AI large model markets are evolving into a tiered competitive landscape, with Chinese AI labs positioned to capture a significant global share due to their pronounced cost advantages.
According to a recent report by Bernstein analysts including Robin Zhu, even when accounting for geopolitical constraints and assuming minimal penetration in the US market, Chinese AI labs could access approximately 35% to 40% of the global AI market's total addressable market (TAM). In absolute terms, this represents a value between $320 billion and $350 billion.
The report suggests that the high token pricing of Anthropic's Claude Fable 5 has prompted developers to reassess AI costs. This event signals a growing market focus on return on investment for AI, which will accelerate the migration of users towards more cost-effective models.
This analysis has profound implications for the global AI industry structure. In Bernstein's base case scenario, leading US labs will continue to dominate highly specialized, premium-priced frontier applications. However, in broader, more commoditized "tail" application areas—such as consumer-grade services, small-to-medium enterprises, and emerging markets—Chinese AI labs, with their lower token prices, are expected to systematically gain market share.
A New Framework for AI Commoditization: Perception Trumps Algorithm
Bernstein proposes a novel framework for analyzing AI commoditization that differs from conventional views. Traditionally, commoditization is seen as stemming from the convergence of underlying model intelligence. Bernstein argues the true driver is human users' perception of model capabilities and the point at which a model becomes "good enough" for a specific application and can operate reliably at scale.
The report ranks AI application scenarios by their speed of commoditization: consumer scenarios (like ordering food or booking hotels) will commoditize first, followed by deterministic enterprise workloads (like Excel modeling), then complex strategic planning and cybersecurity. Frontier scientific fields such as drug discovery, nuclear fusion, and space exploration will commoditize last, as users in these areas have near-infinite willingness to pay, sustaining high premiums for cutting-edge models long-term.
Bernstein notes that initiatives like the AI agent announcements from Tencent's WeChat and explorations within Alibaba's Qwen application indicate that commercial deployment of AI agents for tasks like buying milk tea, flight tickets, or T-shirts is imminent. Once a category of tasks is "solved," the marginal return on additional R&D investment plummets, naturally redirecting lab resources toward more complex frontier challenges.
Market Stratification: US Holds the Frontier, China Targets the Mainstream
Bernstein anticipates the global AI market will develop a two-tiered structure. The top tier, dominated by US frontier labs, will continuously unlock new capabilities for high-paying, increasingly specialized clients. The second tier is the "tail AI" market serving more ordinary enterprise and consumer needs, where competition will pivot to cost-per-task, reliability, and developer trust.
Regarding international markets, acceptance of Chinese AI models is extremely low in the US, somewhat better in Europe, and generally higher in other regions, particularly emerging markets like the Middle East and Southeast Asia.
The report also points out that while the US frontier labs' upgrade to next-generation chips like Rubin from Blackwell may temporarily widen the capability gap with Chinese models, historical patterns of technology diffusion suggest this gap will narrow again. A period of 6 to 12 months is not considered long relative to the inertia of consumer habits and corporate procurement cycles.
Cost Advantages and Profitability Outlook for Chinese AI Labs
Bernstein attributes the cost advantages of Chinese AI labs to several structural factors: lower developer labor costs, a "latecomer advantage" from using global state-of-the-art (SOTA) models as a directional R&D beacon, and flexible utilization of previous-generation chip clusters. These factors collectively ensure that the absolute R&D expenditure of Chinese labs will remain lower than their US counterparts over the long term.
On the path to profitability, the report is cautiously optimistic. Bernstein expects R&D spending at top Chinese AI labs to grow rapidly over the next five years—citing statements from Zhipu AI and Minimax that a 50% compound annual growth rate "would not be surprising." However, as more application scenarios become "solved," the scope of scenarios requiring exponential R&D investment will gradually narrow, allowing R&D expense growth to slow and creating conditions for operating leverage.
On the inference side, the report views the inference margins for most AI labs as "decent to strong." Operating expenses outside of R&D are relatively lean, with marketing spend likely to remain low in an environment where "superior inference capabilities essentially sell themselves." Bernstein's overall assessment is that the evolution of AI commoditization is "actually quite optimistic" for the long-term economic model of AI labs.
Alibaba and Tencent Poised for Early Benefits
Bernstein maintains its Outperform ratings on Tencent and Alibaba, with target prices of HK$780 and US$180/HK$176, respectively.
The report views both companies as key players in the commercialization of AI in China, with Tencent's WeChat AI agent initiatives and Alibaba's Qwen ecosystem pointing toward the early commercialization of consumer-grade AI scenarios.
Regarding the broader AI investment theme, Bernstein states that the market reaction to Claude Fable 5's high token costs may accelerate the timeline for AI developers and users to scrutinize AI ROI.
The report concludes that AI users will choose between models to match the marginal cost of tokens with the marginal benefit of task completion. This logic systematically favors Chinese AI labs that offer "good enough" inference capabilities at significantly lower prices. Bernstein adds that its own experience of consuming approximately 2 billion tokens over the past one to two months has further reinforced this view.
Comments