Alibaba and ByteDance Plan to Order Huawei 950PR AI Chip as CUDA Compatibility Improves
Alibaba and ByteDance are planning to order Huawei's new Ascend 950PR AI chip after testing showed dramatically improved compatibility with Nvidia's CUDA software ecosystem. Huawei is targeting approximately 750,000 units in 2026 — a major step in China's push to reduce dependence on restricted Western AI hardware.
Why the 950PR Matters
The Ascend 950PR represents a strategic shift for Huawei. Previous chips stuck to Huawei's proprietary CANN software system, creating a painful migration barrier for Chinese tech companies that had built their AI stacks on Nvidia's CUDA. The 950PR addresses this directly:
- Better CUDA compatibility: Developers can migrate existing Nvidia-based models much more easily
- Improved response speeds: Better inference performance for running trained AI models
- Optimized for inference: While raw compute improvement over the 910C is modest, the chip excels at the inference workloads that power real-world AI applications
Pricing
Huawei is offering two versions:
- Standard (DDR memory): ~50,000 yuan ($6,900) per card
- Premium (HBM memory): ~70,000 yuan ($9,700) per card
For context, Nvidia's H100 cards sell for $25,000-$40,000 each (when available), making the 950PR significantly cheaper — though likely with lower peak performance.
Who's Buying
Both Alibaba and ByteDance have confirmed plans to order the 950PR after successful testing. These are two of China's largest AI infrastructure buyers — Alibaba for its cloud and e-commerce AI, ByteDance for TikTok's recommendation algorithms and its growing AI product lineup (including the Seedance video generation model).
Samples were sent to customers in January, with mass production beginning next month and full-scale shipments in the second half of 2026.
The Bigger Picture: China's AI Chip Independence
The 950PR is China's most credible answer to US export controls that have blocked sales of Nvidia's A100, H100, and H200 chips since October 2022. While Huawei's chips don't match Nvidia's peak performance, the CUDA compatibility improvement removes the biggest barrier to adoption — software migration.
With 750,000 units targeted for 2026, Huawei is scaling up rapidly. If the chips perform well in production, they could significantly reduce China's dependence on smuggled Nvidia hardware — a market that, as the recent $62 million DOJ bust showed, carries enormous legal risks.
Bottom Line
Huawei's 950PR isn't going to match Nvidia chip-for-chip on raw performance, but that's not the point. By dramatically improving CUDA compatibility and pricing the chip at a fraction of Nvidia's cost, Huawei is making it practical for Chinese tech giants to build domestic AI infrastructure. When Alibaba and ByteDance — two of the world's largest AI companies — commit to ordering your chips, it signals that the chip independence strategy is starting to work.
Frequently Asked Questions
Can the 950PR replace Nvidia GPUs?
For inference workloads (running trained models), the 950PR is increasingly competitive. For training large models from scratch, Nvidia's chips still hold a significant performance advantage.
Why does CUDA compatibility matter?
Most AI software worldwide was built on Nvidia's CUDA platform. Better CUDA compatibility means Chinese companies can migrate their existing AI code to Huawei chips without rewriting everything — a massive practical barrier that previous Huawei chips couldn't overcome.
How does the price compare to Nvidia?
The 950PR starts at ~$6,900 per card. Nvidia's H100 sells for $25,000-$40,000. The significant price difference makes the 950PR attractive for cost-sensitive deployments, even if peak performance is lower.