Chinese demand for Nvidia hardware is not going away. On October 20, 2026, Reuters reported, citing two people familiar with the matter, that ByteDance, Alibaba and Tencent have ordered more than 3 million Nvidia RTX PRO 5500 cards with a combined sticker price above $35 billion. The same sources say Nvidia cannot promise to fill them. We read the syndicated text on The Standard (Hong Kong).
Everything below rests on anonymous sources relayed by Reuters; the three companies did not respond and Nvidia disputed none of it in the text we saw. For the earlier approval story see our post on China's approval of the RTX PRO 5500.
TL;DR
| Item | Reported figure |
|---|---|
| Buyers | ByteDance, Alibaba, Tencent |
| Quantity | More than 3 million cards |
| Price | About $11,400 each; over $35B combined at sticker price |
| Total China orders | More than 6 million, per two sources |
| Nvidia supply | Partners may get 5 to 10 percent of requests; at most 3 million cards globally in 2027 |
| Timing | Early samples December; deliveries from Q1 2027 |
| Use | Mainly inference |
| Status | Anonymous sources; not confirmed by buyers or Nvidia |
Why the RTX PRO 5500
It is a workstation card, not a data-center accelerator. Reuters says it is less powerful than the H200 and B300. The B300 cannot be sold to Chinese buyers under US export controls, and H200 exports are reviewed case by case with licensing conditions. The 5500 is, per the sources, among the most powerful Nvidia chips still available, and similar in performance to the RTX 6000D, a cut-down product Chinese firms largely shunned last year as overpriced. That calculus changed as hopes of looser rules faded and compute stayed scarce. The RTX 6000D has been discontinued and its price in China has doubled, one source said.
The supply squeeze
Sources say Nvidia will devote much of its TSMC capacity to Rubin from 2027, leaving less for products like this. Hence the 5 to 10 percent warning and the 3 million global cap. A separate datapoint: a Shenzhen-listed company, Tansun Technology, said in an October 13 filing it agreed to buy 20,000 cards from an unidentified counterparty, and warned of possible delays.
Claimed versus verified
| Claim | Status |
|---|---|
| 3M+ cards, $35B+ | Two anonymous sources; sticker-price arithmetic (3M x $11,400 is about $34B, so the "over $35B" implies slightly more than 3M) |
| 6M total Chinese orders | Two sources |
| Nvidia 5-10% fill rate | One source; another cites a 3M global cap for 2027 |
| Tansun 20,000-card filing | Public filing, per Reuters |
| Nvidia statement on export controls | On the record |
Why it matters
- Demand is not the constraint. Orders far exceed supply, which supports high prices and gray-market activity; see Taiwan's move to criminalize AI chip server exports and the Super Micro smuggling plea.
- Domestic substitution is partial. Huawei is gaining share but lags; see Nvidia vs Huawei share claims.
- Inference economics. Chinese firms are buying capacity to serve models, consistent with the capacity picture in China's 24 GW vs US 56 GW compute estimate.
- Policy risk. Washington or Beijing could change what ships before Q1 2027; the policy timeline is in our export-controls timeline.
Why a workstation card can serve inference
Training a frontier model needs thousands of accelerators linked by very fast interconnects, which is what the H200 and B300 are built for. Serving a trained model is different. Each request can run on a single card or a small group, and the main needs are memory capacity and bandwidth rather than cluster-scale networking. A workstation card with a large memory pool and a mature software stack is therefore a workable inference part, even if it would be a poor training part. That is the logic Reuters attributes to the buyers: they want to run trained models for chatbot responses, and they accept a weaker chip because stronger ones are blocked.
The trade-off is efficiency. Cards without data-center interconnects and with lower memory bandwidth deliver fewer tokens per second per dollar and per watt than purpose-built accelerators, so a given amount of serving needs more cards, more racks and more power. If the reported 3 million cards were all deployed for inference, the aggregate compute would still be material, but the power and space bill would be larger than for the same throughput on top-end parts. That fits the picture of Chinese firms stacking whatever compute they can obtain while pushing domestic suppliers.
