China’s Domestic GPU Industry 2026: A $100 Billion Substitution Opportunity as US Sanctions Reshape the AI Chip Market

Executive Summary

China’s AI chip market is undergoing a seismic transformation. With a total addressable market projected to reach ¥715.3 billion (~$98 billion) by 2029 and US export controls accelerating the decline of NVIDIA’s market share from ~55% in 2025 to a forecasted 8% by 2026, an unprecedented substitution window has opened for domestic GPU makers. This article provides a comprehensive analysis based on the “Domestic GPU Chip Industry Deep Dive: Computing Power Demand Explosion, Domestic Substitution Accelerating” report published by Zheshang Securities on July 13, 2026, covering market sizing, competitive dynamics, financial performance, technology roadmaps, and the broader industry context.

SEO Keywords: China domestic GPU, AI chip market China, NVIDIA China ban, GPU substitution, Iluvatar CoreX, Tianshu Zhixin, Cambricon, US chip sanctions, AI computing power China, domestic GPU market size, H20 export license, China AI capex 2026

---

Part I: The $100 Billion Market — China’s AI Chip Landscape

1.1 Overall AI Chip Market: Rapid Expansion

China’s AI chip market has entered a phase of explosive growth. According to Frost & Sullivan data cited in the report, the total market (by revenue) has more than quadrupled between 2022 and 2025, with no signs of slowing:

Table 1: China AI Chip Market by Category (Revenue, ¥100 Million)

Year	General-Purpose GPU	NPU	Other ASIC	Total
2022	534	101	34	669
2023	961	236	74	1,271
2024	1,546	456	173	2,175
2025E	3,542	552	143	4,237
2026E	3,500	812	162	4,474
2027E	4,638	1,106	205	5,949
2028E	5,873	1,315	222	7,410
2029E	7,153	1,552	276	8,981

Key takeaways:

• General-purpose GPUs dominate: They will consistently account for 75–85% of the total AI chip market through 2029, far outpacing NPUs and other ASICs.
• Market size at scale: The total AI chip market is projected to approach ¥900 billion by 2029, more than a 13x increase from 2022.
• GPU is the primary battleground: While NPUs are growing, the real prize — and the focus of domestic substitution — remains the general-purpose GPU segment.

1.2 The Domestic Substitution Story: From Marginal to Mainstream

The most consequential shift in the report is not the overall market growth but the distribution between imported and domestically produced GPUs:

Table 2: China General-Purpose GPU Market — Imported vs. Domestic (Revenue, ¥100 Million)

Year	Imported GPU	Domestic GPU	Domestic Share
2022	524	10	1.9%
2023	930	31	3.2%
2024	1,491	55	3.6%
2025E	2,376	166	6.5%
2026E	3,165	335	9.6%
2027E	3,948	690	14.9%
2028E	4,550	1,323	22.5%
2029E	4,932	2,221	31.0%

The growth mathematics are striking:

Metric	CAGR (2025–2029)
Overall General-Purpose GPU Market	19.2%
Imported GPU Segment	Subdued growth
Domestic GPU Segment	91.4%

Domestic GPUs are projected to grow from a ¥166 billion market in 2025 to ¥2,221 billion in 2029 — a 13.4x increase in just four years. By 2029, nearly one in every three yuan spent on GPUs in China will go to a domestic supplier.

1.3 Training vs. Inference: Two Markets, Two Dynamics

The report further segments the GPU market into training and inference, revealing divergent growth trajectories:

Table 3: China Training GPU Market (Revenue, ¥100 Million)

Year	Imported	Domestic	Total
2022	351	8	359
2023	570	23	593
2024	811	34	845
2025E	1,236	96	1,332
2026E	1,452	129	1,581
2027E	1,623	170	1,793
2028E	1,941	274	2,215
2029E	1,927	388	2,315

• Training market CAGR (2025–2029): 14.8%
• Domestic training GPU CAGR (2025–2029): 41.8%

Table 4: China Inference GPU Market (Revenue, ¥100 Million)

Year	Imported	Domestic	Total
2022	173	2	175
2023	359	8	367
2024	680	21	701
2025E	1,140	69	1,209
2026E	1,713	205	1,918
2027E	2,324	521	2,845
2028E	2,609	1,049	3,658
2029E	3,005	1,833	4,838

• Inference market CAGR (2025–2029): 41.4%
• Domestic inference GPU CAGR (2025–2029): 126.9%

Critical insight: The inference market is both larger and faster-growing than the training market. By 2029, inference will represent approximately 68% of the total GPU market (¥4,838 billion out of ¥7,153 billion). Domestic inference GPU growth at 126.9% CAGR is the single most aggressive forecast in the report — and the segment where Chinese companies appear to have the clearest near-term competitive advantage, as inference workloads are generally less demanding than training on raw floating-point performance.

---

Part II: The Catalyst — US Sanctions Timeline

The domestic GPU substitution story cannot be understood without the context of escalating US export controls. The report catalogs the sanctions timeline in detail:

Table 5: US Chip Export Controls Against China — Complete Timeline

Round	Date	Event
Round 1	Oct 7, 2022	US mandates performance-capped A100 shipments to China
Round 2	Oct 2023	US tightens AI chip restriction criteria
Round 2	Oct 18, 2023	US halts A800/H800 chip shipments
Round 3	Nov 17, 2023	US bans RTX 4090 consumer GPU sales to China
Round 4	Apr 9, 2025	H20 export now requires BIS license
Round 4	Apr 14, 2025	H20 export license requirement made indefinite

The progression is clear and escalatory: from performance caps (2022), to model-specific bans (2023), to consumer-grade restrictions, and finally to a blanket licensing regime for the last legally exportable NVIDIA data-center GPU — the H20. The April 2025 indefinite license requirement effectively closes the door on all high-performance NVIDIA data-center GPUs for the Chinese market.

The Market Share Vacuum

According to Bernstein estimates cited in the report:

Year	NVIDIA Share of China AI Chip Market
2025	~55% (accelerator card market) / ~40% (broader AI chip market)
2026E	~8%

This represents a collapse of over 40 percentage points in market share within approximately one year. The report characterizes this as a “historic substitution window” — and the numbers bear this out. Measured against the 2026E total GPU market of ¥3,500 billion, a 40-point share transfer represents roughly ¥1,400 billion (~$192 billion) of addressable market migrating from NVIDIA to domestic alternatives.