TL;DR — China’s AI Is Now Driven by Two Reinforcing Curves
In August 2026, two mutually reinforcing accelerations hit China’s AI industry at the same time. The regulatory curve: as of mid-to-late August, 538 generative-AI services have completed filing and 263 applications/functions have completed registration — up about 55.5% from the 346 filings recorded in March 2025. Beijing alone has 242 filings, Jiangsu 80, and Shandong 49, with the regional coordination pattern already in place. The compute curve: on August 17 the domestic-AI-chip sector rallied, with brokers expecting Cambricon to deliver a Q3 earnings step-up, Huawei Ascend open-sourcing the CANN heterogeneous compute architecture, and Cambricon NeuWare maturing into a full software stack — domestic AI compute has moved from “usable” to “in a positive cycle”. Underpinning both curves, Xi Jinping’s keynote at WAIC 2026 (July 17-20) anchored the global governance stance, and the AI+ Action Plan for 2026 set the “100 industry-LLM benchmarks + 1000 high-value application scenes” target. Filing, compute, governance, and application are landing on all four sides at once.
1. The Regulatory Curve: 538 Filings + 263 Registrations, +55.5% in 18 Months
Filing data is the cleanest window into how many large models are actually in production in China.
| Dimension | March 2025 (baseline) | June 2026 (Beijing) | August 2026 (national) | Change |
|---|---|---|---|---|
| National filings | 346 | — | 538 | +55.5% (~18 months) |
| National registrations | — | — | 263 | Newly disclosed |
| Beijing filings | 209 | 242 | — | +33 (~4 months) |
| Shandong filings | — | — | 49 (incl. 5 new on Aug 4) | Steady growth |
| Jiangsu filings | — | 80 | — | 2nd in Yangtze Delta |
Sources: Cyberspace Administration of China (CAC) public info + Beijing CAC notice (2026-06-18) + Shandong CAC notice (2026-08-04/05) + Jiangsu CAC notice (2026-05-22) + Tencent Cloud Developer Community “Generative AI Large Model Filing Materials” (2026-08-18 19:04).
The split between “filing” and “registration” has been the regulatory design since the 2024 Interim Measures for the Management of Generative AI Services: filing applies to entities that directly provide a large-model service; registration applies to enterprises that call an already-filed model’s capability and serve it to the public. The first regulates the model, the second regulates the product that uses the model. With filings + registrations climbing from 346 to 538 + 263 = 801, large models have shifted from “experiments owned by a handful of vendors” to de facto infrastructure for hundreds of vertical applications.
2. The Compute Curve: August 17 Domestic-AI Rally, Ascend + Cambricon in Sync
On August 17, A-share domestic-AI-compute names opened higher and stayed strong; Cambricon and Hygon were added by multiple brokers to “Q3 earnings likely to step up” baskets. On August 18 at 10:14 the sector diverged intraday, with semiconductor equipment names relatively stronger.
| Date | Event | Industry meaning |
|---|---|---|
| 2025-12 | Huawei Ascend open-sourced the CANN heterogeneous compute architecture | Reduces CUDA dependency |
| 2025-12 | Cambricon polished the Cambricon NeuWare base software platform | Software stack is maturing |
| 2025-02-07 | DeepSeek completed adaptation with 16 domestic AI chip makers (Ascend, Metax, Iluvatar, etc.) | Model × chip decoupling |
| 2026-08-17 | Domestic-AI compute sector rallied | Market repriced the substitution thesis |
| 2026-08-18 | Semiconductor equipment led the sector | Capacity-expansion expectations front-loaded |
Sources: so.html5.qq.com “AI Compute Prosperity Continues to Validate” (2026-08-17) + CLS “Domestic Compute Enters a Positive Cycle” (2026-08-18) + CITIC Securities research note (2025-02-07) + so.html5.qq.com “Ascend + Cambricon Drive Comprehensive Transformation of China AI Compute” (2025-12-30).
