--- title: "Jingjia Micro pushed its GPU from \"usable\" to \"good enough\" — and added an edge AI chip" date: 2026-10-09 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-jingjia-micro-gpu language: en --- # Jingjia Micro pushed its GPU from "usable" to "good enough" — and added an edge AI chip > Changsha-based Jingjia Micro (景嘉微) is shipping its JM11 GPU into government cloud desktops while spinning up a 64-TOPS edge AI SoC through a subsidiary. A look at its "cloud-to-edge" compute strategy and the R&D bill behind it. A graphics chip that only works in a demo is worthless. A Chinese maker most people have never heard of is now shipping one into government cloud desktops — and quietly attaching a second chip for the edge. Jingjia Micro (景嘉微, 300474.SZ) started life supplying graphics and display controllers for specialized, ruggedized equipment. In 2025 it made a sharper turn toward mainstream AI compute with a two-track plan: a high-performance GPU for servers and desktops, plus a small, low-power AI chip for devices out in the field. ## The JM11 GPU and the "32 sessions" test The headline product is the JM11 series of graphics processing chips. According to the company's investor communications and Chinese financial press, JM11 entered small-batch delivery in 2025 and is being used in government cloud desktops, industrial design, and geographic-information rendering. The concrete proof point the company repeats: a single domestic server built around JM11 can stably support 32 concurrent cloud-desktop users, with lossless encoding so each session stays sharp. That is a deliberately "boring" metric, and it is the right one. Cloud desktops are where a domestic GPU lives or dies on compatibility and stability, not on a benchmark score. Jingjia has paired the chip with home-grown operating systems (Kylin, openEuler) and, with partner KylinSec, shipped an all-domestic GPU virtualization solution into power-grid dispatch — infrastructure where "it didn't crash" matters more than "it was fast." ## The edge bet: CH37 and the shift to endpoints The more interesting move is toward the edge. Through a subsidiary, Chenghengwei (诚恒微), Jingjia is mass-producing the CH37 series of edge AI SoCs in small batches. The chip is built on a self-developed architecture and delivers about 64 TOPS at INT8 precision, with a dual-mode image signal processor that can fuse visible-light and infrared feeds at once. The target is exactly where AI is moving: industrial drones, robots, and embodied-intelligence (具身智能) devices that must see and decide without a data center. In August 2025 Jingjia invested about 220 million yuan (≈ US$30.6 million) to take control of Chenghengwei, a clear signal that edge AI is not a side experiment but a second front. The logic is sound: as efficient open models like DeepSeek push inference off the cloud and onto devices, the volume market is at the edge, not the server rack. ## The R&D bill behind the strategy None of this is cheap. Jingjia reported 2025 revenue of about 720 million yuan (≈ US$100 million), up 54% year on year, but a net loss of about 165 million yuan (≈ US$22.9 million) as it poured money into chips. Research and development hit 428 million yuan (≈ US$59.4 million) — nearly 60% of revenue — and its three-year cumulative R&D passed 1.04 billion yuan (≈ US$144 million). By late 2025 its R&D headcount had reached 1,060 people, up 80% in a year. That profile — growing sales, deep losses, soaring R&D — is the standard shape of a Chinese chip designer betting that "strategic" demand (domestic, controllable, secure) will mature faster than the P&L. The 2026 first quarter brought 84 million yuan (≈ US$11.7 million) in revenue as the build-out continued. ## Why "cloud-to-edge" is the real story Most coverage of Chinese chips fixates on the biggest, fastest training GPU. Jingjia is going the other way: a GPU good enough for desktops and virtualization, plus a tiny AI chip for endpoints, wired together into one software-and-ecosystem story. If the thesis holds, the company catches the same trend the whole industry sees — compute flowing from a few giant clouds down to millions of devices. The risk is execution. Edge AI is crowded, and a 64-TOPS part must compete on power, price, and tooling, not on nationality. The GPU side must keep proving compatibility as software stacks evolve. ## What readers can do now - If you spec domestic hardware, judge a GPU by concurrent-user and virtualization proofs, not by peak FLOPS claimed in a launch. - If you build edge devices, watch Chenghengwei's CH37 for drones and robots — dual-mode visible/infrared sensing is a genuine differentiator for inspection and security. - If you follow the sector, track R&D-as-%-of-revenue as a health signal; above 50% means the product is still being bought, not yet self-funding. ## Honest limitations Figures come from the company's investor-relations disclosures and Chinese financial reporting (Xinhua Finance and business press). Performance claims such as "32 concurrent sessions" and "64 TOPS" are company-stated, not independently benchmarked in this article. The edge-AI market-size figures sometimes cited by media are not restated here to avoid unverified numbers. This is analysis, not investment advice, and avoids political framing. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-jingjia-micro-gpu Free to quote with attribution and a link to the original.