How are the cross-border computing power businesses of more than 10 A-share comp

Over 10 A-share listed companies have announced their entry into the cross-border computing power business in the past two years. Among them, two companies, within less than half a year after their official announcements, have already experienced changes in their computing power business.

In the past year, large models have sparked competition in the computing power industry. Both training and inference of large models consume intelligent computing power. As a result, Chinese technology companies and telecommunications operators are procuring AI servers on a large scale.

In addition to the aforementioned professional computing power service providers, since 2023, at least 10 A-share listed companies have entered the field across different industries. These include Hongbo Shares, Jinji Shares, Lianhua Health, Hangjin Technology, Jinzi Ham, An'er, Huafu Fashion, Qunxing Toys, Anuoqi, and Hengli Home Furnishings. They are involved in industries such as printing, chemical, food, home furnishing, property, clothing, dye, and textile.

In October 2023, the United States imposed export controls on advanced AI chips to China. The export controls have made the computing power business more difficult. Theoretically, this should have "discouraged" these companies. However, among the aforementioned 10 listed companies, only Hongbo Shares, Jinji Shares, and Lianhua Health entered the market before the controls took effect. The other seven companies entered after the controls were in place.

Facing bottlenecks in their main business, searching for a second growth curve, and using the computing power concept to boost stock prices—these are more or less the reasons for the above companies to cross over.

We have compiled the performance of the aforementioned 10 A-share listed companies over the past three years. Among them, only Hengli Home Furnishings has seen a positive increase in revenue for three consecutive years, while other listed companies have experienced negative revenue growth. Jinzi Ham and An'er have seen negative revenue growth for three consecutive years. Hongbo Shares and An'er have had negative net profits for two consecutive years.

The computing power business is not a surefire profit, with high capital and technical barriers. Most of the aforementioned companies entering the field lack industry knowledge and technical accumulation, and there are many hidden dangers on the cross-border road. Some companies have encountered difficulties in computing power supply, while others have announced the termination of their cross-border plans. Several companies have also been questioned by regulatory authorities for issues such as non-standard information disclosure and inadequate risk warnings, and have even been ordered to rectify.By the end of 2023, Lotus Health and Jinji Shares successively announced that there were difficulties in delivering the previously purchased Nvidia H800 GPU server orders, necessitating the search for new sources of supply. In April 2024, An'er announced the termination of the acquisition of the computing power service provider Shenzhen Innovation Science. In the same month, Hengli Home Furnishing also terminated the computing power business that had only been officially announced for two months.

We have sent requests for comment on why they are venturing into computing power, which is unrelated to their main business, to Hongbo Shares, Jinji Shares, Lotus Health, Hangjin Technology, Jinzi Ham, An'er, Huafu Fashion, Qunxing Toys, Annoqi, and Hengli Home Furnishing.

Among them, Jinji Shares responded positively, stating that further announcements would be made once the progress of the computing power business reaches the disclosure standard. Annoqi and Jinzi Ham indicated that the investment in computing power is a financial investment, and information should be primarily based on announcements. Hangjin Technology and Huafu Fashion did not respond. As of the time of writing, we have not been able to reach the board secretaries of Hongbo Shares, Lotus Health, An'er, Qunxing Toys, and Hengli Home Furnishing by phone.

There are many questions from the outside world about the aforementioned listed companies, mainly focusing on a few points.

Firstly, why venture into the computing power business, which is far from the main business?

Secondly, what are the risks of venturing into computing power?

Thirdly, what stage have these companies reached in their business progress, and is it sustainable?

Why venture into new fields?

Among these 10 A-share listed companies, some are in industries that experienced sluggish growth in 2023.

Data from the National Bureau of Statistics shows that the business income of printing and dyeing enterprises above designated size in 2023 was 298.6 billion yuan, a year-on-year increase of 1.4%. In 2023, the industrial added value of the clothing industry's enterprises above designated size decreased by 7.6% year-on-year. The business income of furniture manufacturing enterprises above designated size decreased by 4.4% year-on-year in 2023.This has brought performance pressure. The aforementioned companies generally have high hopes for the computing power business. They hope to achieve high revenue and high profits through cross-industry expansion. Hongbo Shares, Hangjin Technology, Huafu Fashion, Qunxing Toys, Anuoqi, and Hengli Home Furnishings have all expressed similar views in their company announcements - the purpose of developing computing power business is to increase company revenue and enhance profitability.

