When people picture a "Chinese GPU company," they usually imagine a challenger to Nvidia's data-center accelerators. Jingjia Micro (景嘉微) is the opposite kind of story — and the more honest one for understanding where domestic chips actually land.
The Changsha firm has long shipped domestically designed GPUs, but its widely shipped JM9 series performs at a level comparable to Nvidia's GeForce GTX 1050, a consumer card from 2016. That is not a failure. It is a window into a different market: trusted, domestically sourced graphics for aircraft, defense systems, and government workstations where independence matters more than peak teraflops.
Key takeaways
- Jingjia Micro (景嘉微), listed as 300474 on Shenzhen, has mass-produced home-grown GPUs for domestic markets.
- Its JM9 series is in volume delivery and performs near an Nvidia GeForce GTX 1050 — a 2016-era consumer GPU, not a modern data-center part.
- The newer JM11 series has reached small-batch shipment and, per the company, has been adapted to domestic CPUs (Phytium 飞腾, Kunpeng 鲲鹏, Hygon 海光, Loongson 龙芯) and operating systems (Kylin 银河麒麟, UnionTech UOS 统信).
- A subsidiary's CH37 edge AI SoC, unveiled in August 2026, integrates CPU/GPU/NPU on one chip with peak 64 TOPS at INT8.
- 2025 R&D spend was about 427.6 million RMB (≈ US$59M), 59.4% of revenue; a GPU industrialization project carries 3.781 billion RMB (≈ US$524M) of planned investment.
Where the chips actually go
Jingjia did not start in gaming or cloud. It grew out of graphics for specialized and embedded systems — cockpits, displays, signal processing — where buyers prioritize chips they can fully audit and source at home. The JM9's GTX 1050-comparable performance is enough for 2D/3D rendering, instrument panels, and workstation graphics in those contexts. It is not a training accelerator, and the company does not claim it is.
The JM11 generation is the bridge toward broader use. Jingjia has emphasized compatibility for the JM11 generation: the GPU now runs atop domestic CPUs and operating systems, forming a "CPU + GPU + OS" trusted stack for government, energy, and industrial-design software. Adapting to four domestic CPU architectures and two domestic OSes is the unglamorous work that makes a Chinese GPU usable outside a niche.
The edge-AI pivot
The more forward-looking move is the CH37, an edge AI system-on-chip from subsidiary Wuxi Chenghengwei (无锡诚恒微), unveiled in August 2026. It crams CPU, GPU, and NPU — ten compute units in one package — with peak throughput of 64 TOPS at INT8, aimed at embodied AI, edge computing, and sensing. That is a different race from data-center training: low power, small footprint, local inference. It is also where Jingjia's "GPU + edge AI SoC" dual strategy is supposed to pay off.
The backdrop is 信创 (domestic IT substitution): a multi-year push for government and state-linked systems to run on Chinese CPUs, operating systems, and GPUs. That demand, not consumer gaming, is what sustains a GPU maker whose flagship performs like a decade-old Nvidia part. It is a market shaped by domestic procurement, and it explains why Jingjia's JM11 compatibility work — adapting to four domestic CPU architectures and two domestic OSes — matters more than raw benchmarks. For a trusted-computing buyer, a chip it can fully audit and source at home is worth more than a faster one it cannot.
Reading the numbers honestly
Jingjia's R&D intensity is striking — about 427.6 million RMB (≈ US$59M), or 59.4% of 2025 revenue — with roughly 1,060 R&D staff, about 68% of employees. A 3.781 billion RMB (≈ US$524M) "high-performance general-purpose GPU" project is in progress. As of an October 2026 report, the company's market capitalization was about 25.8 billion RMB (≈ US$3.6B). These are reported figures, not estimates I re-derived.
The honest framing: Jingjia is positioned in trusted and embedded graphics through domestic substitution, and is an emerging edge-AI player. It is not, and does not claim to be, a competitor to Nvidia's H100 or Huawei's Ascend in large-model (大模型) training.
What readers can do now
- If you evaluate "domestic GPU" claims, ask which workload. A JM9 beating a 2016 consumer card is a real milestone for trusted systems, irrelevant to LLM training.
- If you build edge products, watch the CH37-class SoCs — that is where Chinese GPU makers may win exportable, non-frontier business.
- If you follow the supply chain, track "CPU + GPU + OS" certification as the real adoption metric, not benchmark scores.
Honest limitations
Performance comparisons (JM9 ≈ GTX 1050) and product-status details come from the company's own statements and Chinese technology-media reporting (iJiwei, Baidu Baike), not from independent benchmark publications I could access. The CH37 64 TOPS figure is a company-stated spec. Financial figures (R&D 427.6M RMB, 59.4% of revenue, 3.781B RMB project, ~25.8B RMB market cap) are from the same Chinese media citing company disclosures. I have not verified JM11's exact silicon specifications or benchmarked any part myself, and I deliberately avoid presenting Jingjia as a frontier AI-accelerator vendor.