Reuters Exclusive: ByteDance Commits $5.6B to Huawei Ascend 950PR; Alibaba and Tencent Join Chip Scramble After DeepSeek V4

A Reuters exclusive reported April 29 that China's biggest internet companies — ByteDance, Alibaba, and Tencent — have scrambled to secure Huawei Ascend 950PR chip orders following DeepSeek's V4 model launch optimized for Huawei hardware. ByteDance committed the most consequential figure: $5.6 billion in Huawei Ascend 950PR orders for 2026 — the single largest non-Nvidia AI chip procurement commitment ever made by one company. At approximately $16,000 per unit, ByteDance's commitment implies roughly 350,000 chips, nearly half of Huawei's entire planned 2026 production target of 750,000 units. ByteDance's scale of commitment reflects the strategic logic: the company already runs 120 trillion tokens per day on its Doubao AI platform and needs massive compute expansion; with Nvidia H20 chips restricted and H100/Blackwell unavailable, the Ascend 950PR is the only viable scale option in China's market. Alibaba Cloud deployed DeepSeek V4-Pro and V4-Flash on its Bailian platform the same day as V4's launch — the fastest major cloud deployment of any new AI model in China's history — and has placed significant orders for Ascend 950PR chips to handle projected inference demand. Tencent Cloud similarly rushed to secure allocation. Cloud and GPU rental operators are also pushing to get on Huawei's order queue. The Huawei Ascend 950PR delivers 1.56 petaflops of AI compute and 112GB HBM memory — outperforming Nvidia's H20 (the only Nvidia chip China can legally acquire) and offering CUDA-compatible software tooling that lowers migration friction for ByteDance's existing production systems. Huawei's plan to ship 750,000 units in 2026 with mass production beginning in April and full-scale delivery in H2 faces significant supply constraints: US export controls prevent China from acquiring advanced chipmaking equipment, limiting SMIC's throughput for the chips. The surge in demand following DeepSeek V4's launch — which validated running world-class AI on Huawei hardware — may create a multi-quarter backlog. For context, $5.6 billion is also the exact same figure as DeepSeek's total claimed training cost — but this represents a single chip order, not training, illustrating the difference in scale between model training and production inference infrastructure.