The secretary of Jinji Shares, a dye manufacturer, frankly told us, "The span of computing power is very large, and the company made up its mind after repeated consideration. The national demand for the original main business is relatively fixed, and it is definitely not feasible not to change." Hongbo Shares' 2023 annual report mentioned that the traditional printing industry is facing many challenges, including market saturation, fierce competition, and low gross profit margin. The company is laying out new fields such as AI computing power to achieve diversified development of the company's business and reduce dependence on a single market.

The forms of cross-industry computing power of the aforementioned companies are three: first, bare-metal leasing, second, self-built intelligent computing centers, and third, investment or acquisition of other computing power enterprises.

Bare-metal leasing refers to the direct leasing of AI servers purchased to the outside. This business does not require networking or cloudization, and is the most primitive and simple business form in the computing power market. Lianhua Health, which is mainly engaged in monosodium glutamate, has adopted the form of bare-metal leasing.

Self-built intelligent computing centers refer to the networking and cloudization of AI servers purchased by enterprises, and then leasing them to the outside. This requires recruiting a professional technical team responsible for software and maintenance. Currently, Hongbo Shares, Jinji Shares, and Hangjin Technology have adopted this method.

The approach of investment and acquisition is that the company sends people to participate in the major decisions of the invested enterprise, and does not participate in daily operations. The common practice is for the company to sign a performance commitment letter with the invested enterprise to ensure investment returns. Jinzi Ham and Anuoqi have adopted this strategy.

To do the computing power business, it is necessary to purchase AI servers. In October 2023, the United States implemented export control on advanced AI chips to China. American chip companies such as Nvidia were prohibited from selling advanced AI chips to Chinese enterprises. The Nvidia A800/H800 chips most commonly used by Chinese enterprises to train large models have been cut off.

Control has made the business more difficult, and theoretically it should "discourage" these enterprises. However, of the aforementioned 10 listed companies, 7 announced their entry after the control took effect. One explanation is that export control has led to a rapid increase in the price of computing power. The same risk has also brought a wind, and this business is considered profitable.

After the control took effect, the main way for Chinese enterprises to obtain advanced AI chips is two: one is the inventory products of server distributors, and the other is the products that enter the country through overseas multiple transfers. Against this background, the increase in A800/H800 servers once exceeded 50%. A server agent said that in November 2023, the H800 server was 3.3 million yuan/piece, and the A800 server was 1.8 million yuan/piece. As long as there is goods, there is no worry about no buyers.Another server distributor stated in January this year that in the H800/A800 rental market, the monthly rent for an 8-card H800 server ranges from 90,000 to 110,000 yuan. Customers also have to pay for cabinet fees and operation and maintenance costs. The usual contract period is three years, and theoretically, the investment can be recouped within 2 to 3 years. Universities, research institutions, manufacturing companies, financial institutions, and even small and medium-sized AI startups will rent sporadically.

Against this backdrop, the rental prices for NVIDIA A800/H800 servers have risen. Its profits are enticing. Dongwu Securities calculated the A800 rental market data in September 2023, estimating that the theoretical gross margin for A800 server rentals is about 40%, with the H800 having an even higher gross margin.

The gross margin of A800/H800 server rentals is even higher than that of most of the aforementioned 10 listed companies. Nine out of these 10 listed companies had a gross margin below 40% in 2023. Only the apparel company An'nai had a gross margin higher than the A800 rental business.

Profit margins from computing power services disclosed by some listed companies also show that they are higher than the company's main business. For example, according to the financial report of Lianhua Health in 2023, the revenue from computing power was 765,000 yuan, with a gross margin of 58.6%. The company's comprehensive gross margin for 2023 was 17.2%. Hangjin Technology's announcement showed that in 2023, the revenue from computing power was 130 million yuan, with a net profit of 11.151 million yuan, and a net profit margin of 8.9%. The company's comprehensive net profit margin for 2023 was 2.9%.

After announcing their intention to enter the computing power business, the share prices of several listed companies rose. Hongbo Shares reached its peak for the year on August 22, 2023, with a closing price of 45.29 yuan, a surge of 552.6% compared to the closing price on January 3, 2023. Lianhua Health reached its annual peak on October 13, 2023, with a closing price of 7.89 yuan, a surge of 214.3% compared to the closing price on January 3, 2023. Jinji Shares reached its annual peak on December 1, 2023, with a closing price of 12.26 yuan, a significant increase of 77.7% compared to the closing price on January 3, 2023.

