China Imposes Strict Export Controls on 23 AI Technologies with 2T Parameter and 10^26 FLOP Thresholds

For nearly four years, the geopolitical tech landscape was defined almost entirely by unilateral Western export restrictions designed to choke Beijing’s access to advanced semiconductor fabrication, lithography equipment, and frontier GPU accelerators. This week, China responded by erecting its own formidable software and algorithmic perimeter.

In a comprehensive regulatory filing issued jointly by the Ministry of Commerce (MOFCOM) and the Ministry of Science and Technology, Beijing published draft revisions under Announcement No. 48, establishing aggressive outbound export controls across 23 specific artificial intelligence technologies.

The new regime introduces hard quantitative thresholds for frontier artificial intelligence: restricting cross-border transfers of model weights, architecture IP, and training recipes for systems scaling beyond 2 trillion parameters or trained with computational budgets exceeding 10^26 total FLOPs. Furthermore, the framework imposes an outright export ban on 14 specialized technologies, including autonomous cyber-exploit agents and self-mutating malware synthesizers.

However, in a calculated move designed to preserve international market revenue for domestic tech giants, MOFCOM explicitly carved out API access, cloud SaaS deployments, and prior open-source releases from the definition of controlled technology exports. The dual-regime demonstrates Beijing’s intent to lock down sovereign algorithmic assets while maintaining commercial monetization channels abroad.

1. Anatomy of Announcement No. 48: Quantitative Perimeters

Unlike earlier, ambiguously worded trade notices, Announcement No. 48 establishes explicit technical benchmarks that define what constitutes a state-restricted “frontier foundation system.”

┌─────────────────────────────────────────────────────────────────────────────┐
│                 China MOFCOM Announcement No. 48 Regulatory Filter           │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   Domestic AI Technology Stack (Models, Agents, Weights, Cluster Code)      │
│                                │                                            │
│                                ▼                                            │
│   ┌─────────────────────────────────────────────────────────────────────┐   │
│   │                     Quantitative Trigger Criteria                   │   │
│   │  - Parameter Count: ≥ 2,000,000,000,000 (2 Trillion) Parameters   │   │
│   │  - Training Compute: ≥ 10^26 Cumulative FLOPs                      │   │
│   │  - Training Ingress: Trained on Chinese "Important Data" / CNI     │   │
│   │  - Benchmark Standing: Top-10 placement on global safety/eval lists │   │
│   └────────────────────────────┬────────────────────────────────────────┘   │
│                                │                                            │
│         ┌──────────────────────┴──────────────────────┐                     │
│         ▼                                             ▼                     │
│   ┌───────────────────────────┐                 ┌───────────────────────────┐
│   │ Prohibited Category (14)  │                 │ Restricted Category (26)  │
│   │  - Autonomous cyber       │                 │  - Frontier base weights  │
│   │    weapons & zero-day     │                 │  - Multi-agent control    │
│   │    mutation agents        │                 │    protocols              │
│   │  - Military target auto-  │                 │  - Heterogeneous cluster  │
│   │    nomous tracking        │                 │    compilers & EDA tools  │
│   │  [TOTAL BAN ON EXPORT]    │                 │  [GOVERNMENT LICENSE REQ] │
│   └───────────────────────────┘                 └───────────────────────────┘
│                                │                                            │
│                                ▼                                            │
│   ┌─────────────────────────────────────────────────────────────────────┐   │
│   │                     Strategic Exemptions (Safe Harbors)             │   │
│   │  ✓ Public REST/gRPC APIs, Web SaaS, and mobile apps (Non-Export)    │   │
│   │  ✓ Pre-existing open-source weights published prior to Oct 16, 2026 │   │
│   └─────────────────────────────────────────────────────────────────────┘   │
└─────────────────────────────────────────────────────────────────────────────┘

The 10^26 FLOP benchmark is particularly telling: it mirrors the reporting threshold established in the United States under federal AI governance orders, signaling that Beijing is synchronizing its export perimeter with internationally recognized definitions of catastrophic-risk and frontier compute capabilities.

2. Prohibited vs. Restricted: Where the Line Is Drawn

The regulatory package spans 40 total emerging technologies—encompassing advanced solar photovoltaics and biopharmaceutical synthesis—with 23 items directly governing artificial intelligence infrastructure:

Technology Classification	Regulatory Status	Scope & Covered Capabilities	Strategic Objective
Autonomous Cyber Exploit Agents	Prohibited (Total Ban)	Autonomous models capable of finding zero-day vulnerabilities, mutating payloads, or executing automated penetration testing without human-in-the-loop controls	Halts proliferation of offensive AI capabilities; prevents domestic talent from licensing exploits abroad
Frontier Foundation Weights (≥ 2T Params)	Restricted (Licensed)	Full checkpoint model weights, FP8/FP4 quantized weight sets, and distillation seeds for models above 2T parameters	Prevents foreign acquisition or hostile corporate takeovers of leading Chinese foundation models
Heterogeneous GPU Cluster Schedulers	Restricted (Licensed)	Software fabrics, communication libraries, and topology-aware compilers optimized for clustering 50,000+ domestic accelerators	Guards national breakthroughs in clustering Ascend, Kunlun, and Moore Threads chips
Cloud Inference APIs & SaaS Platforms	Exempt (Safe Harbor)	Querying models via standard HTTPS/WebSocket API endpoints, hosted platforms, and third-party developer sandboxes	Protects billions in recurring global cloud revenue for Alibaba Cloud, Tencent, and Baidu
Pre-Existing Open-Source Weights	Exempt (Grandfathered)	Open-weights models (e.g., Qwen, DeepSeek, Yi) published with open licenses prior to October 16, 2026	Prevents legal chaos for international developers already utilizing Chinese open weights

