Chinese CPU maker Zhaoxin used an information technology application innovation forum in Shanghai to present an AI computing stack built on its own ZX86 instruction set. The company's pitch centers on a single lineup that runs from the KH-50000 server processor to the KX-7000 for PCs and embedded systems, covering AI servers through to AI endpoints. Partner hardware was on display alongside it, giving a sense of how close China's domestic CPUs now are to practical AI deployment.

Staking Out an AI Position at the Shanghai Forum

The event was held as the AI innovation track of the 2026 Shanghai Information Technology Application Innovation Industry Development Forum[1]. Discussion focused on domestic computing platforms, home-grown large language models and AI agents. Zhaoxin, as a Shanghai-based supplier of domestic compute silicon, presented its work on applying ZX86 processors to AI workloads.

Version 2.0 of the Shanghai State-Owned Assets Xinchuang Laboratory was also announced during the forum[1]. Built on ZX86 compute and developed jointly with carriers, it offers a cloud layer that handles multiple chip types at once, with services organized around four pillars: one-stop compatibility certification, digital transformation support for state-owned entities, low-code development and agent applications, and an open-source community. Signing ceremonies for application partners and compute-platform partners were held at the same time.

The ZX86 Instruction Set as Foundation

Zhaoxin repeatedly returns to the ZX86 instruction set as its foundation[1]. The company describes the approach as independent but not closed, compatible but not dependent, and oriented toward new applications. By building its own intellectual property framework while maintaining x86 compatibility, Zhaoxin argues it can keep improving future CPU products and applications on its own terms.

On top of that foundation sit the KH-50000 for servers and the KX-7000 for PCs and embedded systems, forming a continuous product family that spans cloud, edge and endpoint[1]. With AI compute demand spread across everything from on-device inference to datacenter training, sales vice president Chen Xiaomin explained in a panel discussion that the platform supports domestic GPUs, AI accelerators, large models and databases.

A 96-Core KH-50000 and a KX-7000 With Twice the Performance

The KH-50000 server processor is built on a chiplet design, in which multiple dies are packaged together[2]. A single chip integrates up to 96 cores and 96 threads, with a base clock of 2.2GHz or 2.6GHz and a maximum operating frequency of 3.0GHz. Level 3 cache reaches 384MB, memory support extends to 12 channels of DDR5 RDIMM for up to 3TB of capacity, and I/O includes up to 128 lanes of PCIe 5.0, ZPI and CXL plus 16 lanes of PCIe 4.0. The chip also handles acceleration instructions for China's SM2, SM3 and SM4 national cryptographic standards.

For expansion, the processor supports two-socket and four-socket configurations, allowing a single server to reach 384 cores[2]. The updated ZPI 5.0 interconnect is described as widening bandwidth while reducing latency and power draw.

On the client side, the KX-7000 uses Zhaoxin's own Century Avenue microarchitecture together with chiplet interconnects[3]. It supports up to 128GB of DDR5 or DDR4 memory, 24 lanes of PCIe 4.0, and I/O including USB4 and SATA3, with integrated graphics that handle DX12 and dual-stream 4K decoding. Zhaoxin puts compute performance at twice that of its previous generation and graphics performance at four times.

Partner Hardware Makes the Edge AI Case

Products from partners using Zhaoxin silicon were displayed throughout the forum[1]. The most striking is an AI endpoint from Lenovo Kaitian based on the KX-7000. Despite a chassis of just one liter, it delivers 640 TOPS per liter of compute density and can run mainstream local large models directly.

Servers and workstations included Lenovo Kaitian's KR722z G3 with two KH-50000 processors, high-capacity storage models, and AI servers pairing the CPU with domestic GPUs[1]. An all-in-one AI system from People's Daily Online and Xintou Zhike combines ZX86 CPUs with a Chinese GPU, running a proprietary large model on top of Kylin OS and domestic middleware, aimed at handling official documents and reports from drafting through review and management. Liquid-cooled AI inference workstations were also shown, tuned for government and enterprise offices that need quiet operation and easy deployment. Partners including ZTE and Seewo displayed ZX86-based AI PCs, laptops and interactive display panels.

That said, what was presented consists largely of exhibits and forum talks. No third-party performance verification or figures on actual deployment scale have been released. Adoption is clearly growing in sectors under strong domestic-sourcing pressure, but judging competitiveness in the broader market will require independent benchmarks and a track record of deployments.

Summary

Zhaoxin is assembling a continuous product family around its ZX86 instruction set, running from the 96-core KH-50000 to the KX-7000 in compact AI endpoints. The Shanghai forum served to demonstrate that breadth through partner servers, workstations and AI PCs. The shortage of third-party verification remains a real gap, but the effort is worth watching as a measure of how far China's domestic compute platforms can reach into AI workloads.

Source: https://www.zhaoxin.com/news_view.aspx?nid=2&typeid=289&id=2051

Source: https://www.zhaoxin.com/prod_view.aspx?nid=3&typeid=605&id=3601

Source: https://www.zhaoxin.com/news_view.aspx?nid=2&typeid=283&id=1857