China's "AI Six Tigers" Generate Only About 10% of OpenAI and Anthropic's Revenue — Commercialization Inflection Point Looms
China's leading AI companies are reaching a delicate inflection point in commercialization: user scale and model capabilities are growing rapidly, yet hard revenue still lags overseas leaders by an order of magnitude. According to the latest report from research firm Epoch AI, as of September 2026, combined AI-related revenue from six Chinese AI companies — Alibaba, ByteDance, Zhipu, Moonshot AI, DeepSeek, and MiniMax — amounts to only about 10% of what OpenAI and Anthropic generate together.
The report systematically maps five monetization pathways for Chinese AI companies — consumer-facing applications, API model calls, enterprise and government customization, licensing fees, and indirect monetization through existing ecosystems — revealing profound structural differences in profitability across business models.
In absolute terms, Anthropic's annualized revenue stands at approximately $65 billion, and OpenAI's at roughly $40 billion. On the Chinese side, ByteDance leads at about $4 billion, followed by Alibaba at approximately $2.4 billion, Zhipu at around $1.8 billion, Moonshot AI and DeepSeek at roughly $1 billion each, and MiniMax at about $800 million — a combined total in the $10 billion range. Notably, the ByteDance and Alibaba figures primarily reflect their Model-as-a-Service (MaaS) revenue, not the complete picture of their overall AI-related businesses.
The Gap Between User Scale and Paid Conversion
The user base for Chinese AI applications is already substantial. According to QuestMobile data, Doubao has approximately 380 million monthly active users, Tongyi around 170 million, and DeepSeek about 130 million. But massive user scale has not directly translated into profit — weak paid conversion rates and low average revenue per user remain glaring shortcomings.
MiniMax's data is the most illustrative. In the first nine months of 2025, its AI-native applications (spanning video generation, audio generation, and the companion app Talkie) averaged approximately 27.6 million monthly active users, yet only 1.77 million were paying users — a conversion rate of about 6% — and the segment's gross margin was a mere 4.7%. During the same period, its open platform and enterprise services business achieved a gross margin of 69.4%. The stark contrast between the two clearly exposes the difficulty of consumer-side monetization.
Low conversion rates are not unique to the Chinese market. According to The Information, among ChatGPT's roughly 900 million weekly active users, paying users accounted for about 5% in 2025, costing OpenAI $3.9 billion annually in inference for non-paying users. But per-user pricing in China is lower — MiniMax's Hailuo video generation standard plan starts at just CNY 55 (approximately $8.2), and Zhipu's chatbot VIP subscription costs about CNY 79 (approximately $12), while ChatGPT Plus starts at $20 per month. Advertising revenue can partially fill the gap — Talkie's ad revenue accounted for 20.9% of MiniMax's total revenue in the first nine months of 2025 — but it cannot fundamentally reverse the structural low-margin nature of consumer-facing business.
The Pricing Dilemma of Open-Source Strategies
Chinese AI companies generally favor open-weight model releases, a strategy that effectively accelerates ecosystem expansion but also exerts sustained pressure on API pricing power.
Take Zhipu's GLM 5.3 Flash as an example. On August 26, 2026, Zhipu listed the model on OpenRouter at a 50% promotional discount while simultaneously releasing the weights. Within hours, third parties began offering services; within a day, 12 competitors had emerged; by September 15, the number had grown to 26, with 8 offering prices below Zhipu's. Zhipu's token share on OpenRouter plummeted from 88% to 22% within 20 days, and daily token volume dropped 70%, even as overall usage of the model grew 17% during the same period.
Epoch AI's analysis shows that when open-source model developers price above the lowest hosting price, their own API's token share shrinks significantly. Using DeepSeek as an example, the report estimates that only about 8% of total customer spending on DeepSeek models flows to DeepSeek itself.
On gross margins, Zhipu's API business was approximately 18.9% in 2025, improving to 24.6% in the first half of 2026. DeepSeek is a rare exception, with estimated gross margins of 70% to 80% on its V4 series API, primarily because it allocates over 70% of compute to model training rather than inference, and has driven down unit inference costs through efficiency optimization.
New Explorations in Enterprise Customization and Licensing Fees
Enterprise and government customized deployments were once a relatively high-margin business for Chinese AI companies. Zhipu is the archetypal example: in 2025, private deployment accounted for 73.7% of its total revenue, with a gross margin of approximately 48.8%, serving government agencies, state-owned enterprises, and banks. It previously partnered with Hangzhou City Investment Group to build vertical models for public transportation, flood control, and road and bridge maintenance.
However, Zhipu has proactively scaled back this business. Its interim report noted that delivery and maintenance costs for such projects are relatively rigid, limiting economies of scale. Private deployment revenue share fell from 73.7% in 2025 to 13.5% in the first half of 2026, as the company pivots toward standardized cloud services and API offerings.
Alibaba and ByteDance are also active in enterprise and government solutions. In May 2026, Alibaba Cloud partnered with Tongji Hospital on medical imaging and research model training; ByteDance's Volcano Engine collaborated with China International Capital Corporation to develop customized financial models. Compared with pure model startups, the two tech giants hold clear advantages in sustaining this business, given their broader enterprise customer bases and cloud infrastructure.
