Chinese tech giants want 3 million Nvidia RTX Pro 5500 cards, and Nvidia reportedly can fill only 5% to 10% of the orders
Reuters says ByteDance, Alibaba and Tencent ordered more than $35 billion of the 84 GB workstation card at about $11,400 each, a rush that shows how much AI inference demand now sits on a chip built to stay inside US export rules.
The RTX Pro 5500 is a workstation GPU with the same number of CUDA cores as its flagship RTX 5090 GPU, but with 84GB of GDDR7 memory attached, making it a good option for AI applications.
Comprehensive take
Axion take: the story is a supply-and-demand mismatch more than a geopolitics story. A card with 84 GB of memory at roughly $11,400 is the most capable inference part many Chinese buyers can legally get, and Nvidia reportedly cannot make enough of them. Western buyers should read it as a warning that availability, not price, will decide who gets this card in 2027.
Use cases
Single-card inference for 70B-class models
84 GB of ECC GDDR7 can hold a 70-billion-parameter model at 8-bit precision for weights alone, by our estimate, with limited room left for context.
Shared workstation pools
Nvidia lists Multi-Instance GPU support for up to two isolated instances and an air-cooled or rack-mounted liquid-cooled option, aimed at teams centralising GPU power.
Who should wait
Buyers who need guaranteed delivery in 2027: reported allocation is 5% to 10% of requests, and Nvidia lists the specifications as preliminary.
Specs
| GPU memory | 84 GB GDDR7 with ECC source |
|---|---|
| Memory bandwidth | 1,398 GB/s source |
| Max power | Up to 600 W source |
| Bus | PCIe Gen 5 x16 source |
| Reported unit price | About $11,400 source |
| Reported orders from the three firms | More than 3 million cards, over $35 billion source |
| Reported 2027 global production ceiling | No more than 3 million cards source |
| Launch date (per Asia Times) | 14 September 2026 source |
Insights
- Reuters puts total Chinese orders above 6 million cards, about twice the reported 3 million global production ceiling for 2027.
- Memory capacity (84 GB versus 32 GB on the RTX 5090) is the reason this card is attractive for inference, not its CUDA core count, which Tom's Hardware says matches the 5090.
- Every order figure traces to anonymous sources; Nvidia and the buyers have not confirmed them.
- Nvidia's product page lists the specifications as preliminary and the card as coming soon.
Axion Scorecard
Overall: 2.0
Not yet scored — no cited data
Not yet scored — no cited data
Not yet scored — no cited data
Not yet scored — no cited data
Not yet scored — no cited data
Nvidia has reportedly told partners they may receive only 5% to 10% of the quantities they asked for, and it can build no more than 3 million cards worldwide in 2027 (Reuters). source
Not yet scored — no cited data
The number that matters in this story is not $35 billion. It is 6 million. According to Reuters, ByteDance, Alibaba and Tencent together ordered more than 3 million RTX Pro 5500 cards, and total orders from customers in China have passed 6 million. One source told Reuters that Nvidia has indicated it can produce no more than 3 million of these cards worldwide in 2027. Put those figures side by side and China alone is asking for roughly twice what Nvidia reportedly plans to make for the whole planet. That is why partners were warned they might see only 5% to 10% of what they requested.
Nothing here has been confirmed on the record. Tom's Hardware notes that neither Nvidia nor the three companies confirmed the figures, and Reuters relies on two sources for the order size and price and a third for the supply warning. The one on-the-record line came from an Nvidia spokesperson, who said Chinese producers of workstation chips and cards have seen "unprecedented growth since 2022" and that US firms remain restricted by what the spokesperson called outdated export controls. Treat the totals as credible reporting, not audited fact.
Why a workstation card
Reuters says the buyers want the card mainly for inference, meaning running trained models to answer queries, which needs less raw compute than training. The spec sheet explains the appeal. Nvidia's own preliminary specifications list 84 GB of GDDR7 with ECC, 1,398 GB/s of memory bandwidth, PCIe Gen 5 x16, up to 600 W, and Multi-Instance GPU support for up to two isolated instances per card. Tom's Hardware reports the same CUDA core count as the RTX 5090, which carries 32 GB. By our arithmetic, 84 GB is 2.6 times that, and it is the difference between a large model fitting on one card or being split across several. A 70-billion-parameter model at 8-bit precision needs about 70 GB for weights alone, before the key-value cache that grows with context length. That fits in 84 GB and does not fit in 32 GB. That is an estimate, not a benchmark, but it shows why memory capacity rather than peak FLOPS is the selling point.
Reuters adds that the card performs similarly to the RTX 6000D, the reduced product Nvidia launched mid-2025, and that it is less powerful than the H200 and B300. The B300 cannot be sold to Chinese buyers, and H200 exports are reviewed case by case. Asia Times reports that Nvidia launched the RTX Pro 5500 on 14 September, and that China's industry ministry asked large domestic firms to disclose purchase plans after a late-September Trump-Xi summit.
What it means for buyers outside China
If you were planning to buy an RTX Pro 5500 for a local-AI workstation or a rack of shared inference nodes, expect lead times and allocation, not a shelf. Nvidia's own product page still showed "Coming Soon" with a notify button when we checked, and its specifications are marked preliminary. Anyone who needs 70B-class models on a single card should plan alternatives, such as unified-memory systems, and should not budget on the $11,400 sticker price holding for small orders.