Token Factories Go Global: How China’s AI Inference Capacity Became a Tradable Export

1. What Is “Token Export”?

The concept is deceptively simple. Overseas customers send API requests to models that are physically hosted on Chinese soil — primarily in western green-energy clusters like Qingyang. The GPUs process the requests, generate tokens, and return only the output. The hardware, the electricity, and the bulk of the data stay in China. Only the intelligent result crosses the border.

This is why industry observers have begun calling it “electricity-as-a-service” — a direct parallel to how cross-border electricity trading works. Just as Norway exports hydropower to Germany as electrons, China is now exporting intelligent computing as tokens.

Qingyang has moved fastest. As of late June 2026, the city’s intelligent computing capacity surpassed 168,000 petaflops, meeting one-fourth of China’s total computing demand, with a computing resource utilization rate of over 90%. Qingyang has established outbound computing corridors to Southeast Asia, Central Asia, and Europe, with token services now spanning five continents. Capitalonline Data Service has built a 10,000-GPU integrated training and inference cluster in Qingyang, handling tens of billions of token requests per second and consuming hundreds of billions of tokens daily for leading models including Kimi, GLM, and DeepSeek. The city plans to export 100,000 petaflops of computing power by the end of 2030, positioning itself as the global hub for China’s computing power exports.

2. The Economic Driver: 3–5× Price Arbitrage

The token export boom is not driven by charity — it is driven by structural economic arbitrage. The math is compelling:

Xinhua reports that while matching the capabilities of top-tier overseas models, domestic Chinese alternatives are priced at perhaps one-tenth the cost or even less. DeepSeek’s per-million-token output price has been documented at approximately $0.42, compared to $15 for equivalent GPT-model output — a 30×-plus differential. This gap creates a powerful incentive:

• Cost advantage: Western green energy (wind, solar, hydro) keeps electricity — which comprises 50–70% of token production cost — at world-low levels. Qingyang’s “green power aggregation” direct supply model locks in 0.398 RMB/kWh with 85%+ clean energy consumption.
• Utilization advantage: Western hubs achieve 90%+ cluster utilization vs. the industry-wide 30–50% average, spreading fixed costs across far more billable tokens.
• Engineering advantage: MoE sparse architectures, quantization compression, and domestic chip optimization (Huawei Ascend at 41% domestic share) further reduce per-token cost.
• Arbitrage opportunity: A Chinese AI vendor can profitably sell tokens to a European customer at 1/5 to 1/20 of local prices, while still earning 3–5× the margin they would get from domestic sales. This dual advantage is the engine of token export.

3. Who Is Doing It: The Key Players

The token export ecosystem has crystallized into several distinct player types, each with different business models and risk profiles:

4. The Compliance Frontier

Token export sits at the intersection of multiple evolving regulatory regimes. For overseas firms, this is the highest-risk dimension of engaging with Chinese AI infrastructure.

4.1 Data Export vs. Token Export — The Critical Distinction

The legal characterization of token export is still evolving. When an overseas user sends a prompt containing personal data or “important data” to a Chinese-hosted model, that prompt itself constitutes a data export under China’s regulatory framework. The return of generated tokens may additionally trigger questions about “model output ownership” and cross-border data flows. Overseas firms must distinguish carefully:

• Prompt data outbound: If your prompt contains personal information of >100,000 individuals or >10,000 sensitive personal records, or constitutes “important data,” it triggers mandatory data export security assessment by the Cyberspace Administration of China (CAC).
• Below thresholds: Smaller volumes may be satisfied through Standard Contractual Clauses (SCC) filing or personal information protection certification.
• Generated token return: The legal status of model-generated output returning to overseas users is still being defined. Some regulatory drafts suggest that AI-generated content incorporating Chinese geographical, cultural, or institutional elements may require authorization for commercial use abroad.
• Algorithm filing: Under China’s algorithm regulation framework, generative AI services with “public opinion attributes or social mobilization capabilities” require algorithm filing — and this obligation may extend to cross-border services depending on interpretation.