Hongbo Shares, Lianhua Health, An'nai, and Qunxing Toys have issued multiple stock trading fluctuation anomaly announcements. The trigger for this announcement is that the company's closing price has deviated by a cumulative value of 20% over three consecutive trading days. Among them, Hongbo Shares issued 14 stock price fluctuation anomaly announcements from February 9, 2023, to February 27, 2024. The "necessary risk warnings" in the announcements all mentioned the computing power business.

The significant increase in stock prices has objectively brought huge benefits to the executives and major shareholders of the listed companies. During the period of rising stock prices, some executives and major shareholders of the listed companies chose to reduce their holdings and cash out. From January to August 2023, the stock price fluctuation range of Jinji Shares was 6.1 yuan to 7.2 yuan. After announcing in August 2023 that it would operate the computing power business, Jinji Shares announced four times the progress of shareholding and executive reductions, with an average reduction price range of 6.1 yuan to 11.9 yuan.

Where lies the risk?

The computing power business seems to offer high returns but is actually high risk.Firstly, the export control of AI chips from the United States poses a risk common to all computing power suppliers. It directly leads to the disruption of supply chains and significant fluctuations in the prices of computing power. As a result, some companies have had to terminate AI server procurement contracts signed in the early stages.

Secondly, the situation of insufficient AI computing power supply in the past three months has improved. The rental prices for AI computing power have decreased. This implies that the high-priced computing power equipment purchased by the aforementioned listed companies may not be able to be rented out at the expected high prices.

Thirdly, the business of computing power has high technical barriers. Bare-metal leasing can only meet the sporadic needs of a small number of companies. Most companies still prefer to rent computing power from professional computing power service providers. Only a very small number of the aforementioned 10 listed companies have the capabilities for AI server networking, scheduling, and maintenance.

Supply chain risks are the most direct. Lotus Health and Jinji Shares were the first to be affected by the U.S. export controls, and the originally planned purchase of H800 servers did not arrive in full.

In September 2023, Lotus Health's subsidiary, Lotus Technology Innovation, had signed a procurement contract with New H3C for 330 NVIDIA H800 servers, with a contract value of 693 million yuan. However, as of April 2024, New H3C had only delivered 12 servers. Lotus Health announced on December 2, 2023, that there was uncertainty regarding delivery and the delivery date. In December 2023, Lotus Health began to sporadically purchase H800 servers from other suppliers, with a purchase price exceeding 2.8 million yuan per unit, far exceeding the 2.1 million yuan per unit price agreed upon with New H3C.

In October 2023, Jinji Shares' wholly-owned subsidiary, Yingzhi Innovation, originally planned to purchase 150 units of NVIDIA H800 series servers from Shenwan Technology for 312 million yuan. However, on January 3, 2024, Jinji Shares announced that Shenwan Technology was unable to complete the delivery in the short term. The announcement showed that Jinji Shares had originally planned to purchase a computing power of 2000P (PetaFLOPS, which is 1000 trillion floating-point operations per second. Peta refers to 10 to the power of 15, and FLOPS refers to the number of floating-point operations per second). Subsequently, through the method of finding alternative suppliers, nearly 500P of computing power was purchased.

Lotus Health and Jinji Shares failed to purchase enough AI servers. There are two reasons behind this. First, due to U.S. export controls, H800 servers can no longer enter the Chinese market. Second, the signing prices with suppliers before the controls were relatively low. For example, the average purchase price of H800 servers for Lotus Health was 2.1 million yuan per unit, and for Jinji Shares, it was 2.08 million yuan per unit. After the U.S. export controls were introduced, the price of a single server once exceeded 3.2 million yuan per unit. The price difference was too large, and suppliers lacked the motivation to supply.

The risk of price changes brought about by changes in supply and demand relationships is also a significant risk. This risk is becoming increasingly apparent. In 2024, the situation of insufficient supply of advanced AI computing power has slightly improved. The rental prices for computing power are decreasing.There are two main reasons. First, AI computing power that was hoarded by technology companies and telecommunications operators out of panic in 2023 is gradually coming online. Second, the "hundred-model war" is receding, with some small and medium-sized enterprises exiting model training.

Analysts from IDC China, a global market research organization, told us at a small communication meeting in January 2024 that the A100/A800, H100/H800 computing power that Chinese enterprises rushed to purchase in 2023 did not come online immediately. This computing power will be gradually released to the market in 2024. In the short term, there will not be a significant shortage of computing power in China. Jack Ma, the chairman of Alibaba's board of directors, also stated at a public event in early April, "In the short term, I think everyone has already stocked up. For the next year or eighteen months, large model training can still continue."

