China's AI Model Revenue Only One-Tenth of OpenAI's as Price War Shifts to Task-Cost Competition
nashnova research
A Rhodium Group report finds that all Chinese AI models combined generate roughly 10% of OpenAI's annual revenue; meanwhile, Bank of America says China's AI price war is pivoting from token pricing to cost-per-task competition — a shift that will redistribute profits across the entire supply chain.
How much are Chinese AI models actually earning?
OpenAI's annual recurring revenue (ARR) stands at $40 billion; Anthropic's at $65 billion. China's leaders lag far behind — ByteDance at roughly $4 billion, Alibaba at about $2.4 billion.
Smaller players trail even further: Moonshot AI at around $1 billion, MiniMax at $800 million, DeepSeek at just $500 million. Separately, CNBC reports Z.ai disclosed its latest ARR at $1.8 billion.
This means → add up every major Chinese AI company's revenue and you get roughly one-tenth of OpenAI alone. The gap is not "a lap behind" — it is an order-of-magnitude chasm.
Revenue is this low — why are valuations still so high?
Rhodium notes that Moonshot AI trades at a price-to-sales ratio of roughly 50× and DeepSeek at a staggering 163×, versus 34× for OpenAI and 21× for Anthropic.
In plain terms = a price-to-sales ratio divides a company's valuation by its annual revenue. The higher the number, the bigger the bet on future growth. DeepSeek at 163× means investors are wagering its revenue will multiply many times over — if that bet fails, the valuation is a bubble.
Rhodium partner Logan Wright put it bluntly: "The funding gap will make it harder for China's frontier AI labs to scale sustainably." This reflects the core structural risk facing Chinese AI startups — the chasm between low revenue and sky-high valuations.
How did the price war shift from "cheap tokens" to "cheap tasks"?
Bank of America reports that China's large-model price war is moving from "across-the-board token-price cuts" to tiered pricing — basic reasoning is commoditizing fast, while frontier capabilities still command differentiation.
This means → the old benchmark was "how much per token"; the new one is "how much for AI to complete a task." Competition has shifted from unit price to a composite metric of efficiency × cost.
BofA's analyst team argues that models with stronger reasoning, longer context windows, multimodal capability, and agentic workflows can still sustain premium pricing. In plain terms = models that handle complex jobs still command high fees; those limited to simple chat are already losing value.
Who ultimately benefits from this competition?
BofA states explicitly: the key economic metric is shifting from cost-per-token to cost-per-task — and this will reshape profit distribution across China's AI supply chain.
Over the long term, AI chips, semiconductor equipment, memory, and major cloud platforms will capture more economic value than standalone AI-model developers. This means → the "shovel sellers" outperform the "gold diggers" — the infrastructure layer is better positioned to monetize than the application layer.
U.S. export controls have objectively accelerated China's domestic AI development, yet also intensified competition by drawing more players into the race. Who can maintain differentiation at the frontier will be the defining variable of the next phase.
市场有风险，内容仅供研究参考，不构成投资建议。