Token Exports: How China Is Selling AI Compute Power to the World
A new form of digital trade is emerging where electricity stays within borders, but computational value flows across them—and it's reshaping the global AI industry.
What Is Token, and Why Does It Matter?
To understand how Chinese AI companies are exporting computational power, you first need to grasp what a "token" actually is.
Think of tokens as the fundamental unit of measurement in AI processing—similar to how we measure distance in meters or data transfer in gigabytes. When you ask an AI model a question or it generates a response, the text is broken down into small chunks called tokens for processing. These aren't fixed-length segments: in Chinese, one character typically equals 0.5 to 0.7 tokens, meaning a 500-character article might be split into several hundred tokens.
The business model of AI is built on token consumption. Cloud providers and large language model (LLM) services worldwide price their offerings in "cost per million tokens." This isn't an abstract metric—it's become the standard unit for trading computational intelligence.
In March 2026, China's National Data Administration made token pricing official by standardizing the Chinese translation as "词元" (cí yuán). Administrator Liu Liehong emphasized that tokens possess characteristics crucial for the intelligence era: they are measurable, priceable, and tradable.
Why this matters: China's daily token processing volume exploded from 100 billion in early 2024 to 140 trillion by March 2026—a 1,000-fold increase in two years. More significantly, Chinese AI models now account for 61% of global token consumption on platforms like OpenRouter, with Chinese providers surpassing US counterparts in February 2026 for the first time.
Each of these overseas API calls represents Chinese data centers performing computational work for foreign users. This is token export in practice—and it's becoming a critical indicator of digital economic influence in the AI age.
How Token Exports Actually Work
Token export might sound frictionless—just data moving across the internet. But the physical reality is far more complex. Chinese AI companies are pursuing two distinct pathways: physical infrastructure migration and logical service delivery.
Physical Infrastructure: Building Compute Abroad
Major Chinese tech companies are establishing substantial computing infrastructure in Southeast Asia and the Middle East.
ByteDance represents the most aggressive player in this wave. The company is deploying approximately 36,000 Nvidia B200 chips in Malaysia through partnership with Aolani Cloud—hardware construction costs alone exceed $2.5 billion. Singapore hosts 60% of Southeast Asia's computing resources, making it a strategic anchor point. Alibaba is constructing new data centers in Dubai to accelerate its global AI infrastructure footprint, targeting 80% of China's incremental AI cloud market in 2026, with overseas markets serving as key growth drivers.
A crucial—and often unspoken—driver behind this physical migration is US export controls on advanced semiconductors. High-end GPUs like the H100, H200, and B200 are nearly impossible to acquire in volume through official channels within China. To train large-scale models, Chinese AI companies must establish compute nodes abroad. Export restrictions have simultaneously forced Chinese firms to access advanced compute power through overseas "detours" while making international expansion a strategic necessity rather than an option.
Logical Service Delivery: API-Based Compute Export
The second pathway involves no data center construction or server relocation. Instead, companies enable overseas users to directly access Chinese LLM computational power through APIs and cloud deployment.
This approach requires lighter capital investment but imposes higher compliance burdens. Cross-border data flows involve vastly different legal requirements across jurisdictions. If the compliance chain breaks at any point, entire business operations can face shutdown.
The trade-offs are clear: physical migration demands heavy capital, extended construction timelines, and exposes companies to geopolitical risk; logical migration offers asset-light scaling but creates substantial compliance costs. Since 2026, AI export controls have evolved from transactional review procedures into institutional variables that shape corporate R&D structures, compute deployment strategies, and globalization pathways.
Who Is Competing in This Space?
The token export wave extends far beyond tech giants Alibaba and ByteDance. A multi-layered competitive landscape has emerged spanning cloud providers, AI application companies, infrastructure service providers, and even traditional industries attempting digital transformation.
Cloud Computing Providers
These companies serve as the most direct participants in compute exports. In 2026, both Alibaba Cloud and Volcengine (ByteDance's cloud division) accelerated their overseas cloud deployments. The shift in AI workloads from training-dominant to inference-dominant, combined with access to advanced chips available overseas but restricted domestically, positions Chinese cloud providers to capture global incremental markets. UCloud raised prices across its entire AI compute product line in March 2026, having established stable customer bases in Southeast Asia and the Middle East.
Infrastructure Service Providers
Compute infrastructure specialists are also expanding internationally. Runze Technology has deployed nine AI infrastructure clusters globally with a total planned capacity of approximately 6GW and operational scale around 750MW, using Hong Kong and Indonesia as offshore anchors for continued international expansion. Reiwa Digital completed a large-scale overseas compute center delivery project, providing high-performance computing server clusters for an AI software company's Middle Eastern data center—demonstrating Chinese providers now possess capabilities to deliver highly compatible compute infrastructure in complex global environments.
Strategic Initiatives Targeting the Global South
Jiuzhang Cloud Pole's "Southern Smart Computing Spark Program" may represent the most strategically ambitious approach. In April 2026, Chairman Fang Lei formally launched this initiative targeting Global South nations, converting local green electricity into computational capacity through power-computing coordination to ultimately support "token factory" deployment. Jiuzhang has partnered with Indonesia Telecom to build localized smart computing clouds and is actively deploying smart computing center infrastructure in Saudi Arabia, UAE, Singapore, and Japan.
