Chinese artificial intelligence and cloud companies have reportedly placed orders for about six million of Nvidia’s RTX PRO 5500 workstation cards, a book worth roughly $70 billion at reported list prices, yet they expect to receive only a small fraction of that volume in the coming months. The shortfall, put at around 1% of orders over the next four months, would leave tens of thousands of cards against demand measured in millions. Neither Nvidia nor the buyers have confirmed the figures, and no shipment schedule has been published.
What the reports say
The story arrived in two waves. On 20 October, Reuters reported that ByteDance, Alibaba and Tencent together had ordered more than three million of the cards, worth over $35 billion at roughly $11,400 each. It was the first public sign of the scale of Chinese demand for the card.
The same account said Nvidia had warned partners it could fulfil only a fraction of requested volume, somewhere between 5% and 10%. Early samples were expected in December 2026, with commercial deliveries starting in early 2027. That already pointed to a rationed allocation spread over many quarters.
A 4 November Bloomberg report widened the picture sharply. Chinese AI and cloud companies had reportedly ordered about six million cards in total, nominally worth around $70 billion, yet buyers expected to receive only tens of thousands within four months.
That works out to roughly 1% of what they asked for, or about 60,000 cards against a six-million-card order book. Reuters sources had also put Nvidia’s total 2027 production of the card at no more than about three million worldwide, half the reported Chinese demand alone.
The card at the centre
The RTX PRO 5500 is a Blackwell workstation graphics card, confirmed by Nvidia’s own product listing as part of its professional line rather than the GeForce gaming range. Reports describe it with 84 GB of GDDR7 memory, error-correction support and a PCIe 5.0 interface.
Memory bandwidth is reported near 1,398 GB per second, with board power up to 600 watts. Those specifications are reported figures, not numbers tied to a fetched Nvidia specification sheet, and they should be read that way throughout this post.
What makes the card attractive is its memory. A workstation card with that much memory can serve inference workloads, where a trained model answers user requests, even though it was sold as a design and visualisation tool. That is exactly the workload Chinese labs are adding fastest.
Chinese labs are releasing models at a rapid pace, adding inference load with every launch. Tencent’s Hy4 model arrived on 4 November, and Alibaba’s Qwen family now counts hundreds of open-weight models, each release needing somewhere to run. The order book is, in that sense, a measure of compute hunger.
Why supply falls so short
The constraint starts with memory and spreads from there. Each card reportedly carries 84 GB of GDDR7, and the memory market has been tight all year, while advanced packaging, board assembly and testing capacity are shared across Nvidia’s whole product range.
The same suppliers feed a data centre business where demand is also running ahead of supply, so a workstation card competes for the same fabs, engineers and logistics as Nvidia’s flagship accelerators. A realistic outcome is a rationed allocation spread over many quarters, not a single catch-up shipment.
There is also a reason to treat the headline order book with care. Buyers who expect a 1% allocation have an incentive to file larger orders to claim a bigger share, and double ordering across vendors and distributors adds further noise to the total.
An order is a request, not revenue. Nvidia will book nothing close to $70 billion from this card in the near term, because deliveries are so limited, and the reported figures remain expectations rather than shipment data. That distinction matters for anyone reading the $70 billion figure as sales.
Why It Matters
The episode shows how much AI compute Chinese firms still want from Nvidia, and how little the export rules leave them to buy. Washington restricts sales of Nvidia’s most advanced AI parts to China, and the latest data centre accelerators are off the table for Chinese buyers.
Older parts need licences, so a workstation card designed to sit inside the export limits is the clean option left. That is why a product that is not a data centre chip at all has ended up at the centre of a geopolitical argument about compute.
For the wider market, the direct effect lands on professional cards first: long lead times and wider channel premiums as Nvidia rations supply, with Western enterprise buyers feeling it too. Gamers face a smaller, indirect effect, since GeForce and RTX PRO cards use related silicon and compete for memory.
The pattern fits Nvidia’s careful China posture, which already drew a 51% line around its cloud customers, and it underlines how export policy now shapes product strategy at the world’s most valuable chip company. Every allocation decision is now a policy decision as well.
Our read
The 1% figure matters less as a precise forecast than as a signal of how tight the market has become. In two weeks, the reported outlook worsened from 5% to 10% fulfilment to about 1%, and the gap between a six-million-card order book and a three-million-card production ceiling cannot be closed by optimism.
For Nvidia, silence is the safer play. Confirming huge Chinese demand would invite political scrutiny in Washington, while denying it would undercut a story that signals pricing power and uncooled demand for Blackwell-class silicon. Anonymous guidance to partners, as the reports describe, fits that strategy exactly.
The real test comes in stages, and each stage is checkable. Samples are reported for December, commercial deliveries for the first quarter of 2027, and Nvidia’s next earnings call will press management on demand, even if it declines to name customers or give China-specific volumes.
Meanwhile the merchant silicon race is widening on the other side of the market, as Amazon prepares to sell its own AI chips beyond its cloud. Scarcity at one supplier pushes every large buyer to look for alternatives, and that pressure will only grow while allocation stays this tight.
Background
The pattern is familiar from years of rule-bending and rule-writing: Washington sets a performance line, Nvidia designs a chip just under it, Chinese buyers order in volume, and regulators later tighten or loosen the line. Each cycle leaves Chinese firms with a little less capability than they want, and pushes them further toward domestic alternatives.
Source: Bloomberg summary, reported-versus-confirmed analysis
0 Comments