ByteDance, Alibaba and Tencent have placed orders worth more than $35 billion for Nvidia’s RTX PRO 5500 workstation GPUs, according to Reuters, as Chinese technology groups search for powerful processors that remain available under U.S. export restrictions. The orders cover more than 3 million units, but Nvidia is expected to be able to fulfill only a fraction of the demand in the near term.
A workstation chip has become an AI substitute
Nvidia markets the RTX PRO 5500 as a professional workstation GPU rather than a conventional data-center accelerator. The Blackwell-based card has 84 GB of GDDR7 memory and is designed for AI, simulation and graphics workloads.
Chinese companies are increasingly interested in using such hardware for AI inference because access to the most powerful data-center chips is restricted.
Export controls reshaped demand
U.S. rules limit sales to China of advanced accelerators such as Nvidia’s top Blackwell and Hopper data-center products. That has pushed buyers toward products that fall below regulatory thresholds or are classified for different use cases.
The result is unusually high demand for a GPU originally intended for professional workstations.
More than 3 million units have been ordered
Reuters reported that the combined value of orders from ByteDance, Alibaba and Tencent exceeds $35 billion. Each company is building large AI services and needs additional inference capacity.
At current scale, those orders would represent one of the largest enterprise GPU procurement waves ever reported.
Nvidia may supply only a small portion
Manufacturing capacity is the immediate constraint. Reuters reported that Nvidia may be able to deliver only 5% to 10% of the requested volume.
That shortage could force Chinese customers to ration computing capacity or seek alternative domestic chips.
Inference is becoming the larger compute market
Training a frontier AI model requires massive clusters, but operating models for millions of users can consume even more computing over time. Inference workloads therefore represent a growing portion of AI infrastructure demand.
Workstation-class GPUs with large memory can be useful for many inference tasks even if they are less efficient than dedicated data-center accelerators.
Alibaba, Tencent and ByteDance have different use cases
Alibaba operates cloud services and large language models, Tencent runs social, gaming and cloud platforms, and ByteDance deploys recommendation and generative AI systems across its applications.
All three can use additional GPU capacity to support internal products and external customers.
The orders show the limits of export controls
U.S. restrictions can slow access to the fastest hardware, but they also create incentives for buyers to find compliant alternatives. Companies may use more lower-performance chips to approximate the output of restricted accelerators.
That can increase electricity consumption and infrastructure complexity without eliminating AI development.
Domestic Chinese chipmakers could benefit
If Nvidia cannot fill the orders, Chinese companies may shift more spending toward local accelerators. Huawei and other domestic suppliers have been expanding their AI-chip offerings.
Software compatibility remains a challenge because Nvidia’s CUDA ecosystem is deeply embedded in AI development.
TSMC capacity is another bottleneck
Nvidia depends on Taiwan Semiconductor Manufacturing Co. and a complex packaging supply chain. Even products outside the highest-end data-center category compete for wafer and packaging capacity.
That makes rapid production expansion difficult despite the enormous order book.
A test of how effective technology controls can be
The RTX PRO 5500 episode illustrates the cat-and-mouse dynamic created by technology restrictions. Regulators draw performance thresholds, manufacturers design product portfolios around them, and customers adapt workloads to what remains available.
For Nvidia, the demand is commercially attractive but politically sensitive. For Chinese technology companies, the question is whether compliant hardware can support AI ambitions at acceptable cost and power consumption.
Sources
- Reuters, Oct. 20, 2026
- Nvidia RTX PRO 5500 product specifications
Power efficiency will shape the economics
Using workstation GPUs at very large scale can solve an availability problem but may not deliver the same performance-per-watt as purpose-built data-center accelerators. Electricity and cooling costs therefore become an important part of the calculation for Chinese buyers.
If millions of lower-tier GPUs are required to substitute for fewer high-end accelerators, infrastructure requirements could rise substantially.
Software optimization can narrow the gap
Chinese AI developers are likely to invest heavily in model compression, quantization and inference optimization to extract more output from the hardware they can legally obtain. Those software techniques can reduce the penalty from using less powerful processors.
The result could be a more fragmented AI hardware market in which software efficiency becomes as strategically important as raw chip performance.