In 2024, the number of manufacturers capable of continuously training large models has decreased. In the short term, the situation of insufficient AI computing power has improved. A senior executive of a state-owned computing power enterprise revealed that last year, more than 200 large models were connected to their data center. This year, only leading manufacturers are still investing, while small and medium-sized enterprises have exited.

A salesperson from a server manufacturer introduced that since 2024, the situation of insufficient supply of high-performance AI servers has eased. In the past three months, the prices for purchasing and leasing AI computing power have been declining. A server manufacturer stated that in April of this year, the price of H100 servers has dropped to 2.6 million yuan per unit to 2.8 million yuan per unit, with a price reduction range of 10%-20%.

The aforementioned senior executive from the state-owned computing power enterprise introduced that this year, the rental price of high-performance AI computing power is 20% off compared to the off-peak period last year. The national supercomputing internet official website shows that some computing power manufacturers offered a limited-time discount on A800 chips between March and April 2024, with the hourly rental price for a single card as low as 2 yuan.

Small manufacturers have to compromise on computing power prices and customer selection in order to survive. For example, the progress of the computing power business announced by Lianhua Health shows that since 2024, the customer demands it has undertaken have become increasingly "fragmented"—the number of AI server rentals is decreasing, and the rental period is shortening. Some of the new contracts for Lianhua Health in 2024 also have a shorter rental period. From February to April 2024, Lianhua Health has seen some contracts with a rental period of only one or three months, or contracts for renting only 1-2 servers.

Whether it is bare-metal leasing or building one's own intelligent computing center, the final consideration is the cost and benefit account. For cross-border enterprises engaged in the bare-metal leasing business, the key is to look at the server rental price and utilization rate.

The aforementioned senior executive from the state-owned leading computing power enterprise analyzed that the rental price of an H800 server is 120,000 yuan per month, and with an average annual utilization rate of 80%, it usually does not incur a loss. At the same price, with an average annual utilization rate below 50%, the probability of loss is very high.

Bare-metal leasing has a market when there is a shortage of computing power. As the shortage of advanced AI computing power in 2024 eases, the attractiveness of bare-metal leasing to customers is also decreasing. Customers generally prefer professional enterprises with a full range of products and good services. At this time, the advantages of technology companies and telecommunications operators are obvious. They have long established professional technical and sales teams, and can provide a complete set of cloud services such as computing, networking, and storage.Is the computing power business sustainable?

For most enterprises that do not possess professional capabilities, the computing power business is unsustainable.

Among 10 listed companies, only Hangjin Technology's computing power business has progressed relatively smoothly. Hangjin Technology, controlled by the State-owned Assets Supervision and Administration Commission of Wuhan, was originally a chemical enterprise and entered the communications and electronics business in 2017, starting to layout the computing power business in 2023. Before crossing over, Hangjin Technology had accumulated six years of experience in the communications and electronics industry. The company has professional suppliers, stable customers, and has signed long-term contracts. Currently, the company has a strategic cooperation relationship with NVIDIA's agent, Unigroup. Hangjin Technology's financial reports show that in 2023, the revenue from the computing power business was 130 million yuan, with a net profit of 11.151 million yuan. In February 2024, the company signed three contracts with a service period of more than three years, with a total contract amount exceeding 800 million yuan.

However, apart from Hangjin Technology, other listed companies generally have not progressed smoothly. Computing power is a heavy asset business, which brings financial pressure to listed companies, and the payback period is not short.

Jinji Shares, Huafu Fashion, and Hengli Home announced in their company announcements that the purchase amount of computing power equipment is large, which will bring pressure to the company's cash flow and fund-raising. Hongbo Shares disclosed the company's computing power business feasibility report in a reply to the Shenzhen Stock Exchange in July 2023. The report shows that Hongbo Shares estimates that the after-tax static payback period for the first phase investment project is 6.7 years, and the after-tax dynamic payback period is 11.5 years.

The computing power business is unrelated to the main business and the investment amount is huge. The various irregular actions of listed companies have attracted the attention of regulatory authorities.

After Hongbo Shares, Lianhua Health, An'nai'er, Qunxing Toys, and Annoqi announced their entry into the computing power field, they were all issued letters by the securities regulatory department. The issues of concern to the securities regulatory department include the progress of the computing power business, risks, irregularities in information disclosure, and the impact on the company's operations.