Key read: If the 801 filed + registered products are to actually run, each one needs inference compute. A 100% dependence on overseas GPUs would mean that the larger the filing count, the larger the geopolitical exposure. The August 17 sector rally + Ascend / Cambricon software-hardware alignment = domestic compute has moved from “policy slogan” to “buyable and deployable in market” — this is the infrastructure precondition for the filing curve to keep climbing.
3. The Governance Curve: Xi Jinping’s WAIC 2026 Keynote
On the morning of July 17, 2026, President Xi Jinping attended the opening ceremony of the 2026 World Artificial Intelligence Conference & High-Level Meeting on Global AI Governance at the Shanghai World Reception Hall and delivered a keynote titled “Building a Fair and Reasonable Global AI Governance System Together”. This is the highest-level Chinese statement in WAIC history.
| Dimension | Key points of Xi’s speech | Industry meaning |
|---|---|---|
| Global governance | A “fair and reasonable” international governance system | China shifts from AI-governance “follower” to “rule proposer” |
| Development pace | ”A new round of scientific-technological revolution and industrial transformation is accelerating” | AI is elevated to a “century-level change” policy priority |
| Sovereignty and openness | Both “opportunity” and “governance challenge” | Domestic “filing + compute” become governance levers |
| Multilateralism | China has broadly invited governments, academia, and international organizations | Window for Chinese AI governance standards to “go out” |
Sources: Guangming Daily “Lead AI Development with Human Wisdom and International Consensus” (2026-07-19) + China Daily “Building a Fair and Reasonable Global AI Governance System Together” (2026-07-18) + so.html5.qq.com “WAIC 2026 Side Notes” (2026-07-21).
From the close of WAIC to mid-August, the governance signal has been transmitting into industrial policy: CAC filing notices from local branches accelerated, and Beijing / Jiangsu / Shandong formed a “leader + follower” pattern, nested with the AI+ Action Plan’s 100-benchmark / 1000-scene target.
4. The Application Curve: AI+ Action Plan — 100 Benchmarks + 1000 Scenes
The 2026 “AI+” Action Deepening Implementation Plan sets explicit quantitative targets:
| Target | Volume | Focus areas | Mechanism |
|---|---|---|---|
| Industry-LLM benchmarks | 100 | Manufacturing, healthcare, education, transportation, energy | Must reach human-expert level in specific scenes + commercial closed loop |
| High-value application scenes | 1000 | ”Open challenge” (揭榜挂帅) mechanism | Local governments + leading enterprises open the scenes |
| Domestic AI chip subsidies | Continued ramp-up | Huawei Ascend, Cambricon, Hygon | R&D subsidy + application support |
| Internet penetration | 70.4% → 79.7% | Digital infrastructure | Urban-rural gap narrowed by 8.5 pp |
Sources: CSDN “2026 Policy Released” (2026-07-24) + CAC “14th Five-Year” informatization report (2025-12-25) + Shenzhen MIIT “Action Plan for Accelerating the AI Terminal Industry (2025-2026)”.
Key observation: 100 benchmarks × 1000 scenes = 100,000 potential “model × scene” pairings. This gives the 538 + 263 filing-curve products a clear market exit, and gives the domestic-compute side (Ascend / Cambricon) a scale-effect channel through which “1000 scenes repeatedly invoke them”.
