On October 20, 2026, Reuters reported that ByteDance, Alibaba and Tencent have ordered more than 3 million NVIDIA RTX PRO 5500 GPUs — a potential deal worth over $35 billion. A workstation-class graphics card is being snapped up by China’s three largest internet companies at data-center scale.
A $35 Billion Order for a “Workstation Card”
The RTX PRO 5500 is not a traditional data-center accelerator — it targets high-performance professional workstations. But under tightened U.S. export controls, it has become one of the most powerful NVIDIA chips Chinese companies can still buy. The use case is consistent across all three giants: AI inference — serving trained models for chat, content generation and other products. Inference is more tolerant of per-card compute than training, but brutally sensitive to volume and delivery certainty.
NVIDIA has told customers and partners that demand far exceeds planned capacity — some partners may eventually receive only 5% to 10% of what they requested.
6 Million Orders vs. 3 Million Units of Supply
The gap runs deeper than any single order. Reuters’ figures sketch the full picture:
- Cumulative orders from Chinese customers have exceeded 6 million units;
- NVIDIA has signaled 2027 global output for the product may not exceed 3 million units;
- First samples are expected in December, with commercial delivery from Q1 2027.
The crunch is partly self-inflicted: from 2027, NVIDIA will dedicate large shares of TSMC’s related advanced capacity to its next-generation Rubin data-center platform, squeezing production room for products like the RTX PRO 5500.
From “Not Interested” to “Can’t Get Enough”
The irony is sharp. When NVIDIA launched the export-compliant RTX 6000D in mid-2025, major Chinese tech firms showed limited interest — some even declined to order. Now the similarly performing RTX PRO 5500 is being hoarded, and the discontinued RTX 6000D has doubled in price on the Chinese market. In about a year, Chinese buyers completely repriced their urgency for compliant NVIDIA silicon.
The Inference-Era Compute Squeeze
Beneath the rush lies a structural shift: the industry’s center of gravity is moving from a training arms race to an inference consumption war. As daily-active users and call volumes become the competitive front line, inference compute is capacity — and the floor of user experience. With premium supply constrained and deliveries uncertain, extracting maximum value from every card that does arrive — through multi-source compute access, intelligent routing and fine-grained cost operations — is becoming as important as securing the cards themselves.
(Compiled from Reuters and other public reports)