Breaking

Chinese-Made AI Chips Take 41% of Local Market as Nvidia Retreats on H20

July 6, 2026 at 22:36 EDT

Infra & Chips
Policy & Regulation
Foundation Models

US-China technology tensions are reshaping China's AI infrastructure buildout, with Chinese firms shifting from Nvidia's advanced accelerators to domestic silicon. According to IDC, domestic makers accounted for about 41% (roughly 1.65 million units) of China's AI accelerator server market in 2025, narrowing Nvidia's lead.

July 7, 2026 · AI Infrastructure

China Turns Away From Nvidia, Toward Homegrown Silicon

Under tightening US export controls, Chinese firms plan to route nearly half their AI accelerator budgets to domestic chips within a year — reshaping the global map of AI infrastructure.

46%

of AI accelerator budget headed to domestic chips (up from 30% today)

80%

of firms say AI infrastructure spend is running over budget this year

1.65M

domestic AI GPUs shipped in 2025 — 41% of ~4M AI servers

China AI chip market share · 2025 → 2026

Nvidia's grip collapses as Huawei and Cambricon absorb the shift.

40%

Nvidia

40%

50%

Huawei

Cambricon

lighter bar = 2025 solid bar = 2026

Domestic momentum

Ascend 950 (2026): ~1 PFLOP FP8, 128–144 GB memory
Atlas 950 SuperPoD: 8,192 chips → 8 exaflops FP8
Cambricon 2026 target: 500,000 units
GLM-Image & LongCat-2.0 trained entirely on domestic stacks

The gap that remains

Ascend 910C ≈ 60% of H100 inference performance
Weaker for large-scale training workloads
Nvidia's CUDA software ecosystem still leads
"Good enough" for inference — not yet for frontier training

Export controls meant to slow China have accelerated its drive to build a self-sufficient AI stack — from chips to servers to frameworks like MindSpore.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.