China's AI Chips and Export Controls: Ascend, SMIC — 2026-10-01
DeepSeek and Huawei open-sourced chip programming tools on September 30, positioning domestic AI silicon as a viable Nvidia alternative in China. Huawei's rotating chairman claimed Ascend chips now outsell Nvidia in China, while Beijing weighs selective approvals for US chip purchases by top AI firms—signaling pragmatic adaptation rather than pure autarky.
China's AI Chips and Export Controls: Ascend, SMIC — 2026-10-01
Top developments
DeepSeek Open-Sources Huawei Chip Tools, Targeting Nvidia's Dominance
On September 30, DeepSeek publicly released a suite of software tools designed for Huawei's Ascend platform, directly mirroring capabilities previously exclusive to Nvidia. The release includes TileLang (a high-level compilation language), computational libraries, and distributed communication libraries—forming a complete alternative software stack.
This move eliminates a critical software barrier for developers wanting to port models from Nvidia's CUDA ecosystem to Huawei hardware. Chinese domestic media praised the release as breaking Nvidia's "monopoly," with Sina Weibo commentary noting this was "previously basically monopolized, now there's one more choice."
Huawei Claims Ascend Sales Now Exceed Nvidia in China
On September 30, Huawei's rotating chairman Eric Xu asserted that domestic Ascend chip sales have surpassed Nvidia's in the Chinese market. While acknowledging that Ascend chips may not match Nvidia's performance, Xu highlighted supply-chain resilience: "They won't get banned at a moment's notice."
External analysts corroborate this shift: Bernstein data cited by Startup Fortune shows Nvidia's China AI chip market share collapsed to 8% by 2026, while Huawei's Ascend line climbed to 50% share.
Beijing Weighs Selective Nvidia H200 Approvals for ByteDance and Alibaba
On September 28, Seoul Economic Daily reported that China is considering approval for ByteDance and Alibaba to purchase Nvidia RTX Pro 5500 AI chips—a pragmatic signal that Beijing will permit limited US semiconductor access even as it accelerates domestic substitution.
This marks a tactical shift: instead of blanket embargoes, China's Ministry of Industry and Information Technology (MIIT) is gatekeeping Nvidia purchases by selecting which firms can buy. The move suggests confidence in Ascend's sufficiency for baseline workloads while preserving strategic flexibility for cutting-edge tasks.
Local view
Chinese financial and tech media are framing the DeepSeek-Huawei partnership as a decisive moment in "independent" AI supply-chain development. Sina Finance (新浪财经) highlighted that Huawei's new Ascend 960 supernode, unveiled September 17, supports million-card clusters with proprietary UnifiedBus interconnect technology—reducing dependence on Nvidia's networking stack.
Huxiu (虎嗅), a prominent tech analysis platform, framed 2026 as a bifurcation year: China's GPU ecosystem is splitting into four competing paths (Huawei Ascend, Cambricon, Moore Threads, Hygon/Xinyun), each targeting different workloads and customers.
21jingji (21经济网) reported on Huawei's September 17 full connect conference, where the firm disclosed plans to ship 75 million Ascend chips in 2026, targeting 375–525 billion yuan (US$52–73 billion) in AI chip revenue—a 60% uplift from 2025 targets.
Context & numbers
Market share collapse: Nvidia's China AI chip share fell from dominant levels to 8% in 2026, while Huawei's Ascend captured 50% of the domestic high-end AI chip market.
Production targets: Huawei aims for 750,000 Ascend chip shipments in 2026, up from prior guidance. Cambricon Technologies achieved 64.97 billion yuan in 2025 revenue (up 453% year-over-year) and 20.59 billion yuan net profit—the first pure cloud AI chip design firm in China to post a full-year profit.
HBM bottleneck: SMIC's foundry capacity for Ascend production remains constrained by high-bandwidth memory (HBM) availability from CXMT. As of early 2026, Huawei's die bank from TSMC was exhausted; future Ascend volumes depend entirely on SMIC's domestic 7nm/5nm wafers paired with CXMT HBM packaging.
US export policy: Effective January 15, 2026, the US Bureau of Industry and Security (BIS) shifted from "presumption of denial" to "case-by-case review" for Nvidia H200 and AMD MI325X exports to China, provided security requirements are met.
On the radar
-
Ascend 950DT launch Q4 2026: Huawei's training-class successor to 950PR debuts in Q4, with 960DT and 960PR variants planned for Q1–Q3 2027. This cadence concentrates SMIC's wafer allocation cycles annually.
-
Cambricon's 2026 earnings: With 453% revenue growth in 2025, the chip design firm is moving from speculative to cash-generative; watch for H1 2026 earnings to confirm margin sustainability.
-
MIIT selective licensing: Beijing's case-by-case approval for Nvidia chips (ByteDance, Alibaba) suggests a formal licensing framework may be formalized in Q4 2026—worth monitoring for additional firm approvals or restrictions.
-
SMIC yield and timeline: Domestic industry sources suggest SMIC's 7nm Ascend yield rates are stabilizing but remain 5–10% below TSMC parity; any further yield gains could unlock 20–30% additional Ascend volume by mid-2027.
Sources cited:
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