Crucially, while the formal public comment period extends until November 15, 2026, MOFCOM enacted an emergency interim pre-export registration requirement that took effect immediately on October 16. Any pending cross-border licensing agreement, foreign joint venture, or transfer of algorithmic ownership must now be registered and cleared through provincial commerce departments before completion.

3. The Commercial Strategy: Locking the Vault While Renting the Key

Western market observers had feared that aggressive export controls might force Chinese AI innovators to retreat behind an impenetrable digital curtain, severing open-weights ecosystems like Qwen and DeepSeek from the global developer community.

Instead, Announcement No. 48 reveals a nuanced, pragmatic economic strategy:

You Cannot Buy the Factory, But You Can Buy the Product: By exempting API queries and hosted enterprise services, Beijing ensures that multinational enterprises in Southeast Asia, the Middle East, Europe, and Latin America can continue building production systems atop Chinese AI platforms without triggering export licensing reviews.
Preempting Silicon Valley Acqui-Hires: In 2024 and 2025, several prominent European and Asian AI startups saw their intellectual property effectively absorbed by American hyperscalers through talent-acquisition and synthetic licensing deals. The MOFCOM rules make it legally impossible for a foreign entity to acquire the core IP or weights of a leading Chinese frontier AI lab without State Council authorization.
Protecting Domestic Hardware Workarounds: As Chinese tech firms deployed innovative software workarounds—such as specialized distributed schedulers that stitch together heterogeneous clusters of non-standard silicon—Beijing recognized that software compiler ingenuity had become China’s most valuable defensive asset against U.S. chip embargoes.

┌─────────────────────────────────────────────────────────────────────────────┐
│                 The New Bipolar AI Trade Architecture                       │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   United States & Allied Perimeter:                                         │
│   ├── Hardware & Silicon Controls (NVIDIA, ASML, HBM memory export bans)   │
│   └── Regulatory scrutiny on agent containment (FTC Section 5 Subpoenas)    │
│                                                                             │
│   China MOFCOM Perimeter (Announcement No. 48):                             │
│   ├── Software & Algorithmic Controls (2T Parameters, 10^26 FLOP limits)    │
│   ├── Prohibitions on offensive cyber agents & cluster scheduling code      │
│   └── Commercial Safe Harbor: Cloud APIs & SaaS permitted for global export │
└─────────────────────────────────────────────────────────────────────────────┘

The export restrictions arrive during an increasingly tense international climate for frontier models, as Washington intensifies oversight with FTC subpoenas directed at OpenAI and Anthropic and Wall Street reassesses the hundreds of billions committed to neocloud data centers.

4. Developer and Enterprise Impact: What Global Teams Need to Know

For international engineering organizations and developer teams building on Chinese models, Announcement No. 48 provides both stability and clear operational boundaries:

Existing API Integrations Remain Uninterrupted: Development teams consuming APIs from Alibaba Cloud, DeepSeek, Tencent, or Baidu are explicitly protected under the SaaS exemption. No export licenses or regulatory disclosures are required for standard API consumption.
Self-Hosted Open Weights Require Compliance Checks: For companies looking to download and self-host future 2T+ parameter frontier models released by Chinese labs, enterprise procurement teams will need to verify whether the publishing lab secured an export authorization from MOFCOM prior to distributing weights internationally.
Cross-Border M&A and Licensing Frozen: Any venture capital deal, merger, or licensing agreement involving transfer of source code, model checkpoints, or training logs between Chinese AI startups and foreign corporations is subject to immediate mandatory pre-export registration.

Frequently Asked Questions

What are the key thresholds in China’s new AI export control rules?

Announcement No. 48 establishes that foundation models with 2 trillion parameters or more, models trained with cumulative compute of 10^26 FLOPs or higher, or models trained on designated Chinese “important data” require strict government export licenses.

Are cloud APIs and ChatGPT-style services affected by the export ban?

No. MOFCOM explicitly established a safe harbor stating that providing access to AI capabilities via APIs, SaaS web platforms, and mobile apps is not considered an outbound technology export.

Which AI technologies are completely banned from export?

Beijing imposed a total export prohibition on 14 technologies, most notably autonomous AI agents designed for cybersecurity exploitation, zero-day vulnerability scanning, self-mutating malicious payloads, and automated military tracking systems.

Does this impact existing open-source models like Qwen or DeepSeek?

Pre-existing open-source weights and models published under open academic licenses prior to October 16, 2026, are exempt from the new export restrictions. Future releases exceeding the 2T parameter or 10^26 FLOP thresholds will require regulatory review before international weight distribution.