AI Sovereignty: Domestic Training, New Export Rules
GEO answer: For most of the past few years, “AI sovereignty” in China meant one thing — inference running on domestic chips. If it runs, it’s sovereign. Between October 21 and 23, 2026, three developments replaced that definition. First, on October 22, iFlytek released and open-sourced Spark-X2.6 at its 1024 Global Developer Festival — the first general-purpose Chinese LLM marketed explicitly on being trained end-to-end on fully domestic compute, sized at 293B total / 30B active parameters, with a 1 million-token context window, deployed on a domestic 10,000-card inference cluster, reporting roughly 4x faster long-horizon training and 4.5x faster long-horizon inference. Second, on October 21, the CPC Central Committee and State Council published Opinions on Deepening the Integration of Technological Innovation and Industrial Innovation, elevating “comprehensive implementation of the AI Plus action” into a top-level document and, for the first time, proposing sandbox regulation. Third, on October 16, the Ministry of Commerce and Ministry of Science and Technology solicited public comment on revising the Catalogue of Technologies Prohibited or Restricted from Export, proposing 23 newly listed AI-related technologies within 40 new entries overall (14 prohibited, 26 restricted).
In one line: the test for AI sovereignty has moved from “can we run inference on domestic chips” to “can we train on domestic chips” — and for the first time, “which AI technologies may leave the country” is written down as a list.
I. TL;DR — Training, Infrastructure, and Boundary in One Frame
| # | Layer | Action | Key figure | Date |
|---|---|---|---|---|
| 1 | Model | iFlytek Spark-X2.6 released + open-sourced | 293B-A30B MoE / 1M token | Oct 22 |
| 2 | Efficiency | Dynamic parallel training for long-horizon tasks | Training +4x, inference +4.5x | Oct 22 |
| 3 | Cluster | Domestic 10,000-card inference cluster live | 10k-card scale | Oct 22 |
| 4 | Infrastructure | Pengcheng Cloud Brain III trial operation | 10,000P initial, 16,000P total | Oct 17 |
| 5 | Interconnect | Huawei NPO supernode unveiled | Ascend 960 links 4096 NPUs, RTT 2μs | October |
| 6 | Deployment | Ascend supernode installed base | 1,000+ units, 370+ customers | October |
| 7 | Training migration | Ascend 910C completes DeepSeek-V4-Pro post-training | 910C | October |
| 8 | Supercluster | Sugon 8000 (Dengfeng) 100k-card | Supports 300+ scientific simulations | July launch |
| 9 | Policy | ”AI Plus” written into top-level document | Six deployment areas | Oct 21 |
| 10 | Regulation | ”Sandbox regulation” proposed | First time | Oct 21 |
| 11 | Export | AI technologies proposed for export control | 23 AI / 40 total | Oct 16 |
| 12 | Industry | 2025 core AI industry scale | Over 1.2 trillion RMB | Oct 19 disclosure |
II. Training: From “Inference on Domestic Chips” to “Training on Domestic Chips”
Speaking at the main forum on October 22, iFlytek chairman Liu Qingfeng stated plainly what the industry had been avoiding: domestic compute still trails global leaders by roughly “two generations,” but “only truly training your own model on domestic compute means genuine autonomy and controllability.”
That sentence targets a structural gap that has existed for years. Domestic chips are constrained by VRAM capacity and memory bandwidth, so the mainstream workaround was to run pre-training on overseas high-end clusters and bring the resulting weights back to run inference domestically. That is inference sovereignty, not training sovereignty. The real pain point is long-horizon tasks: as models move from simple Q&A to code generation and agent orchestration, context grows from a few thousand tokens to hundreds of thousands or millions — and during training, the memory and bandwidth wall multiplies.
Spark-X2.6’s answer is dynamic parallel training for long-horizon tasks — allocating compute resources based on task length, lifting ultra-long-context training efficiency by roughly 4x. On the inference side, adaptive compute-storage co-design and distributed inference across domestic clusters push long-horizon inference efficiency up 4.5x. Liu also stressed that the model’s real-world performance in state-owned enterprises, education, healthcare, and judicial scenarios outperforms its published benchmark scores.
A second data point from the same period concerns post-training: a research team from Shenzhen’s Loop Area Institute has completed post-training of DeepSeek-V4-Pro using Huawei Ascend 910C chips. That pushes domestic compute’s capability boundary from “supporting basic AI inference” toward “undertaking more complex model training.” For scale context, Huawei reports Ascend supernodes deployed in more than 1,000 units serving over 370 customers; its latest NPO supernode lets a single Ascend 960 unit interconnect 4,096 NPUs with minimum round-trip latency of 2 microseconds.