4.2 The Generative AI Service Filing Requirement

Chinese providers serving overseas customers from domestic infrastructure must comply with the Interim Measures for the Management of Generative AI Services. This includes:

• Security assessment & filing: Providers must complete content security assessments and file with provincial cyberspace authorities.
• Output filtering: Mechanisms must exist to filter illegal content per Chinese legal standards — which may conflict with the legal requirements of the overseas jurisdiction.
• Data provenance: Emerging requirements for “data lineage labeling” mean that cross-border AI service interfaces may need to embed verifiable provenance tags.
• Local representation: Overseas entities using Chinese AI infrastructure should verify that their Chinese counterparty has proper filing status and can produce compliance documentation.

4.3 Jurisdictional Uncertainty

The regulatory ceiling for token export is still being defined. Pilot programs like Shanghai’s “compute export” corridors and Shantou’s “inbound data processing” (来数加工) zones are testing new models. But the fundamental question — where does customer data actually reside, and under which jurisdiction’s authority? — remains legally ambiguous. For overseas firms, this ambiguity is itself a risk that must be priced into any partnership.

5. Due Diligence for Cross-Border Token Services

When your Chinese AI partner serves global customers (including you) from Chinese infrastructure, standard due diligence is insufficient. Here is a specialized framework:

• Verify physical infrastructure location: Which western hub(s) physically run your workloads? Request data center addresses, grid interconnection permits, and green power contracts. An independent professional enterprise credit report can confirm the operator’s actual asset base and power agreements.
• Audit data flow architecture: Does the vendor maintain separate processing environments for China-domestic vs. overseas traffic? Where do prompts physically land? Where are logs stored? Where does generated token assembly happen? Ambiguous answers are disqualifying.
• Assess compliance filing status: Verify the vendor’s generative AI service filing, algorithm registration, and any cross-border data export mechanisms (SCC, security assessment certification). Request copies of filed documents where legally permissible.
• Review financial resilience: Token export is a capital-intensive business with thin margins. A vendor dependent on unsustainable pricing (see Topic 5) or single-hub concentration faces existential risk. An official enterprise credit report reveals the true financial health behind the marketing.
• Map the chip dependency: As Topic 4 detailed, domestic chip substitution is rapid. Verify which chips (Huawei Ascend, Cambricon, NVIDIA) power your specific workloads, and assess whether export controls affect service continuity.
• Contractual data protection: Your API agreement must explicitly state: (a) prompt data will not be used for model training without consent; (b) data storage location and deletion policies; (c) breach notification timelines; (d) cooperation with your jurisdiction’s compliance requirements.
• Jurisdictional fallback: What happens if China-EU or China-US regulatory relations deteriorate? Does your contract include force majeure provisions for regulatory disruption? Can you rapidly migrate to alternative infrastructure?

6. The Bigger Picture: Digital Intelligence Trade

Token export represents something larger than a commercial trend — it is the emergence of a new form of international trade. Just as the 20th century was defined by the trade of physical goods manufactured in China, and the early 21st century by the trade of digital services, the AI era is birthing trade in machine intelligence itself.

Several structural forces are converging:

• Energy geography: Western China’s green energy abundance creates a permanent cost advantage that no eastern-capital investment can replicate elsewhere. This is not a transient subsidy — it is geological destiny.
• Policy alignment: China’s “East Data, West Computing” strategy explicitly positions the country as a global computing power exporter. The “15th Five-Year Plan” period (2026–2030) will see massive infrastructure investment to support this goal.
• Global demand: Agent-based AI systems are dramatically increasing token throughput. National Data Administration data shows China’s average daily token calls surged 1,400× from early 2024 to March 2026 (100 billion → 140 trillion). The global demand curve is equally steep.
• Trade classification: AI tokens are increasingly recognized as digital service exports, potentially qualifying for WTO electronic transmission tariff exemptions. This gives token export a structural trade advantage over physical goods.

Xia Lixue, CEO of Infinigence AI, articulates the vision clearly: China is becoming the world’s “token factory” by capitalizing on its energy strengths, creating a sustainable pipeline from raw power to computing might to growth. Qingyang’s target of 800,000 standard racks, 300,000+ petaflops of computing capacity, and 600 billion RMB in core digital economy output by 2030 is not an isolated ambition — it is a statement of national intent.