HomeAtlasDashboardChartsMetalsResearchPersonsBubble MapGlobeLeadershipLensesMethodologyGlossarySource ↗
Independent research & opinion. Gradings are automated / LLM-assisted and may contain errors or hallucinations; nothing here is a statement of fact, financial advice, or an accusation of wrongdoing by any party. Claims about identifiable people or organizations reflect public records + good-faith interpretation; intent is not inferred from association. Methodology & disclaimer.

China's AI stack beyond DeepSeek — the full domestic LLM field, the Huawei Ascend hardware push, and the built-in state-alignment / censorship layer

Built 2026-06-21 from research/spec-china-ai-stack-censorship.json. Fixes the "efficiency counter-thesis only names DeepSeek" gap. Companion to fin-google-amazon-anthropic-meta (the Jevons/efficiency thesis), spec-semiconductor-logistics-standards (SMIC/export controls), macro-crqc-quantum-landscape (China programs), and the planned local/uncensored-AI block.

Frame. China fields a deep, competitive, largely open-weight LLM ecosystem and a mandatory state-alignment layer baked into all of it. Three layers: (1) the labs, (2) the Huawei Ascend hardware push to escape US export controls, (3) the CAC censorship layer. The twist: many models are open-weight, so the censorship is de-censorable downstream — the bridge to the uncensored-AI counter-ecosystem. Discipline. Labs/models/specs/shares/CAC facts are fact (web-verified). The "capability real and censorship built-in" duality is the graded finding. Composition guard: "China's models" = many independent labs under a common regulator, not one mind. Overlay; excluded from the proofs.

1. The labs (far beyond DeepSeek)

Market signal: by April 2026, Xiaomi+Alibaba+MiniMax+Zhipu+DeepSeek+StepFun > ~45% of weekly token volume on OpenRouter. Fact.

2. The hardware push (sovereign silicon, incomplete)

To escape US export controls China is building a domestic accelerator stack. Huawei Ascend leads: the 910C (96GB HBM2e, ~A100-class FP16, SMIC 7nm) targets ~600k units in 2026 (up to ~1.6M dies), with 950PR (Q1-26) and 950DT (Q4-26) + SuperPoD interconnect next. Cambricon, Biren, Moore Threads round out the field; SMIC fabs them.

Reality check: substitution is incomplete — DeepSeek's R2 was delayed by Ascend training instabilities, forcing a return to Nvidia H800s for critical runs; HBM is the bottleneck. Real-and-improving, not yet achieved. Fact.

3. The censorship layer (alignment by state mandate)

Every public generative-AI service must (a) register its algorithm with the CAC + pass a security assessment (748 services filed by Dec 2025), and (b) uphold "Core Socialist Values" — no content that could "subvert state power, harm national security, or undermine social stability." CAC tests models (ByteDance, Alibaba, …) on Xi Jinping and sensitive topics; models refuse/deflect on Tiananmen 1989, Taiwan, Xinjiang, Hong Kong by design. This is alignment-by-state-mandate — structurally different from Western corporate/voluntary RLHF safety. Fact (Interim Measures 2023; registration count; socialist-values testing).

4. The open-weight paradox

The censorship lives in the weights + serving layer — but most models ship open-weight (Qwen Apache 2.0; GLM/MiniMax MIT; Hugging Face). Open weights can be fine-tuned / "abliterated" to remove refusals downstream (de-censoring tooling like Heretic; uncensored re-releases). So Chinese models are both censored-when-served-from-China and de-censorable raw material once public — the paradox bridging to the local/uncensored-AI counter-ecosystem. Open licenses fact; de-censorability demonstrated but quality-variable (contested).

5. The honest reading

Both true at once: (1) Chinese AI capability is real, plural, fast, and unusually open — not one DeepSeek but a dozen competitive labs, several frontier-adjacent at far lower cost, increasingly on domestic silicon; (2) the censorship is real, mandatory, and built-in — a state-alignment layer Western models lack. The efficiency/Jevons counter-thesis holds (cheap, abundant, open models accelerate diffusion) with the control caveat (abundance is state-aligned at the source) — and the open-weight paradox means that control is leaky once weights escape.

Sources: Interconnects — Chinese open model builders; DigitalApplied — Chinese AI Q2 2026 market share; Huawei 600k Ascend 910C in 2026; SemiAnalysis — Ascend ramp / HBM bottleneck; CAC Interim Measures for Generative AI; Euronews — 'core socialist values' testing.

← Research index · structured data: spec-china-ai-stack-censorship.json · spec-china-ai-stack-censorship.md