RTX PRO 5500 China Price: What Enterprises Are Quoted

In the last days of September, searches for the RTX PRO 5500 China price mix three questions: whether MIIT will allow purchases, what one card is quoted at, and what an AI server costs once the cards are installed. The product page still says Coming Soon. The figures below stay attached to who said them, on which day, and whether they can enter a contract.

Where the MIIT inquiry stands

On 27 September 2026, The Information reported that Beijing had signaled some domestic companies could buy this new chip for high-end professional computers. The Ministry of Industry and Information Technology asked firms for purchase quantities and planned use, and told some of them it intended to approve. The Dong-a Ilbo retelling on 29 September names ByteDance and Alibaba, and says ByteDance is considering an order of about one million chips.

The same relays include a supply cadence: NVIDIA plans to ship about 500,000 units a quarter to China starting in late December, and sales staff asked customers to order by 30 September to lock supply. One million cards is about two quarters of that plan. Today is the last day of that window, and the date comes from interviews, not an NVIDIA newsroom post.

The previous buying round has been stalled for about a year. Dong-a Ilbo writes that in September 2025 China told companies to stop testing and ordering the RTX PRO 6000D; after the United States cleared H200 exports in December 2025, Beijing still limited entry; and in October 2025 Jensen Huang said NVIDIA’s share of China’s AI-chip market had fallen from 95% to 0%. This inquiry uses the same “state quantity and use, then discuss approval” shape. Export permission still depends on Washington. Coverage only says industry contacts expect a workstation card to clear more easily than a top training accelerator. The U.S. government has not published a license decision for the RTX PRO 5500.

For the earlier export rules and how teams split training, inference, and agents, see how China-based teams buy advanced AI compute after NVIDIA export limits.

Specs on the product page that touch the bill

The English product page and the Traditional Chinese page are live, titled RTX PRO 5500 Blackwell Workstation Edition. Both mark specifications as preliminary and subject to change. A 15 September channel piece said the mainland China site did not list the page at that time. Whether a purchase can happen does not depend on that URL loading.

The official FAQ states MIG plainly: one instance of up to 84 GB, or two isolated instances of up to 42 GB each, with dedicated memory, cache, compute cores, and quality of service. Eight cards split in half are 16 instances on the rack. That sentence is usable in a capacity plan. The product page does not publish a CUDA-core count. The 21,760 CUDA figure in the 15 September piece is a channel inference and does not belong on a configuration sheet.

“Up to 3×” Tensor and “up to 2×” ray tracing are NVIDIA’s ceilings versus the previous generation. The page does not publish China-region frame rates or tokens per second. Keep the words “up to” in the purchase appendix.

Two China prices, kept in separate columns

A search for the RTX PRO 5500 price hits two weeks and two methods.

• Buyer quotes after 27 September 2026. Relays put local buyers at RMB 85,000–90,000 per card, about USD 13,000. Dong-a Ilbo writes “up to RMB 90,000.” The quotes come from buyers, not an NVIDIA price list. The same reports say the band sits near Huawei’s Ascend 950PR. This article does not add a separate 950PR list price.
• The 15 September channel guess. Sina Finance, citing a hardware channel: the PRO 6000 workstation was about USD 16,000 official and RMB 160,000–180,000 in China channels, with a higher server version; the PRO 5000 72 GB was generally above RMB 80,000. From that, the guess was that a 5500 in China would land around RMB 130,000. That is a pre-launch comparison, not an invoice.

RMB 85,000 is about a third below RMB 130,000. A contract at RMB 90,000 would mean the 15 September guess overstated the card budget. A later channel markup would mean RMB 85,000 understated it. Until an approval and an invoice exist, both columns stay marked unsold.

The million-unit intent multiplies directly in yuan: 1,000,000 × 85,000 = RMB 85 billion, and 1,000,000 × 90,000 = RMB 90 billion. Relays convert that band to about USD 12.6–13.4 billion. The product is an intent-scale multiplication, not recognized revenue. A quarter of 500,000 cards at RMB 85,000 is about RMB 42.5 billion of goods value, again only the low quote extended.

An 8-GPU AI server: cards and electricity apart

In the procurement relays, buyers want the cards in servers, eight per machine, running chatbot and agent inference. The official form factor is a rack workstation with air cooling or liquid cooling on the RXM. The 8-GPU server is a deployment sketch. The workstation is the product definition. The budget needs both rows.

