One machine costs 20 million yuan, the hidden sky-high GPU business | 36Kr Lite

By Li Zhaofeng

Edited by Zhang Yuxin

In the area of the bonded zone of a city in southern China, Chen Yuan (a pseudonym), a procurement representative of a company, has been waiting for a whole day.

Chen Yuan needs to purchase a batch of high-end GPU servers of a certain model (hereinafter referred to as Model X). Before he came, his boss gave him an "absolute order": the goods must be obtained on the same day. However, Chen Yuan waited from morning to 11 p.m., only to be told by the other party that "the goods cannot be sold".

Lin Yue (a pseudonym), the intermediary who helped him source the goods, immediately contacted another seller. This time, the other party quoted a price of more than 20 million yuan per unit, but still "cannot get the goods immediately".

What Chen Yuan wants is a Model X 8-card server — a machine equipped with 8 high-end GPUs of Model X. Since the beginning of this year, the price of this type of equipment has experienced a sharp surge.

At the end of last year, the domestic quotation for the Model X 8-card server was about 4 million yuan; by September this year, the quotation of the same model had risen to nearly 4 times the original level, reaching about 15 million yuan; and in the past month, the quotation has risen sharply again, exceeding 20 million yuan.

The direct reason for the price increase is that the demand for high-end GPUs continues to expand — around the world, major AI manufacturers are stepping up the "arms race" of computing power. For example, in February this year, Meta and the chip manufacturer that produces high-end GPUs announced a multi-year cooperation, planning to deploy millions of high-end GPUs to build ultra-large-scale data centers for AI model training and inference.

On the other hand, this market with huge demand and transaction volume is not a sufficiently transparent market — a piece of supply information often passes through multiple intermediaries, and the price will be increased once for each transfer. Even if the buyer is willing to pay a high price, they may not know who actually has the goods and when the delivery can be made.

Around this ever-rising market trend, *Intelligent Emergence* had in-depth conversations with several computing power practitioners. From the transactions they handled, we tried to figure out how this hidden high-value business operates.

The Business of Intermediaries

The night before Chen Yuan went to the bonded zone, Lin Yue kept tracking the goods until one or two o'clock in the morning. He remembered that he did not go to sleep until he repeatedly confirmed that the goods were fine.

"The night before, I repeatedly confirmed with the supplier, and they said there was definitely no problem. At more than 7 a.m., I started to follow up again. In the morning and at noon, I kept asking if there was any problem with the goods. The supplier kept saying, 'The goods are here, and the owner has locked them up.'"

However, under such strict precautions, he still failed to help Chen Yuan get the goods, and all the previous repeated confirmations turned out to be empty promises from the other party.

"I got goods from this seller before, and I would definitely not look for them again if I didn't trust them to some extent. But now I suspect that they may just be intermediaries too." Lin Yue said that he felt the other party did not directly control the Model X supply he wanted.

In a GPU server transaction, the most upstream party is the goods owner who directly controls the equipment, that is, the seller, and the most downstream party is the enterprise that buys the equipment for its own use or rents out computing power, that is, the buyer. Between the two, there are several layers of intermediaries: intermediaries who find buyers for the goods owner, intermediaries who find goods for buyers, and procurement agents who are entrusted by buyers to complete the procurement, and so on.

It is the existence of a large number of intermediaries that makes supply information often pass through multiple transfers before reaching the buyer. Lin Yue finds goods for buyers, but the "seller" he contacts may also be an intermediary — he is not the layer closest to the goods owner. "I have more buyer resources, but I am not close enough to the 'real sellers' and lack control over them." This has become an important reason why Lin Yue cannot guarantee the transaction even after repeated confirmations.

The huge demand gap and tight supply in the market have led to the rapid price increase of Model X GPU servers, which is a reason almost everyone can see. What is not visible is that the existence of several layers of intermediaries has also pushed up the price step by step.

"If a goods owner releases the news that 'a batch of goods is ready for sale', the intermediary closest to them will immediately pass the supply information down. Every time the information is transferred, a markup will be added on the quotation of the previous intermediary. Finally, when the quotation reaches the buyer, the price has been raised a lot."