5. Five-Step Enterprise Landing Path
| Step | Action | Resources / platforms |
|---|---|---|
| 1. Filing self-check | Determine whether your service triggers “filing / registration” | CAC’s 538-filing list + Beijing / Jiangsu / Shandong CAC notices |
| 2. Compute inventory | Evaluate whether inference depends on overseas GPUs | Ascend / Cambricon / Hygon / Metax / Iluvatar adaptation matrix |
| 3. Scene mapping | Pick 1-2 vertical pilots within the 1000-scene target | ”AI+” Action Plan + WAIC “ten treasures” coordinate system |
| 4. Governance alignment | Data security + content labeling + user agreement | ”AI-Generated Content Labeling Method” + “whitelist” evaluation bodies |
| 5. Post-filing operation | Publish model name / filing number / generation-content label | Beijing CAC 6/2 notice mandatory publication from the 241st filing onward |
Key Terminology
| Term | Explanation |
|---|---|
| Generative AI service filing | The security assessment + filing procedure that any entity providing a generative-AI service with public-opinion or social-mobilization attributes must complete (Article 17 of the Interim Measures) |
| Generative AI application / function registration | The registration procedure for enterprises that call an already-filed model’s capability and serve it to the public |
| CANN | Compute Architecture for Neural Networks, Huawei Ascend’s heterogeneous compute architecture positioned as a CUDA counterpart |
| Cambricon NeuWare | Cambricon’s base software platform, supporting its self-developed AI chips’ runtime and toolchain |
| ”AI+” Action | The 2026 deepening implementation plan, with the “100 industry-LLM benchmarks + 1000 high-value application scenes” dual-hundred target |
| WAIC | World Artificial Intelligence Conference, held annually in Shanghai since 2018 |
| Open Challenge (揭榜挂帅) | A Chinese-style tech-task mechanism — government publishes demand lists, enterprises / research institutes bid to tackle them |
| ”Ten Treasures” coordinate system | The “ten landmark exhibits” coordinate system of WAIC, used as a reference for evaluating the relative position of head products |
FAQ (Direct Answers to High-Frequency Questions)
Q1: What scale does 538 filings represent? A: 538 completed filings + 263 completed registrations = 801 “licensed” generative-AI products nationwide. Up about 55.5% from the 346 filings in March 2025 — the steepest segment of the past 18 months.
Q2: What’s the difference between filing and registration? A: Filing applies to entities that directly provide a large-model service. Registration applies to enterprises that call an already-filed model’s capability and serve it to the public. The first regulates the model itself; the second regulates the product that uses the model.
Q3: Is domestic compute really usable now? A: DeepSeek has completed adaptation with 16 domestic AI chip makers; Ascend open-sourced CANN; Cambricon polished NeuWare. The market has moved from “policy slogan” to “buyable and deployable” — but per-card performance and software ecosystem still trail NVIDIA, so scene-by-scene evaluation is required.
Q4: What’s the direct business meaning of Xi’s WAIC 2026 speech? A: It elevates AI to a “century-level change” policy priority. Governance cadence will further standardize — compliance costs rise, but the “go out” channel opens through participation in multilateral standard-setting.
Q5: Can the AI+ Action’s 100 benchmarks + 1000 scenes actually land? A: The 1000 scenes map to the “open challenge” mechanism — local governments + leading enterprises open the scenes. It is the refined upgrade of the past “thousand-model war”. The 100 benchmarks require “human-expert level + commercial closed loop”, which is harder than a count target.
Q6: What should enterprises do right now? A: Step one: filing self-check + compute inventory (see the 5-step path in Section 5). Then pick 1-2 vertical pilots within the 1000 scenes, and avoid head-on competition in the “100 benchmarks” red ocean.
References
- CAC “14th Five-Year” Informatization Report (2025-12-25): Over 700 generative-AI products completed filing (baseline)
- Tencent Cloud Developer Community “Generative AI Large Model Filing Materials” (2026-08-18 19:04): 538 filings + 263 registrations
- Beijing CAC Notice (2026-06-02): 241 filings
- Beijing CAC Notice (2026-06-18): 242 filings
- Shandong CAC Notice (2026-08-04): 49 filings (incl. Haidai Public Opinion LLM and 4 others)
- Jiangsu CAC Notice (2026-05-22): 80 filings
- so.html5.qq.com “AI Compute Prosperity Continues to Validate” (2026-08-17): Aug 17 domestic-compute sector rally
- CLS “Domestic Compute Enters a Positive Cycle” (2026-08-18): Cambricon / Hygon Q3 earnings expectations
- Guangming Daily “Lead AI Development with Human Wisdom and International Consensus” (2026-07-19): Xi Jinping’s WAIC keynote
- CSDN “2026 Policy Released” (2026-07-24): AI+ Action 100 benchmarks + 1000 scenes