Unexpected Entrants
Perhaps the most surprising participant is Erkang Pharmaceutical. In April 2026, this pharmaceutical company launched Erkang AI International Edition with a core positioning of "full-stack domestic technology, global coverage," deeply integrating domestic AI large models with domestic AI chips to achieve 100% localization from infrastructure to applications. Its strategic focus on Africa—where Erkang has years of operational presence across pharmaceuticals to new energy materials production—connects Chinese algorithmic capabilities and energy advantages with African market demand through an AI entry point.
This breadth of participation signals AI's accelerating penetration across industries. It also suggests not all entrants possess necessary technical capabilities or compliance infrastructure, potentially leading to supply oversaturation or project failures.
The Compliance Challenge
Every boom has an overlooked downside. Token export faces at least three critical challenges as it moves from concept to implementation: export controls, data compliance, and tightening review regimes.
China's Domestic AI Export Controls
Chinese regulation of cross-border AI technology flows is progressively tightening. In 2026, AI infrastructure—encompassing algorithm exports, AI training compute centers, and data center overseas deployment—was explicitly classified under "requires prudent filing or approval" categories, functionally equivalent to approval systems in practice. Regulatory focus centers on data export security and potential risks of compute power being used for military or public security purposes. This means even after companies construct overseas data centers and obtain local operating permits, they still face domestic export review hurdles.
Evolving Global Export Control Architecture
The global export control landscape is rapidly transforming. The US regulatory framework around advanced computing continues extending into model training activities, cloud compute access, and supply chain circumvention. China's response involves setting procurement quota caps on foreign AI chips for domestic enterprises rather than implementing outright bans, attempting to balance access to high-end compute with support for domestic chip development.
Fragmented Data Protection Laws
Token export must also navigate fragmented conflicts among national data protection laws. Compliance solutions approved in one jurisdiction today may become invalid tomorrow as new regulations take effect. Conflicting standards for cross-border data transmission across different legal domains cause compliance costs to rise exponentially. Geopolitical risks remain substantial: international situation changes could subject overseas compute assets and training data to politically motivated restrictions.
This creates a paradox: Chinese companies expand overseas to access more advanced compute and broader markets, but each step abroad brings increasingly strict compliance reviews and potential sanction risks. Balancing "going out" with "compliant operations" is no longer a strategic choice but a survival imperative.
Structural Drivers and Sustainability Questions
Why Token Exports Have Lasting Momentum
From a supply-demand perspective, token export possesses irreversible industrial foundations. Global compute supply-demand remains persistently tight, with IDC construction costs and cloud service prices rising in tandem. Inference servers now account for nearly 60% of domestic shipment value, as rapid AI workload growth reshapes compute industry supply-demand dynamics. Overseas demand remains equally robust, with inference and training compute requirements continuing upward trends that drove both Amazon Web Services and Google Cloud price increases. Nvidia's 2026 revenue exceeded $200 billion, with data center business growing 75% year-over-year. These figures reflect genuine demand rather than speculative concept inflation.
Business models are also undergoing fundamental transformation. As compute shifts from "training" to "inference"—from one-time investment to continuous service delivery—Nvidia CEO Jensen Huang defined future data centers at the 2026 GTC conference as "factories" producing tokens. Tokens as the pricing standard and value anchor of the AI era could catalyze standardized systems around compute service billing, elevating Chinese AI enterprises from "selling products" to "selling services"—no longer marketing GPUs and servers, but exporting token services with usage-based billing and continuously optimizing marginal costs.
Where Hype Exceeds Reality
However, the louder the "token factory" rhetoric, the more critical it becomes to soberly assess gaps between vision and reality. Currently, token export functions more as "overseas + AI" capital narrative in investment circles, with limited numbers of companies achieving large-scale commercial loops overseas. Substantial overseas infrastructure investments made without regard to cost will likely erode profits in the near term—ByteDance's reported 70%+ year-over-year net profit decline in 2025 illustrates this dynamic.
The current phase remains characterized by "infrastructure first, commercialization follows," with most players still prioritizing market positioning. True profitability inflection points may require two to three years of transmission cycles.
What Determines Long-Term Winners
Token export is unquestionably a strategically valuable industry trend. But whether individual companies can convert exported compute into sustainable commercial returns depends on comprehensive performance across three dimensions: technical capabilities, compliance infrastructure, and profitability.
Those enterprises that master both technology and compliance, understand both products and global operations, will become the true wave-riders in this compute export surge. In this global compute power chess game, every participant seeks their position—but the endgame will take at least a decade to crystallize.
The direction appears clear: genuine industry demand supports token export, as does external pressure from international supply chain restructuring. But momentum doesn't guarantee all participants will succeed or endure. The companies combining technical depth with compliance rigor, product excellence with global operational sophistication, will define the lasting shape of China's AI globalization story.
Related Coverage：
China’s AI Large Models Lead Global Token Usage: Industry Shifts and Implications