--- title: "Inside China's AI Stack: 6,000 Companies and a Full Chain" date: 2026-08-22 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-inside-china-ai-stack language: en --- # Inside China's AI Stack: 6,000 Companies and a Full Chain > Silicon, clusters, models, applications — and the connective tissue between them. A map of how China's AI industry actually fits together, and why the interlocking matters more than any single company. It is tempting to explain China's AI progress by pointing at one company. That is almost always the wrong explanation. The more accurate picture is of a **stack** — and of companies deliberately built to interlock. China has **more than 6,000 AI companies**, with an industrial chain covering intelligent chips, compute clusters, model development and application scenarios. ## Key takeaways - **Silicon:** Huawei Ascend shipped about **812,000 accelerators in 2025** (~49% of the domestic market); Alibaba's T-Head about **265,000**; Baidu Kunlun and Cambricon roughly **116,000 each**; Hygon about **82,500**. - **Compute infrastructure:** the Ascend 384 supernode has been deployed in **more than 500 sets**; an **11,520-card** cluster went live in Shaoguan in July 2026; a fully domestic **100,000-card** cluster has been built; the Atlas 950 SuperCluster is designed to scale to **500,000 cards**. - **Models:** DeepSeek, Alibaba's Qwen, Zhipu's GLM, Moonshot's Kimi and MiniMax ship frontier-class open models. In 2025, the US produced **50 mainstream large models and China 30**. - **Applications:** Kuaishou's Kling has **over 100 million users**; Unitree has produced **more than 18,000 humanoids**; AgiBot about **15,000**. - **Software:** CANN is fully open-sourced with **3,000+ partners**, and new Chinese models ship with **Day-0 support across nine domestic chips**. ## Read the stack bottom-up **Layer 1 — Silicon.** The chip tier is no longer a single point of failure. Huawei supplies the volume; Cambricon and MetaX push specification; Hygon offers a CUDA-compatible migration path; Alibaba and Baidu build for their own clouds. Diversity matters because it creates second sourcing under supply constraints. **Layer 2 — Systems.** Because individual chips are constrained, the cluster became the unit of compute. Supernodes, 10,000-card installations and liquid-cooled pods are where Chinese systems engineering converts a component disadvantage into a system-level advantage. **Layer 3 — Software.** This was the acknowledged weak link, and it is where the most consequential 2026 moves happened: CANN open-sourced, tooling matured, and models shipping with Day-0 domestic-chip support. Without this layer, layers 1 and 2 are inert. **Layer 4 — Models.** Chinese labs compete on capability but win on the combination of capability and cost. The result is that Chinese open models are the default choice for a large share of developers globally, including in the United States. **Layer 5 — Applications.** This is where the stack monetises: video generation, humanoid hardware, industrial inspection, medical robotics, agricultural advisory. ## Why interlocking is the strategy The most interesting structural feature is vertical integration *across* layers, not just within companies. **Alibaba** runs a chip-design arm (T-Head), a model family (Qwen) and a cloud business. The stated effect is a shortening of the cycle from chip design to commercial deployment, because the cloud provides an immediate test bed and the model provides an immediate workload. **Huawei** runs silicon (Ascend), a software stack (CANN), systems (Atlas supernodes) and cloud services — and has been described as the only domestic vendor with a commercially scaled supernode that has trained state-of-the-art models. **DeepSeek and Unitree** signed a strategic memorandum to work jointly on large models and embodied intelligence — a model company and a body company deciding that the seam between them is where the value is. This is a response to a specific constraint. Under export controls, no single layer can be assumed reliable, so the layers are being wired together to reduce dependence on any external link. ## The coordination layer Above the companies sits a coordinating apparatus: the **AI Plus** initiative setting diffusion targets, ministry-level standards for emerging sectors like brain–computer interfaces, and international institutions such as the **World Artificial Intelligence Cooperation Organization**. Western analysis has described this combination — research goals, industrial capacity, sensible pricing and coordinated deployment — as China's distinctive AI path. It is not a command economy producing chips by decree; it is a densely networked industry with a government setting direction and buying the early output. ## The vulnerabilities Three are worth naming. **Manufacturing constraints.** Wafer fabrication, HBM and advanced packaging cap the entire stack. Design capability and capital are not the binding limits. **Software ecosystem depth.** Chinese tooling has improved dramatically but still carries a smaller third-party ecosystem than CUDA-era incumbents. **External demand risk.** A significant share of model adoption comes from outside China, which is both the achievement and the exposure — it is the layer most sensitive to geopolitics. ## What to watch The single most informative indicator is not any benchmark. It is whether **cost parity on domestic silicon holds at production scale**. Early reports suggest per-token serving costs on domestic clusters have reached parity with mainstream NVIDIA GPUs. If that holds as volumes grow, the stack is self-sustaining. If it does not, the system still works — but at a subsidy. *Layer statistics from third-party market estimates for 2025 and company disclosures in 2026.* --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-inside-china-ai-stack Free to quote with attribution and a link to the original.