Article｜Li Zhaofeng
Editor｜Zhang Yuxin
In a bonded area of a southern city, a company’s purchasing representative, Chen Yuan (pseudonym), waited an entire day.
Chen Yuan needed to buy a batch of high-end GPU servers of a certain model (hereafter referred to as Model X). Before coming, his boss gave him a strict order: he must secure the goods that day. However, from morning until 11 PM, he was met with the response, “Not for sale.”
The intermediary helping him find the goods, Lin Yue (pseudonym), immediately contacted another seller. This time, the price quoted was over 20 million yuan per unit, but the response was still “cannot secure the goods immediately.”
Chen Yuan was looking for Model X 8-card servers — a machine equipped with 8 high-end GPUs of Model X. Since the beginning of this year, the price of this model of equipment has experienced a surge.
At the end of last year, the quoted price for the Model X 8-card server in the country was around 4 million yuan; by September of this year, the price for the same model had surged to nearly 4 times its original price, reaching around 15 million yuan. In the last month, the quoted price had risen dramatically again, surpassing 20 million yuan.
The direct reason for the price hike is the sustained expansion of demand for high-end GPUs — globally, major AI companies are hastening their “arms race” for computing power. For example, in February this year, Meta announced a multiyear partnership with the chip manufacturer producing high-end GPUs, planning to deploy millions of high-end GPUs to build large-scale data centers for AI model training and inference.
On the other hand, this enormous market of demand and transaction volume is not sufficiently transparent — a piece of source information often has to pass through multiple intermediaries, and with each transfer, the price is increased. Even if buyers are willing to pay a high price, they may not know who really has the goods and when they can deliver.
Amidst this rise in prices, “Intelligent Emergence” spoke in depth with several practitioners in computing power. From the various transactions they handled, we tried to understand how this secretive high-priced business operates.
The Business of Intermediaries
The night before Chen Yuan went to the bonded area, Lin Yue was monitoring the goods until around 1 or 2 AM. He remembers that he confirmed multiple times that there was no issue with the goods before going to sleep.
“The night before, I repeatedly confirmed with the supplier, who assured me that everything was fine. Then around 7 AM, I started following up again. By morning and noon, I was still inquiring if there were any issues with the goods. The supplier kept saying, ‘The goods are there, the owner has locked them in.’”
But despite such vigilant checking, he still couldn’t secure the goods for Chen Yuan, as all his prior confirmations turned out to be empty promises.
“Previously, I had successfully procured goods from this seller, so there was certainly some trust involved. But now, I’m beginning to doubt whether they are just intermediaries themselves.” Lin Yue said he feels that they didn’t actually have direct access to the Model X source he wanted.
In a transaction for GPU servers, the upstream consists of the owners directly controlling the equipment, i.e., the sellers, while the downstream includes companies purchasing the equipment for personal use or to rent out computing power, i.e., the buyers. Between the two, there are several layers of intermediaries: those who find buyers for the owners, those who locate sources for buyers, and others who are commissioned by buyers to complete purchases, etc.
The existence of numerous intermediaries means that information regarding sources often has to be passed through multiple hands before reaching buyers. Lin Yue finds goods for buyers, but the “seller” he contacts might also be an intermediary — he is not the closest layer to the owner. “I have more resources from buyers, but I’m not close enough to the ‘real sellers’ and lack control over them.” This became one of the main reasons Lin Yue couldn’t secure deals, even after repeatedly confirming.
The market’s huge demand gap and tight supply lead to a rapid price increase for Model X GPU servers, which is a reason almost everyone can see. What is less visible is that the presence of multiple layers of intermediaries has also incrementally pushed prices higher.
“If an owner releases information that ‘a batch of goods is ready for sale,’ the nearest intermediary will immediately pass the source information down, and each time it is transferred, a layer is added to the price from the previous intermediary; by the time it reaches the buyer, the price has been significantly increased.”