How the arithmetic works
Reuters gives about $11,400 per card and "more than 3 million" cards. Multiplying gives roughly $34.2 billion, so the reported "over $35 billion" implies slightly above 3 million or a mix of prices; the report does not break down the split among the three buyers. "Sticker price" matters: large buyers usually negotiate, and partners resell through channels, so actual spend could differ. With the supply cap of 3 million cards globally in 2027 and total Chinese orders above 6 million, even a full allocation to China would cover only half of requested volume, before other regions are served. A 5 to 10 percent fill rate on 3 million cards would be 150,000 to 300,000 cards in the first allocation, which is the number planners should use rather than the headline.
Scenarios from here
| Scenario | What it would mean |
|---|---|
| Nvidia fills 5-10% as warned | A few hundred thousand cards reach China in 2027; gray-market prices stay high |
| Washington tightens rules on workstation cards | Orders could be cancelled or stranded; Nvidia's nominal statement about outdated controls hints at friction |
| Washington loosens rules for stronger chips | Demand for the 5500 could fall as buyers wait for better parts |
| Beijing limits US imports further | The Nvidia spokesperson cited China's own limit on US imports; domestic parts such as Huawei's gain share |
| Nvidia adds capacity | Possible only by taking wafers from Rubin, which the sources say Nvidia is unlikely to do |
None of these outcomes is predicted by the report. They illustrate why a large order book does not translate into predictable revenue or shipments.
How to treat reports like this
Anonymous-source scoops about orders are valuable early signals but should be handled carefully. Check whether more than one outlet confirms independently (here the text is a single Reuters report that others syndicated), whether a filing corroborates part of it (Tansun's October 13 filing supports the existence of demand, not the headline numbers), and whether the company or governments comment. In this case Nvidia commented on policy but not on volumes, and the buyers were silent. Until a filing or earnings call gives numbers, the safest phrasing is that Reuters reports orders of this size, and that supply will limit what actually arrives.
Background: three years of workarounds
Since 2022, US export rules have pushed Nvidia to ship a series of reduced-capability parts to China, each tuned to sit under the thresholds of the moment. The RTX 6000D, launched in the middle of last year, was one such product; Reuters reported at the time that major Chinese firms showed little interest and some declined to order, judging it overpriced for its capability and hoping Washington would eventually permit stronger chips. That hope has faded, according to two sources, and with computing capacity scarce the RTX PRO 5500 now looks like a reasonable substitute. Jensen Huang has said the restrictions cut Nvidia's share of China's AI data-center chip market from near monopoly to effectively nothing, which makes the workstation channel notable: it is a way for Nvidia to keep some presence with products the rules still allow.
Meanwhile Beijing has encouraged homegrown chips, and Huawei has gained share, though Chinese chipmakers generally struggle to match leading American products partly because of limits on access to foreign chipmaking equipment. The result is a market where buyers take any compliant Nvidia part they can get while also building on domestic silicon.
Practical takeaways
- Do not model $35 billion as Nvidia revenue. Treat it as unfilled demand with a likely fill rate far below 100 percent.
- Expect continued price pressure on workstation-class cards in China and spillover into gray markets.
- Watch inference-efficiency work. Quantization, smaller models and better serving stacks become more valuable when hardware is capped.
- Track policy dates. Deliveries from Q1 2027 mean any rule change in the next quarter could alter what ships.
What to watch
- Nvidia's next earnings commentary on China.
- Any BIS or Chinese ministry statement.
- December sample shipments.
- Whether other Chinese firms disclose orders in filings, as Tansun did.
What this means for builders
If you rent compute from Chinese clouds or depend on Chinese open models, expect continued scarcity and more inference-focused efficiency work. For everyone else, it is a reminder that Nvidia's 2027 capacity is being allocated across Rubin, workstation cards and China, which tightens supply for all buyers; see Nvidia's packaging expansion for when relief might come.
Frequently asked questions
Is the 5500 a data-center chip? No, a workstation card repurposed for inference clusters.
Does Nvidia earn $35B? No; orders are unfilled and unconfirmed.
When do cards ship? Samples December, deliveries Q1 2027, per sources.
Related reading
- China approves RTX PRO 5500 purchases
- Taiwan criminalizes AI chip server exports
- Super Micro contractor guilty plea
- Nvidia vs Huawei share: claimed vs verified
- China 24 GW vs US 56 GW compute
- AI policy timeline 2026
- Sources: Reuters via The Standard · Investing.com syndication
Based on Reuters reporting of October 20, 2026 from anonymous sources; figures are unconfirmed.