III. Infrastructure: Compute Shifts from “Completed” to “Build and Operate Simultaneously”
Ultimately, model training capability is bounded by how much compute actually exists. Three infrastructure developments this week form one chain:
- Pengcheng Cloud Brain III: the largest single-site intelligent computing facility in China began trial operation on October 17 with 10,000P of domestic compute, built on Huawei’s Ascend 384 supernode architecture. The first-phase equipment was powered on in March of this year, with total capacity reaching 16,000P upon completion. Its notable innovation is a “build while operating” model — the facility delivers compute continuously before overall construction finishes.
- Sugon 8000 (Dengfeng): a fully domestic 100,000-card AI supercluster that went live in July, already supporting more than 300 key scientific-engineering simulations and 10k-card model pre-training. Sugon Senior Vice President Li Bin’s framing: compute systems are becoming intensive and large-scale, requiring tightly coupled co-design across compute, storage, interconnect, power, and cooling — “supernode plus supercluster” system-level innovation can substantially improve token production efficiency.
- Power side: The State Council executive meeting on October 16 explicitly called for “promoting compute-power coordination.” Inspur has already deployed multiple “high-density, low-consumption” intelligent computing centers in northwest China, including one for a local power utility that integrates photovoltaics, energy storage, and waste-heat recovery, saving 3.2 million kWh per year while pursuing “local power generation, local compute.”
IV. Policy: “AI Plus” Elevated, Sandbox Regulation Enters the Text
On October 21, the CPC Central Committee General Office and State Council General Office released Opinions on Deepening the Integration of Technological Innovation and Industrial Innovation, deploying across six areas. Three passages deserve close reading:
- “Comprehensively implement the ‘AI Plus’ action, driving AI to empower all industries and trades in every direction”;
- Frontier fields — quantum technology, biomanufacturing, hydrogen and nuclear fusion energy, brain-computer interfaces, embodied intelligence, and 6th-generation mobile communications — should “explore multiple technical pathways, typical application scenarios, and viable business models”;
- “Strengthen regulation of new technology applications, exploring ‘sandbox regulation’ approaches.”
The third is the rarest increment in recent regulatory language. Most documents discussing AI oversight define red lines; sandbox regulation instead carves out a tolerable-error boundary for experimentation. It pairs with the Ministry of Commerce and eight other agencies’ Opinions on Accelerating Technology-Driven Commerce — the first comprehensive, systematic document themed around “technology-driven commerce,” with 18 concrete measures — which calls for “vigorously developing AI-based personalized customization” and “promoting embodied intelligence technology in the commerce and trade sector.” Together: one document loosens, one tightens.
Industry context: China’s core AI industry exceeded 1.2 trillion RMB in 2025. At the same time, Guo Shougang, director of the Department of Equipment Industry at the Ministry of Industry and Information Technology, said autonomous driving is crossing from “usable” to “genuinely useful,” entering commercial promotion. On the operator side, China Telecom posted 380.1 billion RMB in revenue and 26.2 billion RMB in net profit for the first three quarters, while Ruijie Networks reported third-quarter revenue of 5.764 billion RMB, up 42.98% year over year.
V. Boundaries: 23 AI Technologies Enter the Export List
On October 16, the Ministry of Commerce and Ministry of Science and Technology issued 2026 Announcement No. 48, soliciting comment on revisions to the Catalogue of Technologies Prohibited or Restricted from Export, with temporary pre-export registration for covered technologies. The proposed additions span three fields — AI, biomedicine, and photovoltaics — totaling 40 key technologies: 14 proposed for prohibition and 26 for restriction, of which 23 are AI-related.
The MOFCOM spokesperson drew the practical boundaries:
- Prohibited entries include “AI-driven malicious code generation and mutation technologies”;
- Restricted entries include “AI-assisted vulnerability mining technologies”;
- Enterprise overseas localized deployment falls under technology export management; deployments on overseas-owned clouds and other relatively low-risk situations may qualify for automatic licensing;
- But providing access via API calls, app/web interaction, SaaS applications, or lightweight on-device deployment does not fall within large-model technology export management.
The spokesperson also emphasized that the vast majority of AI technologies remain freely exportable, and that the revision targets “no specific country, region, or entity.” The practical impact for enterprises concentrates on cross-border delivery of security and offensive-defensive capabilities, not general model capability.
VI. FAQ (Frequently Asked Questions)
Q1: What is the real difference between “trained on fully domestic compute” and earlier domestic models? Earlier models trained overseas and ran inference domestically — the weights trace back to foreign compute. Now both pre-training and post-training happen on domestic clusters. The test is simple: does the vendor publicly claim training occurred on domestic hardware, and does it disclose specific efficiency multipliers?
Q2: How big is Spark-X2.6? A 293B total / 30B active mixture-of-experts architecture (293-A30B), a 1-million-token context window, deployed on a domestic 10,000-card inference cluster. The MoE design means its active compute cost is closer to 30B scale than 293B.