The intermediary strategizing for buyers needs to possess a skill: how to determine if the seller you are connected with truly has the goods, rather than being just another intermediary.
Lin Yue has his methods. “Basic questions like what model it is, how much it costs — can they answer promptly? If I ask a question and have to wait a long time for a response, or if an initial agreement is followed by new conditions, they are likely not the real seller.”
But Lin Yue realizes he “misjudged” this time, “the person he connected with cannot directly decide how to sell and has to ask upstream layer by layer, which results in the buyer waiting too long for reliable information.”
And “reliable information” is crucial for buyers. Most buyers prefer to trade “on the same stage,” which means both parties verify funds and goods in person, confirming that the buyer has money and the seller has the goods before completing the transaction. However, currently, the high-end GPU server supply is a seller's market, and sellers prefer to first screen buyers, ensuring their qualifications and funds are in place before revealing their bottom lines.
Buyers find themselves in a more passive position, needing to act quickly to secure goods. Common practice typically involves: the buyer informs the intermediary of an acceptable price, and the intermediary then keeps an eye on sources in the market at this price; as soon as a source surfaces, they immediately ask the seller to send their business license, pay a deposit, and lock in the goods — buyers still negotiating terms generally don’t get the opportunity to transact.
Instances where buyers are “led on” or even scammed out of money are not uncommon. Lin Yue described this practice: “The other party claims to definitely have the goods, has you sign a contract and pay a deposit, and once they receive the money, they just disappear.” To him, this is also a reason some buyers are reluctant to pay in advance and prefer to look for immediate sources.
In a transaction involving multiple parties, establishing trust among each is challenging, filled with caution and probing.
“I cannot determine how many goods are truly available in the market; I can’t tell who’s offering real prices and who’s not.” Lin Yue suspects that some contacts who confidently claim to have goods don’t actually control the sources directly and have to keep inquiring through other channels.
He doesn’t dare fully trust transaction news circulating within the industry: “It looks very lively, as if many deals were made, but the actual closed deals may not be as many as it sounds.” He observed that high prices have caused some buyers to calm down and take a wait-and-see approach.
Currently, the buyers in the market mainly include: model companies procuring equipment for model training, large internet enterprises, and computing power service providers selling or renting out computing clusters. “These companies usually won’t deal directly but will entrust procurement agents to handle transactions,” Lin Yue said.
Now, faced with quotes exceeding 20 million yuan, most large companies are unwilling to pay. Lin Yue indicated that for Model X 8-card servers, the acceptable price ceiling for domestic AI giants is around 15 million yuan.
However, not all model companies have retreated. According to Lin Yue, recently in the market, there have still been purchasing inquiries from some leading model companies.
Who Pays the High Price
Data from the Ministry of Industry and Information Technology indicates that by the end of June this year, China's intelligent computing power scale had increased by 177% year-on-year, with 52 intelligent computing facilities built containing over ten thousand cards. Besides model companies, companies engaged in renting computing power are also the main force in procuring high-end GPU servers.
In a climate of soaring prices and intermediaries scouring for goods, Zhao Cheng (pseudonym) is interested in which purchases eventually come to fruition, which allows him to gauge the true temperature of this seemingly bustling underground market.
As the price of Model X 8-cards surpassed 20 million, Zhao Cheng said: “Most model companies have given up on purchasing.” Among the buyers he has encountered, the mainstream is companies procuring equipment to build computing clusters, which then provide computing power services to downstream clients. “If I am building a computing cluster and have already signed contracts with clients, promising a delivery timeframe and facing breach responsibilities for delays, then if the compensation cost exceeds the machine's price increase, I have no choice but to buy, even if it hurts.”
In this context, the pressure of delivery is transmitted upstream along with orders, and“being able to obtain the equipment on time” has become a premise for most buyers to accept a premium.
Compared to regular intermediaries, Zhao Cheng insists on directly connecting with buyers. This is dictated by his business model: once buyers purchase GPU servers, he provides subsequent deployment and maintenance services, earning a service fee. “I might give an intermediary one chance to chat, but if the second communication cannot directly involve the buyer, I won’t continue.”
