is getting higher and higher. And with the advancement and popularization of cloud computing technology, computing power leasing will become more convenient and affordable. So I think computing power leasing is a promising and valuable field, and more users and manufacturers will flood into this market in the future.”

It is understood that compared with the traditional self-built data center model, computing power leasing has many advantages.

Xiao Qiaoqiong, founder of Saijianni Brand Consulting, told the reporter of the Financial Associated Press: "Compared with the traditional mode of purchasing servers to build data centers, computing power leasing has multiple advantages such as strong flexibility, low fixed costs, and good professional technical support. One of the mainstream models of computing power supply.”

"Self-built data centers have a long asset investment cycle and need to continuously upgrade hardware, while computing power leasing can dynamically adjust computing power according to business needs at any time, and can effectively optimize cost structures." Angel investor and senior artificial intelligence expert Guo Tao said .

According to Wu Gaobin, “The core costs mainly focus on server R&D, procurement, maintenance and upgrades, etc., which can be controlled through specific technologies and management methods, such as intelligent management. Cost management, strategic cooperation with chip manufacturers to reduce procurement costs, etc.” In addition, Jiang Han said that the initial cost situation varies with the size of the company, large manufacturers usually use centralized procurement and economies of scale to reduce costs, while small manufacturers mainly Meet customer needs through customized solutions.

From the perspective of users, Zhang Hongrun told reporters from the Financial Associated Press that based on computing power leasing, users only need to pay on demand, and do not have to bear the cost of purchasing, maintaining, and upgrading hardware equipment, and do not have to worry about waste caused by idle or outdated equipment; Access the required computing power resources through the cloud anywhere, and quickly start training and application; users can choose different computing power platforms and models according to their needs, and can also make more attempts without geographical or time constraints on models, tools and other resources and explore.