Three major Chinese tech companies order over 3 million NVIDIA RTX PRO 5500 cards, with order value exceeding $35 billion

[[**Three major Chinese internet companies have ordered over 3 million NVIDIA RTX PRO 5500 workstation graphics cards, with an order value exceeding $35 billion—more than twice the global planned production capacity for 2027—but actual deliveries may be only 5%–10%.**

**Key points**

**1. Compliance is the biggest draw**

The RTX PRO 5500 workstation card sits on the edge of the control list; it is one of the most powerful NVIDIA GPUs Chinese buyers can currently purchase, and procurement has been approved by the Ministry of Industry and Information Technology.

**2. Performance designed specifically for AI inference optimization**

The card uses the same GB202 chip as the RTX 5090, but memory has increased from 32GB to 84GB, bandwidth is correspondingly lower, and it supports MIG partitioning, giving it outstanding price-performance for AI inference.

**3. Replacement value far exceeds similar products from last year**

Compared with last year's lukewarm RTX PRO 6000D, the 5500 is quoted at about 85,000 yuan; paying about 50% more buys 2.6 times the memory, making it a better alternative against the backdrop of unchanged U.S. controls and a greater shortage of compute power.

**4. Ordering early to lock in limited capacity**

NVIDIA needs to allocate chip capacity among the RTX PRO 5500, the 5090, and next-generation products. The major companies know full delivery is unlikely, but they still report demand to compete for priority supply rights. 2026-10-21 13:38 More than 200 billion [yuan] scrambling for workstation graphics cards; the major companies are waiting for it to come to the rescue. Alphabet List Quick Overview

This article is from the WeChat public account: Alphabet List, author: Sibo

Recently, according to Reuters, three major Chinese internet companies collectively ordered more than 3 million NVIDIA RTX PRO 5500 graphics cards. At a list price of about $11,400 each, the potential value of these orders exceeds $35 billion.

You should know that the RTX PRO 5500 is a workstation graphics card. Such cards were previously used more for rendering, simulation, and 3D modeling, and belong to a different product line from chips designed specifically for hyperscale AI clusters. But now, it has become one of the most powerful NVIDIA GPUs that Chinese buyers can still purchase.

According to reports, total orders from Chinese customers for this card have already exceeded 6 million units, while NVIDIA has told partners that global production of this card in 2027 may not exceed 3 million units.

In other words, demand reported by the Chinese market alone is already more than twice the planned global annual output. Even if every card produced next year were sold to China, it would cover at most half.

That is only the ideal scenario; actual deliveries will be even lower. NVIDIA has already warned partners that it may ultimately be able to fulfill only 5% to 10% of the ordered quantity.

But even 5% to 10% is still 150,000 to 300,000 cards. In normal times, that would also be a large-scale purchase. But for China's major companies today, being able to get them at all is already worth it.

Major companies are eyeing workstation cards first and foremost because their choices are limited.

For Chinese customers, the RTX PRO 5500 has a very practical advantage: it is compliant and can be bought.

U.S. export restrictions on high-end chips are relatively tight. NVIDIA's B300 data center chip based on the Blackwell architecture cannot be sold to China, and the Hopper-architecture H200 also requires case-by-case approval. "Workstation cards" happen to sit on the edge of the control list.

According to reports, major Chinese companies are buying the RTX PRO 5500 mainly for AI inference, that is, letting already-trained models answer users' questions. Compared with training large models from scratch, inference has relatively lower compute requirements, and workstation graphics cards can also take on part of the task.

Restricted supply of high-performance chips, combined with continued growth in domestic AI compute demand, further highlights the replacement value of the RTX PRO 5500.

This order also meets local market requirements. At the end of September, the Ministry of Industry and Information Technology required some companies to report their procurement plans and uses for the RTX PRO 5500, and informed some of them that procurement was proposed for approval. Clearly, this purchase also complies with domestic regulations.