He admits that in this secretive underground trading market, every role is almost invisible, which directly relates to buyers’ demands — most buyers are unwilling to share trading channels, as this would increase competition with other buyers who might also be seeking the same cards. “There’s only so much to go around; if five people are dividing it, it’s manageable; if ten or twenty people come to divide it, it certainly won’t be enough.”
In the domestic high-end GPU market, besides Model X cards, Y cards, launched earlier and also having good AI computing performance, are another major trading model. Purchasing this model of server can involve looking for market stock or directly applying for supply quotas from the GPU manufacturer, who arranges the supply themselves. According to Zhao Cheng, the current market price for Y-type 8-card stock is about 6.5 million yuan per unit, while purchasing through the original factory quota can lower the price by around 2 million yuan.
However, lower-priced original factory quotas come with stricter application thresholds. Zhao Cheng said that some of the clients he interacts with, while eager for goods, are temporarily unable to meet the verification requirements for funds, data centers, or purpose proof, making it difficult to advance procurement, and they ultimately still have to turn to the high-priced market.
Currently, although Y-type cards are about two-thirds cheaper than X-type cards, this does not mean they are easier to obtain. This circles back to the persistent issue raised repeatedly by the previous intermediaries: “being able to obtain goods in a timely manner” is key to completing these transactions, but commitments for “immediate stock” in the market are not reliable.
Zhao Cheng gave an example: “For instance, if a server was previously quoted at 4.5 million yuan, and later the market price rises to over 6.5 million yuan, I'd rather pay you a penalty and sell it at a higher price to someone else, deducting the penalty, and earn more than before. Why wouldn’t I breach the contract?” Thus, even if a buyer has already locked in a deal, issues may still arise.
Intelligent Emergence learned that some transactions also come with bundled conditions: buyers wishing to purchase in-demand models like Model X must also buy Y-type or other models of GPU servers. “It’s just like buying an Hermes bag, you have to take the accompanying goods.” One buyer remarked.
“It’s not enough to just say that chips can be mass-produced; that doesn’t guarantee large quantities of supply.” Zhao Cheng traced the supply issue upstream: insufficient supply of high-bandwidth memory (HBM) and tight capacity for advanced packaging are both limiting mass production of high-end servers. He estimates that there are not many ready modules available globally for assembling Y-type 8-card servers, and in an optimistic scenario, they may only support supply for about six more months. Therefore, prices will continue to rise.
The upstream industry is still expanding production, but new supplies will take time. SK Hynix stated in October this year that “supply has not been able to keep up with market growth,” and the technical difficulties of advanced processes and the time needed for new plant construction are two major constraints. The company plans to maximize production capacity at its M15X factory by year-end and increase supply of the next generation of high-bandwidth memory HBM4.
At a performance release event held in the same month, TSMC disclosed that nearly 20 factory projects are progressing globally, involving advanced processes and advanced packaging, but “equipment supply is tight,” and expansion still requires prior coordination with equipment suppliers.
Companies are also seeking other options. Reports have revealed that since the beginning of this year, some companies have turned to leasing due to the rising prices of Model X 8-card servers, but at that time, the monthly rent for a year-long lease had already risen to as high as 190,000 yuan, amounting to over 2 million yuan for the entire year.
Additionally, some large firms have recently started inquiring about the market for consumer-level graphics cards like RTX 5080.
Chen Yuan, who was sent by his boss to purchase Model X 8-card servers, still hasn’t gone back; he is still waiting for the goods. He asked the seller to at least pull out one or two units for him to see the equipment. Initially, the other party told him to wait a bit longer, then later said the goods couldn’t be coordinated and that old customers were also competing for this batch of products.
This purchase ultimately did not conclude. When Chen Yuan continued to ask whether he could obtain goods the next morning, the seller replied, “No, wait a few more days.”
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。