Card value (price math, not a hands-on test):

• 8 × 85,000 = RMB 680,000
• 8 × 90,000 = RMB 720,000
• 8 × 130,000 ≈ RMB 1,040,000 (only if you still use the 15 September guess)

Power (assumption first): the official cap is 600 W. Eight cards at that cap are 4.8 kW, GPUs only. At RMB 0.8 per kWh, 24 hours, 30 days: 4.8 × 24 × 30 × 0.8 = RMB 2,764.8 per month, about RMB 2,800. PSU conversion, CPUs, switching, and cooling are outside that product. If your commercial tariff is not RMB 0.8, scale this line.

Still off the sheet: chassis and redundant PSUs, CPU and DRAM, NICs, rack space, liquid cooling only on RXM, import and use-case review, and NVIDIA software licenses. For a small team with low concurrency, the same inference often lands on API tokens first. An owned 8-GPU node pays back only after utilization spreads the RMB 680,000. Always-on agents, signing, and the repo do not have to live on that GPU. See Mac mini as an AI agent server: memory is the bottleneck.

How far enterprise procurement can go now

The product-page action is a notify-me signup or a conversation with an NVIDIA product expert. The China path in the reports is different: large companies tell MIIT the quantity and the use, then wait for an approval line. The names in print are ByteDance and Alibaba. Smaller firms are not on the public list for this round of questions.

• Large internet companies: prepare quantity plus use. A use case written as a training flagship does not match the official inference and graphics positioning.
• Workstation teams in graphics, simulation, and video: 84 GB ECC, up to four DisplayPort 2.1b outputs, 9th-gen NVENC, and 6th-gen NVDEC are already on the product page. The ship date is still Coming Soon.
• Budgets built at RMB 130,000 or RMB 90,000: keep both as scenarios until an approval and a quotation exist.
• 30 September: the lock-in date in the relays. Missing it does not erase later quarters of the 500,000-unit plan. It expires the “order now to queue” sentence.

Which workloads this card covers

The product page lists six blocks: agents and generative AI including LLM inference, physical AI and Omniverse simulation, AI rendering, scientific computing, video production, and MIG so one card serves two users. The 84 GB of GDDR7 is there for longer context, several models, and large 3D scenes. The page does not describe an HBM training card.

The comparison with Ascend 950PR is one sentence so far: buyers say the price band is close. Whether operators, collectives, and toolchains swap is a question for shipping drivers and sample cards. Equal unit price does not price a migration.

The PRO 6000 comparison stays inside the 15 September channel description: RMB 160,000–180,000 for the workstation in China channels, 96 GB of memory. If the 5500’s 84 GB and the lower buyer quote both hold, the gap is a lower spec and also whether this round of approvals actually opens. Until both cards can be invoiced, the gap is a scenario.

Three checks from here

• The approval text: whether MIIT publishes a quantity cap, a company list, and use limits.
• The U.S. license: whether Washington writes this workstation card into a license or an exception. Industry expectation does not replace the Federal Register.
• The first invoice: whether the cleared price lands at RMB 85,000–90,000 or back near the RMB 130,000 channel guess. Rebuild the 8-GPU budget from the invoice. Keep the RMB 0.8/kWh power assumption until your tariff replaces it.

Late December is the first shipment window in the relays. Before that, a server plan can lock chassis power (leave headroom at 600 W per card) and the MIG split (42 GB or 84 GB). The unit-price cell keeps two scenarios.

Questions

What is the official China price?

The product page has none. The citable figures are the RMB 85,000–90,000 buyer quotes and the RMB 130,000 guess from 15 September.

Has ByteDance bought one million cards?

The verb in the reports is “considering.” RMB 85–90 billion is quote times quantity, not a confirmed contract.

What does an 8-GPU server cost?

Cards only: RMB 680,000–720,000 at the buyer quotes, or about RMB 1.04 million at the RMB 130,000 guess. GPU electricity at a RMB 0.8/kWh assumption is about RMB 2,800 a month at the 600 W cap.

Can a smaller company order now?

The public action is the notify-me form. This round of questions is large tech firms reporting demand to MIIT.

Can it replace an H200 for large-model training?

The official sentence is inference, agents, simulation, and graphics on a workstation. Training flagships stay on the datacenter line.