ByteDance, Alibaba and Tencent Reportedly Seek 3 Million Nvidia RTX Pro 5500 GPUs
A reported $35 billion order far exceeds Nvidia’s planned supply, highlighting demand for workstation GPUs with 84GB of memory.
The reported orders for Nvidia’s RTX Pro 5500 show how a workstation GPU can become an important option for AI development when more specialized accelerators are difficult to obtain. ByteDance, Alibaba and Tencent have reportedly sought more than 3 million cards, with a combined value above $35 billion. Sources cited by Reuters say Nvidia has told customers it can’t meet that volume with its planned production.
The scale is striking against the reported price: about $11,400 per card. At that price, 3 million units would amount to roughly $34.2 billion, close to the reported order total. Those figures remain based on sources, however. Nvidia and the three Chinese companies haven’t confirmed the orders on the record.
Why this workstation card fits AI workloads
The RTX Pro 5500 has the same number of CUDA cores as Nvidia’s flagship RTX 5090, according to the source, and comes with 84GB of GDDR7 memory. That large memory pool can matter for AI tasks: keeping more of a model or its working data in GPU memory can reduce the need to split workloads across cards or move data repeatedly between memory and storage.
It’s still a workstation GPU, rather than one of Nvidia’s higher-end enterprise accelerators. The source says the RTX Pro 5500 is less powerful than the H200 and B300, which are designed for enterprise AI use. But when those products are unavailable to Chinese buyers or subject to tighter limits, a less powerful card that can be shipped becomes more useful. The comparison is about practical access as much as peak performance.
That doesn’t make the RTX Pro 5500 a direct substitute for an H200 or B300 in every workload. The article provides no benchmark results, and CUDA core counts alone don’t establish how quickly a card will train or serve a particular model. Memory capacity, software, interconnects, power use and the size of the job all affect performance. Buyers would need workload-specific testing to judge how many workstation cards a deployment requires and what operating costs that brings.
Supply is the immediate constraint
One source told Reuters that Nvidia may be able to deliver only 5% to 10% of the reported demand. Another said the company expects to produce 3 million RTX Pro 5500 cards globally in 2027. Those accounts leave open how much of that planned output is already committed to other customers, and when any Chinese orders could be fulfilled.
If supply falls that far short, the reported order total describes demand, not delivered systems. Teams planning AI capacity around the card would have to account for uncertain availability and the possibility that deliveries arrive in stages. A large order can signal strong interest, but it doesn’t show how many GPUs customers will receive or how quickly they can put them to work.
Production volume also matters beyond the chip itself. A GPU rollout depends on complete workstations or servers, power and cooling capacity, memory and networking, plus the engineering effort to install and operate the equipment. The source gives no details on the buyers’ intended deployments, system designs or delivery schedules, so it’s not possible to translate the order into a specific amount of usable AI computing.
Demand persists amid limits on access
The reported buying reflects a persistent appetite for Nvidia hardware in China, even as access to some AI-focused products remains constrained. An Nvidia spokesperson told Reuters that demand from the Chinese market has seen “unprecedented growth since 2022,” following the introduction of export restrictions. The spokesperson also pointed to constraints on U.S. firms and limits on U.S. imports in China.
The RTX Pro 5500’s appeal rests on a particular combination: substantial memory, a CUDA core count matching the RTX 5090, and a reported path to buyers who can’t readily obtain more capable enterprise GPUs. That makes it worth watching as a component in AI systems, while leaving performance and economics to be established by real deployments. The key next questions are how much of the reported demand Nvidia can serve, when cards could arrive, and how they perform on the workloads buyers actually run.
For related context, see the data center cooling state of play, custom AI ASICs, changes to AI chip controls, an Ian Buck Q&A on compute products, and data center CPU demand.
