By |
Edited by | Wen Shuqi
These past two days in Shanghai, it’s safe to say the city has the highest concentration of Tokens in the country.
At the 2026 World AI Conference (WAIC), the term “Token factory” popped up everywhere. In the H2 computing power zone, one exhibitor even set up a dedicated English-only “Token for Export” area, showing off how to package computing power into standardized Token services ready for global markets. This drew a crowd of international visitors, especially from the Middle East. Over in the H1 hall, besides major model companies and tech giants, you could find plenty of these “Token factories” making a business out of this new trade.
Tokens are the basic unit for processing text in AI models. Every time a user asks an AI a question or writes code, they’re burning through tokens. NUPIAO noticed something new this year at WAIC: Tokens have become a digital commodity you can price, trade, and ship across borders. The whole lifecycle—production, pricing, management, and delivery—is spinning up a brand-new industry chain.
At WAIC, Zheng Weimin, an academician of the Chinese Academy of Engineering and a professor at Tsinghua University, laid out what it takes for big models to go global. He said you need three things: a high-quality model, a “Token factory” that can produce top-notch tokens, and cheap electricity. Get those three right, and your tokens can hit the world stage.
In fact, China is already the world’s biggest Token supplier. Recent data from the National Data Administration shows that daily AI token calls in China have exploded from about 100 billion in early 2024 to over 140 trillion by March 2026—a mind-blowing 1,400-fold increase in just two years. OpenRouter data reveals that tokens from Chinese models now account for 54.1% of global consumption, outpacing the 41.5% share from American models. In coding scenarios, the Chinese model share jumps to a massive 67.6%.
PPIO is one of the fastest players in this “Token for Export” race. At their booth, NUPIAO learned that they’ve already signed up over 660,000 registered developers and listed more than 200 big models. Their network of over 5,000 computing nodes spans six continents. In 2025, their average GPU utilization rate was over 75%.
“We can deliver a full suite of Token services through our computing platforms, both in China and overseas,” a PPIO rep told NUPIAO. Thanks to this global compute setup, model makers and users don’t need to build their own overseas infrastructure to produce and use tokens.
So, with production running smoothly, who’s buying these tokens?
Right now, the main customers for Token factories are individual developers and businesses, both at home and abroad. The biggest use case is AI Coding, which is also where the most money is being made.
“Coding involves complex contexts and puts a lot of demand on the inference engine,” the PPIO rep explained. Some programming tasks can require users to input data that goes beyond a million tokens. This means Token factories need to be super competitive in areas like caching tech, inference performance compression, and context compression.
Another explosive area is AI Agents, one of the hottest topics at this year’s WAIC. “The token consumption for a single task from these tools is growing exponentially,” the rep said. “Some companies want to set up secure, isolated sandbox environments for their deployment.” That’s exactly where Token factories add serious value.
As for pricing, the Token business model is similar to public cloud services: you pay for what you use. NUPIAO found that the standard model in the industry is pay-per-token, often called MaaS (Model as a Service). There are also some options for private deployment and custom fine-tuning, which are charged on a project basis.
The real reason China’s Token exports have reached this scale comes down to a simple formula: high performance at a low cost.
“The gap between domestic models and overseas ones is shrinking, but our cost advantage is clear,” a PPIO exec told NUPIAO. China’s deep talent pool and engineering dividend mean that model R&D and service costs are way below those of foreign competitors. Another game-changer is open source. Major Chinese model players like DeepSeek, Alibaba’s Qwen, Kimi, Zhipu, and MiniMax are almost all taking the open-source route.
The beauty of open source is that companies everywhere—whether in Europe, the US, Southeast Asia, or the Middle East—can deploy Chinese models locally and use their own data to power their AI. This sense of “control and autonomy” is something closed-source models can’t easily offer, especially in today’s world where data sovereignty is a big deal.
“A Token is a standardized product, but the competition behind it touches everything,” another PPIO exec pointed out. Some clients want speed, others need long context, some are price-sensitive, and others require high concurrency. The real moat for a Token factory is to excel in three areas at once: the scale of computing power, network optimization, and the inference engine.
At the same time, the adaptation of domestic chips is speeding up. This year’s WAIC also featured a major showcase of homegrown computing power innovations.
NUPIAO heard from many exhibitors that domestic chips, particularly those from Huawei, are already being used at scale for model inference services. “In MaaS scenarios, they’re already working really well.” While training scenarios still rely heavily on high-end chips, the “domestic alternative” for inference is becoming a reality.
For a long time, the main story for Chinese companies going global was about “product export”—selling hardware and software overseas. At the SenseTime booth, NUPIAO learned that they’ve been doing AI exports since 2018, with a footprint in Southeast Asia and the Middle East. But most of their orders are project-based, which requires deep, long-term local engagement. This stands in stark contrast to the standardized, pay-as-you-go Token model.
Looking at WAIC today, a new division of labor is emerging in China’s AI push abroad. At the bottom layer, you have the computing infrastructure. In the middle, Token factories provide standardized Token products. And at the top, industry-specific Agents and AI applications offer brands and services directly to end users.
AI is helping product exports climb a higher value ladder.
David, the product head at Titan Technology’s Navos, sees that in the past, Chinese products going global were mostly about low-cost supply chains. But today, AI is leveling the playing field, giving more Chinese brands a shot at the world market.
Building a brand requires marketing, but in the past, only big clients or big businesses could afford to do overseas marketing. David ran the numbers for NUPIAO: planning a single overseas marketing campaign costs a ton—you need a dedicated team and a serious budget. “But AI is changing all that. Now, small and medium businesses, even individual sellers, can also do brand exports.”
NUPIAO saw a demo at their booth. Someone told the Navos AI agent: “Sell a coffee brand targeting young women in Southeast Asia.” The system automatically broke down the task, matched it with overseas influencer styles, generated multiple video versions, and finally offered a full marketing plan. A project that used to need a team of 10 people is now handled by AI itself.
At the Token booths, the crowd was mostly techies, discussing latency and pricing. Over at the Titan Technology booth, it was e-commerce sellers and brand owners looking for opportunities, talking about ROI and conversion rates.
These two booths were right next to each other, and the crowds didn’t seem to mix. But they share the same logic: when computing power and AI become as cheap and accessible as water and electricity, the barrier to going global drops dramatically. Today, China’s new AI business for global markets has moved into the practical “how-to” stage.