Intermediaries who plan for buyers need to master a skill: how to judge whether the seller you are connecting with is a real seller who has the goods in hand, not just another intermediary.

Lin Yue has his own way. "Can the other party explain basic questions clearly immediately, such as what model is being sold and how much it costs? If I ask a question and have to wait for a long time to get a reply, or if they just agreed to a transaction condition and then put forward a new one, they are most likely not the real seller."

But Lin Yue knew he "misjudged" this time, "The person I was connecting with cannot directly decide how to sell, and has to ask layer by layer towards the upstream, so the buyer can't get a definite reply for a long time."

And a "definite reply" is so important for buyers. Most buyers hope to trade "face to face", that is, both parties verify capital and inspect the goods on the spot, confirm that the buyer has enough money and the seller has the goods, and then complete the transaction. But at present, the supply of high-end GPU servers is a seller's market, and sellers prefer to screen buyers first, confirm the buyer's qualification and whether the funds are in place, and then show their cards.

Buyers are in a more passive position, and they have to act fast to get the goods. The common operation is often: the buyer tells the intermediary the acceptable price, and the intermediary keeps an eye on the supply in the market at this price for a long time. Once the supply appears, the intermediary immediately asks the seller to send the business license, pay the deposit, and then lock the goods for them — those buyers who are still negotiating conditions repeatedly can't get the transaction opportunity at all.

There are also many cases where buyers are "led around" or even defrauded of money. Lin Yue described this operation: "The other party says they definitely have the goods, signs a contract with you first, asks you to pay the deposit, and then runs away with the money." In his opinion, this is also the reason why some buyers are unwilling to pay in advance to reserve goods and insist on finding spot goods.

All parties in a transaction can hardly trust each other, and caution and temptation fill the whole process.

"I can't be sure how much goods there are in the market, or whose quotation is real and whose is fake." Lin Yue suspects that some contacts who keep promising that they have goods do not directly control the supply, and have to ask other channels for information.

He doesn't even fully believe the transaction news coming from the industry: "It looks very lively, as if many deals have been closed, but the number of transactions that are actually completed may not be as large as it sounds." In his opinion, the high price has also made some buyers calm down and wait and see instead.

At present, the main buyers in the market include: model companies that purchase equipment for model training, large internet manufacturers, and computing power service providers that sell or rent computing power clusters outward. "Most of these companies will not show up directly, but entrust procurement agents to complete the transactions." Lin Yue said.

Now, facing the quotation of more than 20 million yuan, most large manufacturers are no longer willing to pay. Lin Yue said that for the Model X 8-card server, the upper limit of acceptable price for domestic leading AI manufacturers is around 15 million yuan.

However, not all model manufacturers have backed down. According to Lin Yue, recently in the market, the purchase demand of a leading model manufacturer still comes from time to time.

Who Pays for the High Price

Data from the Ministry of Industry and Information Technology shows that by the end of June this year, China's intelligent computing power scale increased by 177% year on year, and 52 intelligent computing facilities with more than 10,000 cards have been built. In addition to model manufacturers, companies engaged in computing power rental business are also the main force purchasing high-end GPU servers.

Under the market trend of rising prices and intermediaries looking for goods everywhere, Zhao Cheng (a pseudonym) cares about which purchases can finally be completed, which can help him figure out the real temperature of this seemingly lively underwater market.

With the price of Model X 8-card exceeding 20 million yuan, Zhao Cheng said: "Most model manufacturers have given up purchasing." Among the buyers he contacted, the more mainstream ones are companies that purchase equipment to build computing power clusters and then provide computing power services to downstream customers. "If I am building a computing power cluster, I have already signed a contract with the customer, promised to deliver within a certain period of time, and I will bear the corresponding liability for breach of contract if I fail to deliver. Then if the compensation cost is higher than the premium of the machine, I have to bite the bullet and buy it."

In this context, the delivery pressure is transmitted upstream along the orders, "Being able to get the equipment on time" has become the premise for most buyers to accept the premium.

