Home / Deployment Hardware

Deployment Hardware

What models run on: nine domestic chip roadmaps, LLM all-in-one machines, and supernode cluster solutions.

Domestic AI chip roadmaps

Training / inference positioning and 2026 status (compiled from public information)

Vendor	Chip roadmap	Positioning	2026 status
Huawei Ascend	910B → 910C → 950PR / 950DT → 960 / 970	Unified training and inference, full-stack CANN	910C in volume production; 950 series launching in two steps in 2026, with in-house HBM
Cambricon	MLU370 → 590 → 690	Dual-use for training and inference	590 in commercial volume use, 690 ramping to mass production, revenue growing rapidly
Hygon	DCU Gen1 → Gen2 → Gen3	DCU with a CUDA-like ecosystem	Gen3 DCU scaling up; integration with Sugon completed
Moore Threads	MTT S4000 → S5000 + KUAE clusters	Full-function GPU	Listed on the STAR Market in January 2026; S5000 targets AI training
MetaX	Xiyun C500 → C600	Unified training and inference	C500 in volume production; listed on the STAR Market in February 2026
Biren	BR100 series → Bili 106E/106M	Shift toward inference	Thousand-GPU clusters delivered to telecom operators; HK listing in progress
Enflame	Yunsui T20/T21 + i10/i20 + S60	Separate training and inference product lines	Steady shipments; embedded in the Tencent ecosystem
Kunlunxin	P800 → M100 → M300	Baidu internal use + external sales	30,000-GPU cluster lit up; M100 in early 2026
Iluvatar CoreX	Tiangu 100 / Zhikai 100	Separate training and inference product lines	Steady shipments; IPO tutoring in progress

LLM all-in-one appliances

Out-of-the-box private deployment solutions (2025–2026)

Huawei FusionCube A3000: Dual-node support for full-parameter (671B) or quantized DeepSeek deployment; Atlas 800I A2 inference server (8×910B).
Chipmaker-backed offerings: Hygon, Moore Threads, MetaX, Biren, Enflame, Cambricon, Kunlunxin, and Iluvatar CoreX have all launched DeepSeek all-in-one appliances / integrated hardware-software solutions.
Server and industry players: Inspur YuanNao EPAI, H3C Lingxi, Lenovo WenTian, FusionScene, plus industry appliances from Baidu Qianfan, Tencent Cloud, and Alibaba Cloud.
Price range(channel estimates): Distilled / small-to-mid-parameter inference appliances run roughly in the hundreds of thousands of RMB; full 671B appliances roughly in the millions of RMB. Subject to vendor quotes.

Supernodes and clusters

The mainstream architecture in the 10,000-GPU era

Solution	Vendor	Scale	Key points
CloudMatrix 384	Huawei	384×910C	All-optical interconnect supernode; commercial deployments across multiple AI computing centers
Atlas 950 SuperPoD	Huawei	8,192 GPUs	Planned for 2026 Q4; 960 reaches 15,488 GPUs (2027)
Kunlunxin supernode / Tianchi	Baidu	256–30,000 GPUs	Self-built 30,000-GPU cluster lit up in 2025
KUAE (Kua'e)	Moore Threads	1,000–10,000 GPUs	Full-Featured GPU Cluster Solutions

Selection Tip: The model–chip compatibility matrix is key — mainstream open-source models like DeepSeek / Qwen / GLM / Kimi have all been adapted to Ascend and major domestic chip inference stacks. Before purchasing, be sure to confirm the target model's official compatibility list and operator coverage.