China built the cheapest AI models in the world. Now it's learning that cheap doesn't mean profitable.
At a September 1 briefing for the UBS Securities A-Share Seminar, analysts laid out a stark new reality for China's AI industry: the model capability arms race is over, and what matters now is token optimization, not token maximization. Chinese developers have caught up to — and in some cases surpassed — their Western counterparts on performance metrics. The question on everyone's mind is whether anyone will pay enough to make it worth their while.
The Token Optimization Shift
Xiong Wei, UBS Securities' China internet analyst, identified three themes defining China's AI sector in 2026: model capability, "token ROI," and monetization. The first is no longer the bottleneck. Chinese open-source models have improved dramatically, particularly in coding and agentic capabilities, closing the gap with GPT-5 and Claude Fable.
The shift is from "token-maxxing" — encouraging maximum AI usage at any cost — to "token optimization," where buyers carefully balance performance against price. After rising AI bills made it harder for enterprises to measure the economic value generated by their token consumption, procurement teams began demanding more from less.
This trend favors Chinese open-source models precisely because of their cost advantage. Xiong estimated that leading Chinese models cost less than one-tenth as much to develop as overseas peers, with API pricing at just 10% to 20% of international competitors. For repetitive or lower-risk workloads — content moderation, customer service chatbots, data classification — the cheaper models are "good enough," and the margin difference is decisive.
The Content Problem
But cheaper AI doesn't automatically solve China's fundamental constraint: user attention is finite. Kenneth Fong, UBS's head of China internet research, noted that major Chinese platforms have hit a wall. User traffic and time spent on mobile devices are no longer growing significantly. AI can lower content-production costs and make advertising more effective, but consumers still have limited hours in the day.
The short-drama industry illustrates this perfectly. China already produces 470 AI-generated short dramas daily — more than Hollywood's annual output. Yet producing more shows at lower cost doesn't mean viewers will watch more of them. Attention is the scarce resource, not content.
Mango TV's recent experiment captures both the promise and the paradox. "The Later Journey to the West," an AIGC-produced fantasy series, premiered on August 31 on Mango TV and Hunan Satellite TV's prime-time slot — a first for a Chinese AI-generated long-form series. The show logged 27.57 million plays by September 2, ranking first among provincial satellite channels in its time slot.
The production used Mango Lingchuang, the platform's in-house AIGC production system, which had served over 40,000 professional users and supported more than 3,900 projects by mid-2026. The platform generated 109 character assets and 143 scene assets for the series. It's also testing a "produce, review and broadcast in parallel" model, allowing later episodes to remain in production as earlier ones air.
Yet whether this efficiency translates into sustainable business remains unclear. Mango Excellent Media's stock surged 44% on the news, but the underlying economics of AI-generated entertainment are still unproven. Can you build a profitable studio model on content that costs 90% less to produce but competes for the same limited audience?
The Monetization Gap
The challenge extends beyond entertainment. Chinese AI companies are shipping models at scale but struggling to convert usage into revenue. The open-source model that powered DeepSeek's rise gave away frontier capability for free, training the world on Chinese efficiency — and then wondering why nobody was paying.
ByteDance's $29.6 billion syndicated loan, reported this week, suggests the company is betting big on AI monetization. But loans don't pay for themselves. The real question is whether Chinese AI can find paying users at scale — not just in China, but globally, where cost competition is fierce.
The open-weight movement offers one path. By publishing models on Hugging Face — now owned by Nvidia — Chinese labs can reach global developers who might build commercial products on top of their work. But Nvidia's acquisition raises questions about whether the platform that hosts China's AI rebellion will remain neutral ground.
The Hardware Constraint
Even as software costs plummet, hardware remains a bottleneck. Chinese carmakers face rising costs for printed circuit boards and other components as global supply chains strain under AI demand. CXMT's HBM3E production, while a milestone, operates at 25% yields — far below Samsung and SK Hynix. Without EUV lithography, China's path to advanced chip independence remains long and expensive.
Yet the cost advantage persists at the model level. Chinese labs are proving that you don't need the most expensive chips to build competitive AI — just smarter engineering, better optimization, and a willingness to ship open weights that others can improve upon.
What This Means
China's AI story is shifting from "can we build it?" to "can we sell it?" The capability gap has closed. The cost advantage is real. But the monetization puzzle remains unsolved.
The short-term drama industry experiment shows that AI-generated content can capture attention — at least temporarily. The long-term question is whether that attention translates into recurring revenue, or whether Chinese AI becomes the world's most efficient commodity: everywhere, indispensable, and underpriced.
In the race to AI sovereignty, China has won the cost war. Now it's learning that winning the cost war isn't the same as winning the business war.
🔥 Hot Takes
1. China proved you can build frontier AI for a tenth of the cost — and now realizes that's both its greatest strength and its biggest problem. The DeepSeek effect showed the world that American AI pricing was a rent-seeking monopoly. But undercutting prices by 90% doesn't build a sustainable business; it builds a commodity trap. China's AI labs are the world's most efficient manufacturers, but manufacturing is the lowest-value activity in the AI stack.
2. The Mango TV drama proves AI content works — but 27 million plays for a single premiere doesn't equal a business model. Yes, it's cheap to produce. Yes, it got ratings. But does anyone subscribe to Mango TV specifically for AI dramas? Or is this just more content competing for the same shrinking attention pool? The 44% stock surge suggests investors are excited, but excitement without unit economics is just speculation.
3. Nvidia buying Hugging Face just as China's open-weight revolution peaks is the ultimate geopolitical irony. Chinese labs publish their best models on the platform to reach global developers — and now that platform is owned by the American company that profits most from keeping the world dependent on Nvidia hardware. The open-weight movement wanted to democratize AI; instead, it's being acquired by the very monopoly it sought to circumvent.