Licensing fees represent an emerging monetization avenue, but publicly available data remains extremely limited. On July 27, 2026, Moonshot AI released Kimi K3 weights under a modified MIT license, stipulating that if a licensee operates a Model-as-a-Service business with annualized revenue exceeding $20 million, it must enter into a separate commercial agreement with Moonshot AI. According to Reuters, Moonshot AI is seeking up to 30% revenue sharing from K3-related services on Microsoft Azure, AWS, and Google Cloud. But the license faces practical enforcement uncertainties: the legal status of model weight copyright remains unsettled, tracing commercial use of specific models and licensee revenue poses operational difficulties, and internal-use scenarios are explicitly exempt from the agreement.
Zhipu is exploring a similar path. Amazon integrated Zhipu's GLM-5.3 into Bedrock in October 2026 and will share revenue with Zhipu. Zhipu has also signed revenue-sharing agreements with multiple major cloud service providers both in China and abroad, though specific ratios have not been disclosed.
The Indirect Monetization Logic of Tech Giants
For Alibaba and ByteDance, the strategic value of AI models lies substantially in their pull-through effect on existing businesses, rather than direct monetization of the models themselves.
Compute demand is the most direct beneficiary. The widespread adoption of AI applications has driven cloud computing demand. Alibaba's AI cloud and compute services revenue reached $7.1 billion in the second quarter of 2026, with year-over-year growth accelerating to 45%; AI-related product revenue hit $1.824 billion, marking 12 consecutive quarters of triple-digit year-over-year growth. Alibaba Group CEO Eddie Wu said the thriving open-source model ecosystem is driving greater demand for cloud computing services, creating a virtuous cycle.
The advertising ecosystem is where ByteDance benefits most prominently. The Seedance video generation model has significantly reduced production costs for short-drama content. In the first quarter of 2026, approximately 128,000 micro-dramas went live, with over 95% being AI-generated content, compared with roughly 33,000 for all of 2025. Short-drama producers then advertise on Douyin to acquire traffic. According to an industry insider, peak daily spending by AI short-drama advertisers on ByteDance's platforms reached approximately CNY 120 million (approximately $17.9 million).
According to IDC data, Volcano Engine handled 49.5% of China's public cloud large-model token call volume in 2025, with MaaS revenue of approximately $221 million that year. ByteDance recently raised its 2026 target to $4 billion.
Epoch AI's report does not point toward a pessimistic conclusion. For large tech companies with cloud infrastructure, advertising platforms, and content ecosystems, AI models function more as amplifiers — generating indirect commercial returns by driving compute sales, enhancing platform stickiness, and boosting ad monetization. For pure model companies, the tension in path selection is more acute: open weights buy ecosystem influence and model diffusion speed, but also cede pricing power; closed-source models help preserve profit margins, yet risk losing first-mover advantage in ecosystem competition. Modified licenses and cloud platform revenue sharing are emerging compromise solutions, but their scale contribution remains to be validated over time.
Capital Markets' Sober Reassessment
Commercialization pressure is being directly reflected in capital markets. Listed Zhipu and MiniMax provide a clear reference point: both companies' interim reports for 2026 show that revenue in the past six months exceeded all of last year, yet after excluding book items such as fair value revaluation of preferred shares arising from their IPOs, adjusted net losses are widening. The reasons are strikingly similar — persistently high training costs for next-generation models, and escalating compute and inference costs driven by growing user numbers and API call volumes, are eating into profits.
The two companies listed on the Hong Kong Stock Exchange in January this year, with peak market capitalizations exceeding HK$1 trillion (approximately $127.5 billion) and approaching HK$400 billion (approximately $51.0 billion) respectively. They have since fallen back to approximately HK$300 billion (approximately $38.2 billion) and over HK$70 billion (approximately $8.9 billion). The fundamental reason for the valuation correction is not market sentiment but financial fundamentals: revenue is rising, but profitability has not kept pace.
This reality is squeezing the valuation space for unlisted large-model companies. Moonshot AI, according to Bloomberg, has completed its latest funding round at a valuation of approximately $50 billion and plans to IPO in Hong Kong in the first quarter of 2027. Its annualized recurring revenue (ARR) climbed from $100 million in March 2026 to $1 billion in August, and according to Bloomberg sources, ARR could reach $2 billion by year-end. But an industry consensus has formed: whoever controls more compute resources, produces tokens at lower cost, and commands smoother business models holds the true moat for large-model companies.
Peter Thiel, accepting the Axel Springer Award in Berlin, made a contrarian judgment: the Google of the model era — a company that can capture monopoly value — may never emerge in the AI age. He drew an analogy to the airline industry: airlines create enormous social value, yet profits are destroyed by competition. The value creation X at the model layer may be very large, but the Y that companies can capture may be very low.
Thiel's core argument: models are more like substitutable intellectual commodities. Switching costs for users are low, product differentiation is often a matter of months, users want results rather than specific model brands, and model switching happens in APIs and workflows rather than social relationships. Network effects, if they exist, are more likely to accrue at the workflow, data, and permissions layer rather than in the base model itself. Epoch AI's data corroborates this view: since 2023, the cost of achieving the same capability level has fallen by approximately 47% per quarter, equivalent to roughly 13 times cheaper per year.
But Thiel also concedes this is merely a probabilistic judgment. Chips, power, and scarce capacity remain highly profitable; a handful of companies that embed models into irreplaceable workflows could still achieve monopoly in Thiel's sense. For Moonshot AI, founder Yang Zhilin once said, "An IPO is a means to accelerate AGI — we will move when the timing is right." That moment has now arrived, and the company must board this train.
As Chinese AI model capabilities continue to improve, observing how each company seeks balance between diffusion and monetization will become a critical dimension for assessing the evolution of China's frontier AI landscape.