Over the past six months, commercialization of China's large language models has suddenly accelerated.
In December 2025, annualized revenue of China's major AI model companies stood at just $4 billion; now it's close to $13 billion. Growth has picked up notably, especially since June.
Counterintuitively, this growth wasn't achieved through price wars.
Zhipu's average API prices rose 101% this year, yet token usage surged more than 40-fold. DeepSeek also raised prices substantially without a significant drop in usage.
Overseas expansion has become a key growth driver. MiniMax generates over 60% of revenue from abroad, Kimi's overseas revenue has surpassed domestic, and Zhipu's overseas share is estimated at nearly 40%. This marks the first time Chinese AI has produced software exports at a meaningful scale.
As revenue grows, model company valuations are diverging. Based on latest annualized recurring revenue, DeepSeek trades at under 70x ARR, while Zhipu and MiniMax are at 36x and 20x, respectively.
Why has domestic large model revenue suddenly accelerated? And why do valuations differ so much among similar model companies? Today, we explore both questions.
China's model revenue growth is accelerating.
Data show that growth in China's large models has been accelerating since June.
In December 2025, the annualized revenue of China's major AI model companies was only about $4 billion. By March of this year, it had risen to $6 billion. Now, that figure is approaching $13 billion.
In other words, in eight months, the annualized revenue scale of Chinese AI model companies has grown from $4 billion to $13 billion.
And the pace of growth is still accelerating.
From December 2025 to March of this year, it increased by about $2 billion in three months. From March to June, it added another $4 billion. From June to August, it took only two months to add about $3 billion more.
In other words, at the start, ARR was adding roughly $600 million to $700 million per month. In the last two months, it has become more than $1 billion per month. The commercialization of large models is clearly accelerating.
Among them, the most representative is Zhipu. In March of this year, Zhipu's MaaS, or API business, had an ARR of only about $250 million. By July, it was approaching $1 billion. In August, based on annualized monthly revenue, ARR reached $1.6 billion. Based on the latest week's annualized revenue, it has even reached $2 billion.
The API revenue in August alone has already exceeded Zhipu's total API business revenue for the first half of 2026.
If current trends continue, Zhipu's API ARR could approach $2.5 billion by year-end.
MiniMax shows a similar growth curve. In December 2025, its ARR was only about $100 million; by February 2026, it exceeded $150 million; from April to May, it had grown to $300 million-$430 million. By August, based on the latest weekly revenue annualized, ARR had surpassed $800 million.
Management previously set a target of exceeding $1 billion in revenue by the end of 2026.
Behind this round of growth are two counterintuitive developments:
First, the growth was not driven by price cuts.
Over the past two years, one of the most notable labels for China's large language models has been price wars. Models are getting cheaper, and tokens are getting cheaper. So when seeing growth in usage and revenue, many people's first reaction might be: is this again trading price for volume?
But the latest data shows the opposite.
Starting in August, API prices for various DeepSeek products increased roughly threefold to twelvefold.
Based on peak-hour calculations, compared with the old prices at the end of July, various billing items generally rose by about 3 to 12 times.
Among them, the V4 Pro cache-hit input price rose from $0.003625 to $0.044 per million tokens, an increase of more than 12 times; the output price rose from $0.87 to $3.96, up about 4.6 times.
Despite such sharp price increases, usage has not declined significantly.
Zhipu is in the same boat. Since the start of this year, the average price of Zhipu's API has risen 101%, roughly doubling. At the same time, token usage has grown more than 40-fold.
The second change is overseas revenue.
According to estimates by FD, author of Robonomics, overseas revenue of Chinese AI model companies has now exceeded $2 billion.
Among them, MiniMax has the highest share of overseas revenue, with 60.8% of its revenue coming from overseas.
Kimi's overseas revenue share is also high. As early as February this year, reports said Kimi's overseas revenue surpassed domestic revenue for the first time. By June, the company's ARR had grown to roughly $300 million, with overseas developers seen as a key source of growth.
Overseas revenue already accounts for roughly 40% of Zhipu's total, by the author's estimates. Even for large companies with a predominantly domestic revenue mix such as Alibaba and Tencent, overseas model revenue likely contributes 5%–10%.
This marks the first time Chinese AI has achieved software exports at a meaningful scale.
Valuations diverge, from 20x to 70x ARR.
Even as revenue growth accelerates, a new issue is emerging:
Large-model companies are becoming increasingly difficult to value.
For now, valuations of domestic large-model companies have clearly diverged.
DeepSeek is valued at around $74 billion based on the latest reported funding round, with ARR above $1 billion, implying a multiple of under 70x. Zhipu has a market value of about $71 billion and latest weekly ARR of roughly $2 billion, for a 36x multiple. MiniMax is valued at approximately $15.6 billion with ARR of about $800 million, implying around 20x.
The range from 20x to 70x represents a more than threefold gap, far exceeding the valuation spread among overseas large-model companies. By comparison, since the second half of 2025, the valuation gap between OpenAI and Anthropic has remained around 1x.
Why do Chinese AI model valuations differ so much?
A key reason is that strategic divergence among Chinese vendors is even greater than between OpenAI and Anthropic.
Both OpenAI and Anthropic have top-tier model capabilities and ample computing, capital and R&D resources.
Although their early strengths weren't identical, once a direction is validated, both can quickly reallocate resources.
Coding is the most typical example. After coding proved to be one of AI's most important high-value scenarios, OpenAI quickly focused resources on catching up.
Third-party tracker TickerTrends estimates Codex's annualized revenue was only about $1 billion in February, rose to $5.6 billion by June, and reached $8.8 billion in August. Over the same period, Claude Code was about $15.1 billion.
Claude Code still leads, but the revenue gap has narrowed from about 2.4 times in early July to 1.7 times in August.
But Chinese vendors are a completely different situation.
Given limited resources, no company can simultaneously push every direction to the cutting edge, so each must place early bets on a finite set of fronts. For example, Zhipu initially focused on MaaS, enterprise APIs, and domestic compute adaptation; Kimi bet on long-context models; MiniMax committed earlier to multimodal, video, and overseas markets.
Moreover, Chinese model companies cannot easily throw billions of dollars at a new direction the way OpenAI can once it spots a misstep. That means when consensus shifts, their strategic pivots often take months or longer.
So Chinese model companies' operating results depend far more on early strategic calls than their overseas peers. The wider the strategic divergence, the greater the revenue gap, and the harder it becomes for valuations to converge.
Beyond strategy, the companies also differ in efficiency.
The Information previously reported that DeepSeek's API business gross margin reached 82.9% in the first seven months of this year, while Anthropic's is estimated at over 60%. That means DeepSeek's gross margin is even higher than Anthropic's.
By contrast, Zhipu's API gross margin is only 24.6%, and most domestic text-model APIs still sit in the 40%-60% range.
The gap is striking. At the same $1 billion ARR, an 80% gross margin leaves $800 million in gross profit, while a 20% margin leaves just $200 million — a fourfold difference.
Combined, these factors make it hard to value Chinese model companies using a single uniform ARR multiple.
Only two things truly separate valuations: the growth trajectory and the quality of revenue.
The former determines how fast a company can run; the latter determines how valuable that growth really is.