NVIDIAThe Story
ByteDance, Alibaba and Tencent have ordered more than 3 million of Nvidia's RTX PRO 5500 graphics cards, with a combined list price exceeding $35 billion, according to a Reuters report citing two people familiar with the matter.
The cards are priced at roughly $11,400 each. Total orders for the card from Chinese customers have exceeded 6 million units, more than double what Nvidia has planned to produce. The company has told partners it cannot guarantee fulfilment and that recipients may receive only 5 to 10 per cent of requested quantities, with global production capped at around 3 million cards in 2027.
The RTX PRO 5500 is a workstation card carrying the same CUDA core count as the RTX 5090 with 84GB of GDDR7 memory. It ranks among the most capable Nvidia parts available to Chinese buyers under current United States export restrictions. The Blackwell-based B300 data centre chip cannot be sold to Chinese customers, and exports of the Hopper-based H200 require case-by-case review and licensing conditions.
The orders mark a sharp reversal for a market that had largely closed. On Nvidia's most recent earnings call at the end of August, chief financial officer Colette Kress said less than 1 per cent of data centre revenue came from Chinese customers, and that it related to older-generation Hopper H200 parts. The company has projected revenue of about $108 billion for the third quarter of its 2027 financial year.
The orders come days after Beijing issued Announcement No. 48, a draft proposal to prohibit export of 14 technologies and license 26 others across artificial intelligence, photovoltaics and biomedicine.
Why It Matters
The $35 billion is the least reliable number in this story, and it is the one in every headline.
Six million units have been ordered against a production ceiling of three million, and Nvidia has already told partners to expect 5 to 10 per cent of what they asked for. Apply that to the big three and the order book converts to something between $1.75 billion and $3.5 billion. Everyone placing these orders knows the arithmetic. They are ordering multiples of what they need precisely because allocation will be pro-rated, which is standard behaviour in any shortage and makes the aggregate figure a measure of anxiety rather than demand.
What the number does establish is the shape of the workaround. The RTX PRO 5500 is a workstation card. Nobody buys three million workstation cards to put in workstations. These are being bought to be racked, in volume, as a substitute for data centre silicon that cannot legally cross the border.
That is the export control regime working exactly as designed and failing at what it was for. The design worked: China cannot buy B300s, and Nvidia's China data centre revenue had fallen below 1 per cent by August. The purpose has not been served: the compute is being assembled anyway, out of parts that are permitted, in quantities that compensate for their individual weakness.
The timing is the sharpest detail. Beijing published its own draft export controls on AI and solar know-how five days before these orders surfaced. One country is restricting the artefact while the other restricts the recipe, and in the same week China placed the largest order for American hardware it has been allowed to place in two years.
The Strategic Read
The substitution being made here has a physical cost, and it is paid in electricity.
A workstation card with GDDR7 memory is markedly less efficient for large-scale AI work than a data centre part built around high-bandwidth memory and designed to sit in a dense rack. Running six million of them to approximate the capability of a far smaller number of B300s means more boards, more floor space, more cooling and considerably more power for the same output.
China can absorb that. It has spent two decades building generation and transmission ahead of demand, and its grid connection timelines are measured in months rather than the thirteen years now quoted in parts of Britain. The constraint that forced Microsoft out of its Essex site this week is not the binding one in Hangzhou or Shenzhen.
That is the part of the export control calculation that gets least attention. Denying China the efficient chip pushes it toward the inefficient one, and the penalty for inefficiency is energy. A country with surplus power and a fast-moving grid pays that penalty more easily than the country imposing the restriction would.
For India the lesson is narrower and more immediate. Whatever Nvidia does end up shipping, 6 million units of Chinese demand sit ahead of everyone else in the queue for the same fabrication capacity. India's 10,000-GPU procurement and its sovereign compute ambitions are bidding into a market where the overflow from one restricted country exceeds global supply.
For daily, sharp analysis of the biggest moves in the Indian business and startup ecosystem, follow StartupFox.