Topics

China AI chip and compute updates (which signals matter for builders)

Evergreen topic pages updated with new evidence

Export-control verified Hardware specs 2026 Updated

Last reviewed: 2026-06-23 · Policy: Editorial standards · Methodology

Decision in 20 seconds

China AI chip and compute updates matter most when they change deployment math — what hardware builders can access, at what price, and with what regulatory constraints. As of May 2026, Huawei's Ascend 910C is the primary domestic GPU alternative for Nvidia H100-class workloads in China, delivering roughly 60–70% of H100 throughput at comparable or higher cost. U.S. export controls (BIS rules updated October 2023 and May 2024) have accelerated domestic chip adoption but created a two-tier compute market: Chinese-domestic deployments increasingly use Ascend or Cambricon; international cloud access still leans on Nvidia H100/A100 via overseas entities.

Use this page when

  • You need to understand what hardware Chinese AI labs are actually using for training and inference.
  • You're planning a deployment inside China and need to know the realistic GPU options.
  • You want to track export control developments that affect China AI compute capacity.

This page is not for

  • Detailed GPU benchmark comparisons — use MLCommons or vendor spec sheets for precise numbers.
  • Consumer GPU purchasing advice — this page focuses on datacenter/cloud deployment hardware.
  • TSMC and semiconductor manufacturing supply chain — that's a different monitoring track.

Key points

  • Huawei Ascend 910C (available Q4 2024): ~60-70% of Nvidia H100 FP16 throughput for transformer workloads; CANN software stack (Huawei's CUDA equivalent) has improved PyTorch compatibility significantly since 2023.
  • Biren BR100 (2022–present): ~320 TFLOPS FP16, comparable to Nvidia A100 at launch; limited software ecosystem uptake vs. Ascend; primarily used in Biren's own cloud offerings.
  • Cambricon MLU-590 (2024): enterprise inference focus, strong INT8/INT4 performance, better software support for deployment vs. training; used by some Chinese cloud providers for inference serving.
  • BIS export controls (Oct 2023, updated May 2024): Nvidia H100, A100, and AMD MI300 series restricted for direct export to China; Nvidia H20 (China-specific downgraded variant) released as compliant option.
  • Nvidia H20 dynamics: H20 is China's legally available Nvidia GPU — ~33% of H100 compute at similar price; Chinese hyperscalers (Baidu Cloud, Alibaba Cloud, Tencent Cloud) ordered billions of dollars of H20 in 2024.
  • ByteDance and Alibaba are building domestic compute infrastructure using Ascend for training and H20 for inference, creating a hybrid deployment architecture.

What changed recently

  • April 2026: Reports of Huawei Ascend 910D development — targeting H100-parity performance with hardware scheduled for late 2026 production.
  • March 2026: Alibaba Cloud expanded Ascend 910B-based instances to general availability; pricing at ~30% premium over equivalent H20 instances.
  • January 2026: BIS added new entity list updates affecting Huawei supply chain components; impacts Ascend 910C volume production timeline.
  • Late 2025: Chinese hyperscaler Baidu disclosed that >40% of its new AI training capacity uses domestic chips (primarily Ascend).
  • November 2025: Nvidia H20 orders from China totaling ~$12B reported in 2024 annual supply chain disclosures — largest single-country chip order outside US.

Explanation

The practical impact for builders in 2026: if you're deploying on Chinese domestic cloud (Alibaba Cloud, Baidu Cloud, Tencent Cloud), you will encounter Ascend 910B/910C instances alongside H20 instances. Ascend has better price-performance for Huawei's own model families (PanGu) but may require CANN-specific code for optimal performance. H20 is generally more portable with standard PyTorch/CUDA code.

Export controls have split the compute ecosystem into three tiers: (1) Nvidia A100/H100 — only available in China via pre-October 2023 inventory or overseas deployments; (2) Nvidia H20 — legally available, limited compute; (3) Domestic chips (Ascend, Cambricon, Biren) — full compute but software ecosystem friction.

The builder-relevant question is: does your model training or inference workload need to run inside China? If yes, plan for Ascend or H20 as the realistic hardware options. If no, H100 via international cloud remains available.

China AI compute landscape — available hardware tiers

Three compute tiers define what Chinese AI deployments can access. Each tier has different software compatibility and cost-performance profiles.

How to verify the answer

For export control updates, BIS.gov is the authoritative source. For chip availability and pricing, check Alibaba Cloud, Baidu Cloud, and Huawei Cloud instance catalogs directly.

Tools / Examples

  • Baidu's ERNIE training in 2024–2025: used a hybrid Ascend 910B + H20 cluster — Ascend for pre-training (lower cost, available at scale), H20 for fine-tuning (better software compatibility with research tools).
  • Alibaba Qwen3 training: confirmed Alibaba used Ascend infrastructure for domestic training runs; Apache 2.0 release enabled international researchers to replicate on H100 clusters.
  • ByteDance Doubao deployment: ByteDance's proprietary inference stack runs on H20 at scale; announced plans to shift 30%+ to Ascend by 2026.

Evidence timeline

Sources

FAQ

What is the best available GPU for AI in China today?

For international-standard training: Nvidia H20 (legally available, limited compute vs H100). For domestic alternatives: Huawei Ascend 910C (~60-70% of H100 throughput). Ascend requires CANN software stack familiarity; H20 works with standard PyTorch/CUDA.

How do U.S. export controls affect Chinese AI development?

BIS rules (Oct 2023, updated May 2024) restrict H100, A100, and AMD MI300 direct export to China. Chinese labs must use H20 (downgraded Nvidia) or domestic chips (Ascend, Cambricon, Biren) for new capacity. Pre-2023 H100/A100 inventory is still in use. The controls have accelerated domestic chip investment but have not stopped Chinese AI capability growth.

Is Huawei Ascend competitive with Nvidia for AI training?

For Huawei-optimized models (PanGu series): Ascend 910C delivers comparable training throughput. For general PyTorch workloads: ~60-70% of H100 FP16 throughput, with CANN compatibility improving. The software ecosystem gap is narrowing but remains the primary friction for teams used to CUDA.

What is Nvidia H20, and why does it matter for China?

H20 is Nvidia's China-compliant GPU — a downgraded H100 with ~33% of the compute at similar price. It runs CUDA normally, making it the easiest Nvidia option for Chinese deployments. Chinese hyperscalers ordered ~$12B of H20 in 2024, making it the de facto standard Nvidia chip in China's new AI infrastructure.

What signals should trigger a review of your China compute strategy?

Action triggers: (1) BIS rule updates that change what hardware Chinese entities can access; (2) Ascend 910D or successor announcement — may shift the cost-performance calculus; (3) Major hyperscaler disclosure that a model was trained primarily on domestic chips — signals ecosystem maturity.

Are Biren and Cambricon real alternatives to Huawei Ascend?

Biren (BR100) and Cambricon (MLU series) are production-grade but have limited software ecosystem adoption vs. Ascend. Biren is primarily used in its own cloud. Cambricon is strong for inference (INT8/INT4) and used by select Chinese cloud providers. Neither has the enterprise scale or software support of Ascend.

Search angles this page supports

China AI chip Huawei Ascend Nvidia H20 China export controls AI chips China compute AI GPU China 2026 Biren Cambricon

Related

Go deeper

Last updated: 2026-06-23 · Policy: Editorial standards · Methodology