In November 2023, Lianhua Health received a warning letter from the Henan Regulatory Bureau of the China Securities Regulatory Commission. One of the reasons was the untimely and insufficient risk warning of the computing power server purchase contract, and inaccurate information disclosure. At the end of April 2024, Lianhua Health was once again ordered to correct by the Henan Regulatory Bureau of the China Securities Regulatory Commission. The reason was that from August 2023 to January 2024, when Lianhua Health used idle raised funds to temporarily supplement working capital, 21.34 million yuan was used as long-term equity investment funds for the computing power business subsidiary and was not disclosed.

In April 2024, Hongbo Shares received a warning letter from the Fujian Regulatory Bureau of the China Securities Regulatory Commission. The reason was that the company included part of the revenue from the computing power equipment purchase contract that had not been fully executed in the 2023 performance forecast released in January 2024. This action was not recognized by the audit institution. In April 2024, Hongbo Shares issued a corrected performance forecast announcement. After excluding this transaction, the company's net profit turned from positive to negative.In March 2024, Annair received two letters of concern from the Shenzhen Stock Exchange in succession. The letters inquired about the company's stock price fluctuations and whether due diligence was conducted on the computing power service provider, Innovation Technology. On April 7, Annair received a decision from the Shenzhen Securities Regulatory Bureau requiring it to rectify its practices. The decision demanded that Annair conduct a comprehensive review of its weaknesses in business management, contract management, and capital control.

Some listed companies have decided to suspend their computing power business. On April 7, 2024, Annair announced the termination of the acquisition of the computing power service provider, Shenzhen Innovation Technology. The reason was that during the due diligence process, Annair discovered that the company owed 16.45 million yuan in taxes. Annair believed that there were significant risks associated with the recovery of accounts receivable at Innovation Technology, which severely affected the continuation of the transaction. Therefore, it decided to terminate the deal.

Henglin Home Furnishings announced on April 17, 2024, that the delivery progress of computing power did not meet expectations, and there was no synergy between the computing power business and the main business. The company terminated the purchase of equity in related computing power enterprises and server procurement orders.

Overall, looking at the aforementioned ten cross-industry listed companies, most companies face significant challenges. Some companies can make short-term profits through bare-metal leasing. However, the appeal of this model is declining. In the long run, these enterprises generally plan or are building their own intelligent computing centers. At this time, the shortcomings of high procurement costs, weak sales channels, and insufficient technological accumulation will be exposed one by one. Most cross-industry listed companies have a huge gap in capital, technology, products, and talent compared to professional computing power companies.

Operating an intelligent computing center is challenging, as it involves forming computing power clusters with hundreds, thousands, or even tens of thousands of cards. There are professional thresholds for optimizing communication between cluster graphics cards and operational management. Unprofessional companies operating intelligent computing centers may even earn less than bare-metal leasing.

Intelligent computing power is mainly divided into two parts: training and inference. In May of this year, several executives of Chinese cloud vendors expressed a viewpoint to us simultaneously — the main demand for intelligent computing power in 2023 was training computing power. After 2024, the growth rate of inference computing power will exceed that of training computing power. The reason is that as various industries use large models in their business, a batch of AI-native applications will be rapidly born.

The problem most cross-industry listed companies are facing now is that they have no technical advantages in training computing power and no price advantages in inference computing power.

Cross-industry manufacturers cannot compete with professional companies in the field of training computing power. A person from a large model manufacturer said in October 2023 that large models are not just a competition of computing power scale and the number of graphics cards. The biggest technical challenge is to ensure that large-scale training tasks are stable and uninterrupted. This is because large model training often requires thousands or tens of thousands of cards, and the training time can often be several weeks or even months. During this period, graphics cards, networks, and systems may encounter failures at any time.

A person from a leading cloud vendor said bluntly that to achieve large-scale commercialization, 5,000 cards are the threshold. Players who can only provide small-scale computing power clusters will be eliminated. Manufacturers with weak product capabilities and do not meet the needs of leading model companies will also be eliminated. In the long run, computing power manufacturers must also provide cost-effective inference solutions, supported by professional cloud platforms. Otherwise, the business cannot expand, and it will not be the first choice for corporate customers in the future.

Cross-industry manufacturers also have no price advantage in the field of inference computing power. ByteDance and Baidu's inference prices in 2024 have dropped by more than 90% compared to 2023. The reason is that technology companies often manage clusters of tens of thousands or even hundreds of thousands of servers. They have a large number of businesses that can level peaks and fill valleys, maximizing utilization efficiency, and thus reducing the price of computing power.Most publicly-listed companies that cross over into the computing power market will find it difficult to establish a foothold in the future if they cannot quickly build competitive barriers. If progress is not smooth, the computing power business may even drag down the original main business of the listed company.