Compared with ordinary intermediaries, Zhao Cheng insists on connecting with buyers directly. This is determined by his business model: after the buyer buys the GPU server, he provides subsequent deployment and maintenance services for the buyer and earns service fees. "I may give an intermediary a chance to talk, but if I still can't talk directly to the buyer after the second communication, I will not continue the cooperation."

He admitted that in this hidden underwater trading market, almost every participant is anonymous, which is directly related to the demands of buyers — most buyers are unwilling to share transaction channels, as that will add other buyers who compete with them for GPUs. "There is only one bowl of rice, which is almost enough for 5 people to share. If 10 or 20 people come to share it, it will definitely not be enough."

In the domestic high-end GPU market, in addition to Model X cards, Model Y cards, which were launched earlier and have good AI computing performance, are another main transaction model. To purchase servers of this model, you can either look for spot goods in the market, or directly apply for supply quota from the company that produces the GPU, which will arrange the supply directly. According to Zhao Cheng, the current spot quotation for Model Y 8-card servers in the underwater market is about 6.5 million yuan per unit, and the price can be about 2 million yuan lower than the spot price if purchased through the original manufacturer's quota.

However, the original manufacturer's quota with lower price has stricter application thresholds. Zhao Cheng said that some customers he contacted are in urgent need of goods, but cannot meet the audit requirements such as capital, computer room or usage certificate for the time being, so it is difficult to promote the procurement, and eventually they have to turn to the high-price market.

At present, although Model Y cards are about two-thirds cheaper than Model X cards, it does not mean that they are easier to buy. This returns to the problem repeatedly mentioned by the several intermediaries earlier: "Being able to get the goods in time" is the key to completing the transaction, but the "spot goods commitment" in the market is not reliable.

Zhao Cheng gave an example: "For example, a server was originally negotiated at 4.5 million yuan, and later the market price rose to more than 6.5 million yuan. Then I would rather pay you the liquidated damages, and sell it to others at a high price. After deducting the liquidated damages, I can earn more than before. Why wouldn't I break the contract?" Therefore, even if the buyer has locked the goods, problems may still occur in the end.

*Intelligent Emergence* learned that some transactions also come with tied sales conditions: if buyers want to buy tight models like Model X, they also have to buy other models of GPU servers such as Model Y. "It's just like buying Hermès, you have to buy matching goods too." A buyer said so.

"It doesn't mean that as long as the chips are mass-produced, the supply can be large enough." Zhao Cheng traced the supply problem to the upstream: insufficient supply of High Bandwidth Memory (HBM) and tight advanced packaging capacity are restricting the mass production of high-end servers. He estimates that there are not many ready modules in the world for assembling Model Y 8-card servers. Under optimistic circumstances, the supply can only be supported for about half a year more. Therefore, the price will be further pushed up.

The upstream of the industry is continuing to expand production, but the new supply will take time. SK Hynix said in October this year that "supply has failed to keep up with market growth", and the technical difficulty of advanced processes and the time required for new factory construction are the two major restrictive factors. The company plans to raise the production capacity of the M15X factory to the maximum level before the end of the year, and increase the supply of the new generation of high bandwidth memory HBM4.

At the performance press conference held in the same month, TSMC disclosed that nearly 20 factory projects around the world are under construction, involving advanced processes and advanced packaging, but "equipment supply is tight", and production expansion still needs to be coordinated with equipment suppliers in advance.

Enterprises are still looking for other alternatives. Public reports have disclosed that since the beginning of this year, some enterprises have turned to leasing due to the price increase of Model X 8-card servers, but at that time, the monthly rent for a one-year lease contract had risen to a maximum of 190,000 yuan, and the annual rent would be more than 2 million yuan.

In addition, some large manufacturers have even begun to inquire about the market conditions of consumer-grade graphics cards such as RTX 5080 recently.

Chen Yuan, who was ordered by his boss to buy the Model X 8-card server, still did not go back, and he is still waiting for the goods. He asked the seller to help allocate at least one or two units, so that he can see the equipment first. The other party first asked him to wait a little longer, and then said that the goods could not be allocated, and old customers were also snatching this batch of goods.

This procurement was not closed in the end. When Chen Yuan still asked if he could get the goods the next morning, the seller replied to him: "You can't get them, wait a few more days."