ByteDance and Alibaba Are Ditching Nvidia for Huawei's New AI Chip — What It Means for Your AI Tools

March 28, 2026  ·  5 min read

What Just Happened

On March 27, 2026, Reuters reported exclusively that customer testing of Huawei's Ascend 950PRAI chip has gone well, and that ByteDance and Alibaba plan to place substantial orders. The development is significant because Huawei's previous flagship chip, the Ascend 910C, struggled to gain traction with China's private-sector tech giants precisely because it lacked compatibility with Nvidia's CUDA software ecosystem.

The 950PR changes the calculus. It is not dramatically more powerful than the 910C in raw compute terms — but it is optimized for AI inference workloads(running trained models in real-time), and critically, its improved CUDA compatibility means developers can migrate existing Nvidia-trained models to Huawei hardware without rewriting their entire software stack.

ByteDance will use the chips for Douyin's recommendation engines and voice assistants. Alibaba will integrate them into Alibaba Cloud and its internal model development. Combined, these two companies represent a substantial fraction of China's AI inference demand.

The US Export Controls Behind the Switch

This story does not happen without US export restrictions. Since 2022, successive rounds of US export controls have blocked Chinese companies from purchasing Nvidia's most advanced AI chips — the H100, A100, and H800 — without special licenses that are rarely granted. The restrictions were tightened in late 2023, further limiting access to the H200 and other next-generation accelerators.

In early 2026, the Trump administration approved limited H200 sales to China under specific conditions — but timing and availability remain uncertain, and Chinese tech giants are not willing to stake their AI roadmaps on US regulatory goodwill. The Ascend 950PR represents the viable domestic alternative they have been waiting for.

The AI Chip War: A Timeline

What This Means for Anyone Who Uses AI

Most AI users think about models and interfaces — not chips. But hardware shapes everything: which models get the training resources to improve rapidly, which AI companies can afford to serve the most users, and which geographies become hubs for AI capability development.

As ByteDance and Alibaba shift to Huawei chips, their AI models — Douyin's recommendation AI, Alibaba's Qwen model series, ByteDance's Doubao assistant — will increasingly run on a separate hardware ecosystem from Western models. Over time, this hardware divergence shapes training pipelines, inference costs, and capability priorities.

For Western users, the immediate impact is indirect but real: Chinese AI models are increasingly competitive (Qwen 3.5's 9B model outperformed OpenAI's gpt-oss-120B on GPQA Diamond benchmarks as of March 2026), and accessing them through a multi-model platform gives you the full benefit regardless of which chip trained them.

AI Hardware Landscape: Who Runs on What

The practical conclusion is simple: locking your productivity to a single AI provider means locking it to that provider's hardware supply chain, regulatory environment, and geopolitical exposure. A multi-model platform abstracts all of that away. You use the best model for each task — wherever it runs.