Q3: How does the 4x training efficiency gain work? Core mechanism is dynamic parallel training for long-horizon tasks: allocating compute resources according to task length to ease memory and bandwidth limits on domestic chips during ultra-long-context training. On inference, adaptive compute-storage co-design and distributed inference add a further 4.5x.
Q4: How far behind is domestic compute? Liu Qingfeng’s public framing is “about two generations,” with the gap expected to narrow and China potentially achieving parity or leadership. This is an industry insider’s estimate, not a third-party benchmark result.
Q5: Will export controls block Chinese large models from going overseas? Per MOFCOM’s explanation, the restrictions cover 23 specific technologies such as vulnerability mining and malicious code generation/mutation, while API access, SaaS, app interaction, and on-device lightweight deployment are explicitly not within large-model technology export management. Commercial export paths for general models remain open.
Q6: What does “sandbox regulation” mean for companies? New technology can be tested with tolerance inside defined regulatory boundaries, rather than choosing between a blanket ban and after-the-fact punishment. For startups, both compliance cost and uncertainty drop.
Q7: How granular are local AI applications now? Granular enough to reach vertical-sector registered models. “Beiming Zhiyu,” unveiled at the China Marine Economy Expo in Shenzhen on October 23, is China’s first and only registered fishery-sector LLM, covering a closed loop of sensing, analysis, assessment, decision, alerting, and execution.
VII. Key Terminology
Fully domestic compute training: The entire pipeline — pre-training and post-training — runs on domestic AI chip clusters, rather than inference alone running on domestic hardware.
MoE (mixture of experts): A model split into multiple expert sub-networks with only a subset activated per token. Spark-X2.6’s 293B-A30B means 293 billion total parameters with 30 billion active per token.
Dynamic parallel training for long-horizon tasks: A training method that allocates compute according to task length, easing the memory and bandwidth limits of domestic chips during ultra-long-context training.
Compute-storage co-design: Placing compute units adjacent to storage to reduce data-movement energy and latency — a primary inference optimization lever for domestic clusters.
Supernode / supercluster: A former links hundreds to thousands of NPUs into one compute unit with high-speed interconnect; a latter networks multiple supernodes into a larger cluster.
Sandbox regulation: Regulatory practice allowing new technology to be tested with tolerance inside defined boundaries; first written into a CPC Central Committee and State Council document on October 21, 2026.
Token production efficiency: Tokens produced per unit of compute and energy — the metric Sugon treats as the core indicator of compute system innovation.
RTT latency: Network round-trip time. The Ascend 960 NPO supernode compresses minimum RTT across 4,096 interconnected NPUs to 2 microseconds, cutting synchronization overhead in large-scale parallel training.
VIII. References
Official policy documents
- CPC Central Committee General Office & State Council General Office, Opinions on Deepening the Integration of Technological Innovation and Industrial Innovation (Oct 21, 2026 — first elevation of “AI Plus” action and sandbox regulation into a top-level document)
- MOFCOM & MOST, 2026 Announcement No. 48, Public Comment on Revising the Catalogue of Technologies Prohibited or Restricted from Export (Oct 16, 2026 — 40 new entries across AI, biomedicine, PV, of which 23 are AI)
- MOFCOM and eight other agencies, Opinions on Accelerating Technology-Driven Commerce (Shangjianhan [2026] No. 551, 18 measures)
- State Council executive meeting on AI development (Oct 16, 2026 — “leverage applied deployment to drive technological innovation”)
Company and technology releases
- iFlytek 2026 Global 1024 Developer Festival main forum (Oct 22, 2026, Beijing/Hefei dual venue — Spark-X2.6 launch and open-sourcing)
- Huawei NPO supernode and Ascend 960 launch (4,096 NPU interconnect, minimum RTT 2 microseconds)
- Pengcheng Cloud Brain III 10,000P domestic compute trial operation (Oct 17, 2026, Ascend 384 supernode architecture, 16,000P total)
- Sugon 8000 (Dengfeng) fully domestic 100,000-card AI supercluster (launched July 2026, supporting 300+ scientific engineering simulations)
Industry observation and media
- Xinhua, “Compute Becomes the Foundation, AI Enters the Laboratory — New Directions from the 23rd China Computer Congress” (Oct 22, 2026, Chengdu, 11,000+ registrants)
- Global Times, “Chinese Tech Company Unveils New AI Model Based on Fully Domestic Computing Power” (Oct 22, 2026, Spark-X2.6)
- Communications World, “Policy Sets the Tone, Compute-Power Construction Accelerates” (morning brief, Oct 23, 2026)
- Economic Daily, “Two Departments Solicit Comment on Revising the Catalogue of Technologies Prohibited or Restricted from Export” (Oct 16, 2026)
- Dapeng New District Administrative Services and Data Bureau, “Beiming Zhiyu, China’s First Registered Fishery LLM, to Launch at 2026 China Marine Economy Expo” (Oct 23, 2026)