On the other hand, the hot demand for the RTX PRO 5500 has in turn pushed up the market for the gaming flagship RTX 5090. According to media reports, some manufacturers have already provided AI compute through modified 5090s or custom servers, while some traders are hoarding stock and holding back sales. Channel prices for the 5090 have also continued to rise, breaking past 50,000 yuan.

However, the RTX PRO 5500 was only quietly posted on the official website in September this year, and there is still some time before mass delivery; its performance in large-scale use remains to be verified.

The positioning of the RTX PRO 5500 itself is very "clever," much like a compliance-approved "special edition" that Jensen Huang built for major companies short of cards.

NVIDIA officially lists this card's maximum power consumption at 600W. Combined with its rack-oriented deployment design and air-cooled and liquid-cooled options, this card is clearly not intended for ordinary companies; power supply and heat dissipation are the first threshold. There is only one place for it—enterprise data centers.

It uses the same GB202 chip as the gaming-oriented RTX 5090, with the same 21,760 CUDA cores, but has stacked memory from 32GB to 84GB. On the other hand, memory bandwidth has dropped from about 1,792GB/s on the 5090 to 1,398GB/s. By cutting bandwidth, it gains memory capacity.

This trade-off is aimed squarely at inference. Model training is a one-time investment, while inference burns money every time it answers, and answer quality depends entirely on memory. The RTX PRO 5500 also supports MIG (Multi-Instance GPU), allowing one card to be split into two 42GB instances and run two inference tasks in parallel.

From a price perspective, it is even easier to understand its appeal.

According to media reports, the early quotation for the RTX PRO 5500 to major customers was about 85,000 yuan, while during the same period the spot price of the 5090 in some domestic channels had reached 57,000 yuan. Based on this quotation, paying about 50% more increases memory from 32GB to 84GB, 2.6 times the original, but it also has GDDR7 ECC memory as well as a design and reliability aimed at compute centers.

In fact, last year Jensen Huang launched the RTX PRO 6000D workstation graphics card, even specifically dividing it into server, workstation, and other versions, but although it was also an NVIDIA product, the 6000D received a much colder reception last year.

At the time, some major customers complained that its performance had been reduced while the price was still about 50,000 yuan, so its cost-effectiveness was not good enough, and some simply did not place orders. At the time, buyers were still hoping that export restrictions might loosen and that stronger chips could eventually be purchased.

Now, controls have not loosened, but compute is even scarcer, so the newly launched RTX PRO 5500 this year has become a more attractive alternative.

After DeepSeek went viral in 2025, local deployment and all-in-one machines became popular as well. At the time, desktop-level deployment solutions had already begun using professional cards such as the RTX 5000 Ada and RTX 5880 Ada.

The approach at the time was to install several professional cards into the same workstation to build up larger memory capacity, and then run models in parallel through software. For example, there was already a deployment solution consisting of four RTX 5000 Ada cards, with a total of 128GB of memory, used to run DeepSeek-R1's 32B distilled model.

But in this year's round, the contradiction is no longer whether it can run, but whether it can be bought. How many orders Chinese customers have placed and how many cards NVIDIA plans to produce are two different things.

For NVIDIA, it must decide who gets the limited chips first. Previously, rumors spread on overseas social media that GB202 would shift more toward RTX PRO professional cards, and that the 5090 might be discontinued. NVIDIA executive Jeff Fisher later denied the discontinuation, but indirectly acknowledged that supply chain capacity was limited.

The RTX PRO 5500, RTX 5090, and RTX PRO 6000 all use the GB202 chip, and different products must share this capacity. And in terms of the broader production arrangement, NVIDIA also has next-generation products to take care of.

Therefore, there is another very practical consideration behind major Chinese companies placing orders early: first submit demand to fight for supply volume, then arrange deployment.

These 6 million orders will be very difficult to fully fulfill, but a step late may leave even less room for choice.

Already published 788 articles on Huxiu. This content is authorized for publication by the author; the views represent only the author and do not represent Huxiu's position. If you have objections or complaints about this manuscript, please contact tougao@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]]]