Amnesty International NafidiaThe most valuable company in the world, by the market, is still highly dependent on a few unknown customers who collectively contribute to tens of billions of dollars.

AI Chip Darling once again warned investors in its 10-Q file to the Supreme Education Council that it has so fundamental accounts that their orders crossed the 10 % threshold of the unified global NVIDIA rotation.

For example, the elite trio of customers who enjoy a particularly deep occurring, separately between 10 billion dollars and 11 billion dollars of goods and services during the first nine months that ended in late October.

Fortunately for NVIDIA investors, this It will not change any time soon. Mandib Singh, global head of technology research in Bloomberg, says he believes that the founder and director of the CEO Henang will not stop.

“The data center training market may reach $ 1 trillion without any real decline,” he says. By this point, the NVIDIA share will definitely decrease significantly from the current 90 %. But it can be in hundreds of billions of dollars in revenue annually.

Nvidia is still restricted

Outside the defense contractors who live off the Pentagon, it is very unusual for the company to have the focus of risks between a handful of customers – let it be ready to become the first to deserve the astronomical amount 4 trillion dollars.

Looking at NVIDIA accounts on a three -month basis, it was there Four whales unknown This, in total, consists of almost every dollar of sales in the second financial quarter; This time, only one of them leaked only three people still avoid these standards.

Singh said luck The unknown whales will include microsoft, DeadPerhaps Super Micuro. But Nafidia refused to comment on speculation.

Nafidia refers to them only as customers A, B and C, and they all reported that they bought a group of collective goods and services worth $ 12.6 billion. This was more than a third of 35.1 billion dollars from NVIDIA registered for the third fiscal quarter until late October.

Its share was also divided into equal footing with every accounting by 12 %, indicating that they are likely to receive a maximum amount of chips allocated to them instead of being perfectly wanted.

This would match the comments of the founder and CEO JenSen Huang that his company The offer is restricted. NVIDIA cannot simply pump more chips, since it has a wholesale manufacturer of the leading chips in the industry to TSMC in Taiwan and has no own production facilities.

Bridges or final user?

More importantly, the appointment of NVIDIA for the main unknown customers as a customer A, the customer B, and so it has not been repaired from one financial period to another. They can change places, while keeping Nafidia their identity is a commercial secret for competitive reasons; There is no doubt that these customers will not like their investors, employees, critics, activists, and competitors, able to know the exact amount of money they spend on NVIDIA chips.

For example, one of the two parties bought about $ 4.2 billion of goods and services during the last quarterly financial period. However, it seems that it was less in the past, because it does not exceed the sign of 10 % during the first nine months in the total.

Meanwhile, it seems that the “Customer D” has done the exact opposite, which has reduced the purchases of NVIDIA chips in the past fiscal quarter, however 12 % of the rotation of the year until now.

Since their names are confidential, it is difficult to determine if they are brokers like troubled Super micro computerWhich provides data center devices, or final users such as Xai. The latter has come out of anywhere, for example, to build a new Memphis Compute Cluster In just three months.

NVIDIA long -term risks include a transformation from training to reasoning chips

However, in the end, there are only a few companies that have capital to be able to compete in the artificial intelligence race, as large language models can be expensive. These are usually cloud computing blades like Microsoft.

OracleFor example, I recently announced plans to build a Zettascale Data Center With more than 131,000 modern Blackwell AI training chips in NVIDIA, which will be stronger than any existing individual site.

It is estimated that the electricity needed to operate such a huge group is an accountTwo nuclear power plants.

The Bloomberg Intelligence Singh analyst sees only a few long -term NVIDIA risk. No one, some players are likely to reduce requests recentlyReducing its share on the market. One of these potential candidates alphabetThat has king Training chips are called TPUS.

Second, its dominance in training does not correspond to inference, which manages artificial intelligence models after it has already been trained. Here, technical requirements are not almost on their latest cases, which means that there is more competition, not only from my competitors like AMD but also companies that have a custom silicone like Timing. Ultimately, the inference will be more feasible as you use more and more artificial intelligence companies.

“There are a lot of companies trying to focus on this deductive opportunity, because you do not need the highest GPU treatment chip for it,” Singh said.

When asked whether this long -term shift to reasoning was a greater risk of losing the market in the market, he replied, “Certainly.”

This story was originally shown on Fortune.com

https://fortune.com/img-assets/wp-content/uploads/2024/11/GettyImages-2179640297-e1732217254905.jpg?resize=1200,600
2024-11-22 01:01:00