🇨🇳 Just how big is China’s AI market?
Reading into China's token-omics and its biggest players
Two quick announcements before I jump into today’s post!
One, I will be chatting with in my live show tomorrow at 12pm ET / 9am PT. You’ll need the Substack app to join the conversation. Two, I am speaking at SXSW next week. If you’re there, join me for an Exponential View meet-up.
OK, now onto our regular programming.
In today’s landscape of artificial intelligence, the token has emerged as the fundamental unit of machine intelligence. Tokens are the building blocks of AI communication—the bite-sized pieces through which humans and machines exchange information. When you type “What is the weather today?”, an AI system breaks this into individual tokens like “What”, “is”, “the”, “weather”, etc., processes them, and then generates new tokens that form its response. Tokens transform your question into something the AI can understand, and the tokens it produces in return represent the machine’s carefully constructed answer.

Understanding the scale of the token market is incredibly valuable, as this insight provides a clear, intuitive grasp of the market size of the genAI economy.
It helps estimate the infrastructure necessary to sustain growth, particularly in terms of computing resources like GPUs and other specialized chips. While the initial demand for AI training capabilities fuelled Nvidia’s historic market capitalization surge past $3 trillion in 2024, sustained token usage for inference — representing demand for AI — will be crucial to maintaining this valuation going forward.
But accurately assessing this market, especially in the US, can be challenging. Major technology firms typically remain guarded and reveal only limited information about their actual usage. So analysts often rely on reported revenue figures, which, while helpful, give only partial visibility into true market dynamics.
Now valuable insights have begun to emerge from the Chinese market. Recently disclosed data points from China — like the DeepSeek numbers we referenced in the Monday email — offer a rare insight into actual usage patterns. This can help us construct a more comprehensive picture of the global token economy.
Software markets commonly exhibit a distinct pattern of concentration: a small number of companies or products tend to dominate a given segment or geographic market. Because this dominance is so pronounced, even limited data about the largest players can significantly illuminate the overall structure and size of the market.
Helpfully, this power law dynamic considerably simplifies market analysis. Even a few data points, like those emerging from China, can provide useful insights into the global AI token economy.
This approach has notable limitations, too: it isn’t bottom-up and captures only a snapshot in time—specifically, the token market from late 2024 to early 2025. However, it helps us understand the scale of the numbers involved and the implications for infrastructure and investment.
Over to China
Looking at this particular report, Baidu’s Wenxiaoyan service was delivering 1 trillion tokens per day in August 2024, with API calls increasing 30-fold over one year. Bytedance’s Doubao exceeded 4 trillion tokens per day following several price cuts, with token usage growing 33 times in a single year.
The report also hinted that there were 200 Chinese firms delivering at least 1 billion tokens per day. We know that Doubao sits at 4 trillion per day, while the 200th-ranked firm delivers around a billion tokens per day. DeepSeek recently announced they served 750 billion tokens daily to a diverse range of customers, both domestically and internationally.
Using these three data points, we can begin to estimate the market’s overall size. I used a modified Zipf distribution, a discrete probability distribution that follows a power law where the frequency of an element is inversely proportional to its rank. In market analysis, Zipf’s law often manifests when the market share of the nth largest company is approximately proportional to 1/n. I’ve adapted this distribution to account for the specific characteristics of the token market, allowing us to estimate the entire market from limited data points about the largest players.
The chart above provides five different distributions of token usage by the largest Chinese genAI firms, ranging from the most concentrated market (orange 🟠) to the least (green 🟢). Each resembles distribution curves observed in other markets: orange like web search, blue like enterprise software, red like cloud computing and yellow and green like non-technology markets.
Summing across all samples allows us to estimate the market’s total size and make informed guesses about market concentration.
My best guess is that the Chinese genAI market today resembles something between the orange and blue curves, with Doubao, Wenxiaoyan and DeepSeek as the largest players. The market might be less concentrated publicly, though probably not significantly below the blue curve.
I’ve built a preview calculator that lets you play around with some of the parameters here. This simplified version isn’t the actual model I used above.
Let’s do a sense-check. We know that DeepSeek has said that they served 750 billion tokens a day and ranks as China’s second-largest AI app behind Doubao. Against my orange curve, this would place DeepSeek third—not bad, considering user leaderboards don’t account for API usage.
Okay, so this is truly a napkin exercise, but it lets us say that for public-facing genAI applications, the Chinese market is running somewhere between nine and eleven trillion tokens per day—remarkable, given that number was likely zero just four years ago. To contextualize this scale: if these tokens were represented as standard English text, the daily Chinese token processing would be equivalent to processing the entire Library of Congress—approximately 51 million documents—every single day.
Is this likely to remain the shape of the market?
Likely not, as the market continues evolving. The US cloud computing market exhibits high concentration, forming a triopoly. Conversely, global social media and messaging markets are more heterogeneous, clustering around several large players. Web search evolved into a winner-take-all market, becoming a near monopoly.
What does it tell us about the compute fabric available in China?
Currently, with only 2,200 H800 GPUs, DeepSeek processes 750 billion tokens daily. If all Chinese companies matched DeepSeek’s efficiency, the entire Chinese market could run on 26,000–32,000 H800 GPUs. While far from all companies currently achieve this level of efficiency, there’s no barrier preventing them from doing so soon, especially since DeepSeek openly shares its efficiency techniques. (Indeed, Alibaba just released a reasoning model even more computationally efficient than DeepSeek’s.)
For context, Nvidia produced over one million H20 GPUs (the successor to H800) for the Chinese market last year, indicating significant inference capacity.
This suggests there is plenty of additional capacity for inference coming. Companies may hit hitting limits today, but the combination of growing supply (not just form Nvidia but also Huawei) and software optimisations will create headroom to accommodate growing demand.
However, considering Wenxiaoyan and Doubao’s 30-fold demand increase within one year, if this trend continues—as I believe it will—they’ll require nearly one million H800-equivalent GPUs, assuming no further algorithmic gains. Transitioning from chatbots to reasoning agents drastically increases token consumption per interaction. Thus, substantial additional chip requirements are imminent. Jensen Huang has suggested that reasoning models demand 100 times more compute than traditional ones, with future needs potentially millions of times higher.
We can, and I probably will, apply a similar analysis to the US market. When I do this analysis, I’ll actually split the market in two. One the one hand, I’ll look at the consumer research market (ChatGPT, Claude consumer, Perplexity) and on the other the enterprise/API-driven market to get some kind of insight into the scale of usage and likely concentration of market power. This bifurcated analysis will likely reveal different concentration patterns—consumer markets tend toward winner-take-most dynamics due to direct network effects, while enterprise markets often support multiple significant players due to industry-specific requirements and relationship-based selling.
This approach, admittedly, is a rough sketch. It doesn’t fully capture internal enterprise usage of tokens, and reported data spans six months. It’s possible Doubao isn’t the market’s largest player. Moreover, the distribution of new entrants remains uncertain, and new entrants can significantly disrupt markets, as Google once did to Excite, Lycos, AltaVista and others. But it does help us explore the contours of the market.
Do the recently announced diffusion LLM influence any of this