Due to the export restrictions on multiple NVIDIA chips, the phenomenon of “low prices quoted by some people and customers unable to buy goods” in the AI computing power market is no longer present. Since Huina Technology (300609.SZ), a concept stock for computing power leasing, announced a doubling of fees for NVIDIA A100 computing power in mid-November, there have been continuous reports of price increases within the industry.
According to interviews conducted by Cailian Press reporters recently, it has become normal for NVIDIA A and H graphics cards to increase by 200%. As AI computing power has become a focus of attention in the industry, it has also attracted network security companies to enter into cross-industry layouts. Under the current bottleneck of computing power, diversified computing power has become a focal point for many manufacturers. However, objective problems such as long adaptation cycles, high investment costs for customized development, and lengthy business migration still exist.
Doubling graphics card prices is normal
“In recent times, the average increase in NVIDIA compute cards is about 1.5-2 times. Even just now, even the 4090 increased fivefold but then came back down,” said an AI manufacturer representative interviewed by Cailian Press reporters.
According to an announcement from Huina Technology, there has been a significant increase in demand for computing power and continuous shortages of computational resources. As a result, fees for their high-performance server compute services embedded with NVIDIA A100 chips will be increased by 100%. In response to this news, someone from a server manufacturer stated that “the market is irrational; it’s normal to raise prices by 100% when supply falls short.”
How are other companies raising their prices? When contacted as an investor by Cailian Press reporters via telephone call with Zhongbei Communication (603220.SH), someone from their securities department stated that “based on what we previously disclosed (contract order price), it’s probably around 50%.” According to Cailian Press reporters’ analysis, Zhongbei Communication is the only company in the previous A-share computing power leasing sector that has detailed disclosures of contract unit prices. On September 7th, October 25th, and November 16th, their compute service unit prices were respectively RMB 120,000/P/year (H800), USD 25,000/P/year (approximately RMB 178,400/P/year), and RMB 180,000/P/year.
It is worth mentioning that all of the aforementioned orders from Zhongbei Communication are shown as framework contracts. Company representatives told Cailian Press reporters that even if these disclosed framework contracts are considered formal contracts, they will not sign any more formal contracts in the future.
An insider from Hengrun Co., Ltd. (603985.SH) also informed Cailian Press reporters that “now there is only inventory left for graphics cards; our company’s compute leasing prices have also been adjusted according to market changes. The increase in spot prices may be around 40%-50%.”
It is worth noting that AI computing power has attracted much attention in the industry, leading to enterprises with GPU resources venturing into cross-border layouts. A person from a leading network security listed company told Caixin reporters, “We are also engaged in computing power leasing, mainly using A-cards. Sales are dynamic and there is still availability, so place your orders as soon as possible.”
The network security expert told Caixin reporters that among their clients for computing power leasing are “Tencent, Alibaba, Huawei, UCloud.” And among the clients of listed companies are also companies such as Zhipei AI, Baichuan Intelligence, 360 and Qinghai Unicom who have increased investment in large models.
Companies target diverse computing power
Caixin reporters noticed that Hengrun shares are at the forefront of planning for computing power among listed companies. However when asked by Caixin reporters how long can the company’s graphics cards meet customer demand? The aforementioned company personnel said,”This may not be quantifiable because the market for computing power follows market supply and demand.”
The increase in parameters of large models and training data requires larger-scale deployment of computing power. Especially with current growth in inference scenarios and demands, some large model manufacturers claim that “the current hardware foundation for computing power cannot keep up.”
Caixin reporters learned from insiders that under the current bottleneck of computing power diversity has become a development trend. Companies like Hengrun shares and Zhongbei Communication have all stated to reporters that they will maintain close communication with domestic high-end chip companies amidst continuous upgrades and technological breakthroughs of domestically produced chips. Personnel from Runjian shares (002929.SZ) recently told reporters,”We have always had a cooperative relationship with Huawei; both Nvidia’s products and Huawei’s products were used in our previous plans for calculating capacity.”
In addition,Land Information (000977.SZ) has created G7 multi-computing platform which is compatible with 15 types of AI chips both domestically and internationally. Reporters learned that after the release of Source 1.0 large model, Land Information has already established cooperation with some domestic inference chip manufacturers and will continue to optimize training in collaboration with domestic computing power providers.
Start-up company Luchen Technology revealed that their Colossal-AI has been adapted to various computing powers,”Our system itself is based on Nvidia architecture, and we have also adapted it for domestic hardware such as Huawei Ascend, Bitmain, Moor Thread.”
It is worth noting that at the recent 2023 Artificial Intelligence Computing Conference, Lin Yonghua, Deputy Dean of Beijing Zhigen Artificial Intelligence Research Institute stated that there is a three-year gap between China’s AI chip large model training performance and foreign countries. “The current cluster training performance of China’s AI chips for large models only reaches close to A100/A800 in a few cases; most are less than 50%.” However, some industry insiders believe that the view regarding the three-year difference in performance may be too optimistic.
In fact, many interviewees agree that achieving diverse computing power poses significant challenges.
“There are already over 40 high-end AI training chips domestically and abroad. However, because each manufacturer adopts different technological routes during development process, there are still many incompatibilities in terms of interface interconnection. This will result in long development adaptation cycles for AI computing systems as well as high costs for customized development and lengthy business migration periods.” Liu Jun,senior vice president of Land Information said at the 2023 Artificial Intelligence Computing Conference.