What was reported
Bloomberg, citing people familiar with the matter, says Alibaba, Tencent, ByteDance and other Chinese AI and cloud firms have ordered about 6 million RTX PRO 5500 chips. Those companies reportedly expect only some tens of thousands, or roughly 1%, within the next four months. The sources asked not to be named because they were discussing a private matter.
This is not a confirmed shipment tally. Coverage of the story notes that the reports describe private orders and internal production estimates, and that NVIDIA has not publicly confirmed either total. The roughly $70 billion figure is an estimate of order value. It is not a record of completed sales.
How this builds on the October Reuters report
Reuters reported on October 20 that ByteDance, Alibaba and Tencent alone had ordered more than 3 million cards. Those orders were at about $11,400 apiece, for a combined sticker value above $35 billion. Reuters' sources said total orders from Chinese customers had passed 6 million. The Bloomberg figure builds on that earlier reporting.
Two supply numbers appear in the coverage, and they are easy to confuse:
| Figure | Source | What it describes |
|---|---|---|
| 5% to 10% of requested quantities | Reuters, one source | What NVIDIA reportedly told partners they might receive |
| Roughly 1% within four months | Bloomberg, unnamed people | What Chinese buyers are planning for in the near term |
| No more than 3 million cards worldwide in 2027 | Reuters, one source | The production ceiling NVIDIA reportedly indicated |
The figures cover different periods and come from different sources. They are not contradictory estimates of the same thing. Reuters also said NVIDIA told customers to expect early samples in December, with commercial deliveries starting in the first quarter of 2027.
The scale mismatch is striking. Reported Chinese orders of more than 6 million are double the reported global 2027 output ceiling of 3 million. That does not mean all global output is earmarked for China, and it does not mean every order is firm or paid. Some orders may be speculative, duplicated across suppliers or padded in the hope of getting a bigger allocation. That last point is my own inference, not something the reports state.
Why a workstation card?
The reporting describes the card as "not powerful enough to fall under those restrictions." Reuters framed it as a less powerful but export-compliant option. Its comparison points:
- The Blackwell-based B300 data-center chip is not allowed to be sold to Chinese buyers.
- The Hopper-based H200 is reviewed case by case and subject to licensing conditions.
- The RTX PRO 5500 is less powerful than both, and Reuters said it is similar to the RTX 6000D in performance, based on product specifications it saw.
Some outlets have gone further. Tom's Hardware, for example, says the card faces no export limits at all. That is probably too strong. The Reuters framing is narrower: the card is one of the most powerful NVIDIA parts still available to Chinese buyers, with no claim that every sale is blanket-approved. The source material also says no regulator has publicly confirmed a product-specific ruling covering all the reported orders. Treat "legal option" as the reported market understanding, not a verified legal determination for each transaction.
What the card is
NVIDIA's own product page lists the RTX PRO 5500 Blackwell Workstation Edition with these specifications:
- Blackwell architecture with fifth-generation Tensor Cores, which add FP4 precision support.
- 84 GB of GDDR7 memory with ECC.
- 1,398 GB/s of memory bandwidth.
- PCIe Gen 5 x16 and up to 4x DisplayPort 2.1b outputs.
- Maximum power consumption of up to 600 W.
- Air cooling, with a liquid-cooled option in the RXM form factor.
- Multi-Instance GPU support: one instance of up to 84 GB, or two instances of up to 42 GB each.
NVIDIA positions the card for rack-mounted workstation deployment. It says IT teams can centralize GPUs in managed racks and share them across users. It lists LLM inference, AI agents, generative AI, physical AI simulation, rendering, scientific computing and video among the workloads.
That 84 GB of memory is the most plausible reason for the appeal. Inference is often limited by how much of a model fits in memory, and a large pool on a single card reduces the need to split models across devices. A secondary report also lists 21,760 CUDA cores, the same count as the consumer RTX 5090. NVIDIA's page, as captured, does not list a CUDA core count.
These are manufacturer specifications. Neither the reporting nor NVIDIA's page offers benchmarks for particular models, so claims about real throughput or cost-effectiveness against an H200 or B300 are unsupported. Reuters said the buyers want the cards mainly for inference, meaning running trained models to generate outputs such as chatbot responses. Inference generally needs less computing power than training. Reuters did not say every card would be used that way.
Why supply is tight
Reuters' sources point to NVIDIA's manufacturing priorities. As NVIDIA and TSMC shift capacity toward next-generation products such as Rubin, supply of the RTX PRO 5500 is expected to stay tight. Two Reuters sources said NVIDIA is expected to devote much of its TSMC capacity to Rubin from 2027. TSMC declined to comment.
There is also a small public data point. Reuters noted that Shenzhen-listed Tansun Technology said in an October 13 filing that it had agreed to buy 20,000 RTX PRO 5500 cards from an unidentified counterparty. The company warned that supply shortages, industry demand and overseas technology controls could delay deliveries or reduce volumes. That is a corporate filing, separate from the anonymous-source estimates.
Reuters reported that the RTX 6000D has been discontinued and its market price in China has doubled, according to one source. Major Chinese buyers had initially shown little interest in that card. Reuters said hopes that Washington would allow more powerful chips have faded, which made the RTX PRO 5500 a more attractive substitute.
Some secondary outlets cite a buyer's internal calculation that Chinese hyperscalers could receive about 700,000 units by the end of 2027 as production rises. That number is a projection relayed second-hand and should be treated cautiously.
NVIDIA's position
An NVIDIA spokesperson told Reuters that US firms remain restricted by "outdated US export controls" covering gaming products released nearly half a decade ago, along with China's own limit on US imports. That is NVIDIA's characterization, and it is not an independent assessment. The company also declined to comment on shipment numbers when asked by Bloomberg. Reuters said ByteDance, Alibaba and Tencent did not respond to requests for comment. The same was true of the US Commerce Department's Bureau of Industry and Security and China's Ministry of Industry and Information Technology.
What it means for PC and IT readers
Outside China, the effects are indirect but real:
- Allocation pressure: If a large share of a limited production run is spoken for, buyers elsewhere may face longer lead times. The reporting does not say this, so treat it as a plausible risk rather than a finding.
- Rack-mounted workstation GPUs: NVIDIA is selling this card as shared, centrally managed workstation capacity. That shows how blurry the line between workstation and data-center hardware has become.
- Policy and supply risk: The story shows that a GPU's status can hinge on export-rule thresholds and foundry priorities as much as on engineering.
The unanswered questions are how many of the 6 million orders are real, whether regulators change their treatment of the card, and whether NVIDIA's Rubin ramp squeezes supply further after 2027. Until NVIDIA or the regulators say more, the numbers here come from anonymous sources.