China's AI Rally Puts US Labs on the Defensive as the Race Shifts to Cost

Why China's AI and Chip Index Is Up 29% While the CSI 300 Stalls

China's technology-heavy Star 50 Index has surged 29% this year, dwarfing a 0.9% gain in the CSI 300 and a 1.5% decline in Hong Kong's Hang Seng Index. Even after recent selling, the index trades at more than 150 times earnings, compared with roughly 35 for the Nasdaq 100. Fund managers say the move is not simply speculative: investors are betting that China's AI ecosystem can close the technological gap with US rivals faster than previously expected.

That repricing is being driven by concrete milestones. Moonshot launched Kimi K3 to compete with Anthropic's models, domestic chipmaker CXMT overtook Tencent to become China's most valuable listed company, and individual investor demand for humanoid robot maker Unitree's planned IPO exceeded the retail allocation by more than 5,500 times. Beijing has reinforced the trend through state-backed national team share purchases and by loosening listing rules: more than half of mainland IPOs this year were technology or industrial companies, while the share in Hong Kong reached 77%.

The optimism sits alongside genuine economic weakness. The property crisis that began in 2022 has weighed on household wealth and consumer confidence, leaving consumer-sensitive internet names down sharply: Alibaba has fallen 17% this year and Tencent 26%. Instead of broad consumer stimulus, Beijing is steering capital toward high-value-added sectors such as chips and AI.

The pressure is now flowing into global pricing. OpenAI cut GPT-5.6 Luna's input price from $1 to $0.20 per million tokens and its output price from $6 to $1.20, while Anthropic introduced Opus 5 at roughly half the price of its more advanced Fable 5 model and scrapped a planned Sonnet 5 price increase. According to Silicon Data, prices paid by customers for leading US lab models have fallen roughly a quarter since mid-July.

The Real Shift: From Model Superiority to Cost per Completed Task

Beijing has turned the AI race into a state-supported capital formation project

China's rally is not only about model quality. Loosened IPO rules, state-backed buying and a policy push for semiconductor self-sufficiency are concentrating capital in AI and chip companies. That support is visible in the numbers: Shanghai IPOs this year have averaged a price-to-earnings ratio of 268, against 67 for last year's listings. The implication is that Beijing is willing to accept frothy valuations in strategic sectors if they accelerate domestic technology development.

OpenAI and Anthropic are defending share with price, not just benchmarks

The cuts from OpenAI and Anthropic show that competition has moved beyond which model is strongest. OpenAI's 80% reduction in GPT-5.6 Luna's per-token prices and Anthropic's decision to release Opus 5 at about half the cost of Fable 5 are direct responses to Chinese labs such as Moonshot and DeepSeek, whose lower-cost models are attracting corporate users including DoorDash and Airbnb. Silicon Data's estimate that leading US model prices have dropped roughly 25% since mid-July suggests the pricing pressure is broad rather than isolated.

Cost per completed task matters more than the headline token price

A simple token-price comparison can mislead. Artificial Analysis found that Anthropic's Opus 5 at a medium effort setting delivers results close to Moonshot's Kimi K3 at maximum effort and at similar cost per task. At maximum effort, OpenAI's GPT-5.6 Luna performs similarly to DeepSeek's V4 Flash but costs about twice as much per task. Because a more capable model can finish a job with fewer tokens or attempts, total task cost, not the sticker price, is the right metric for enterprise buyers.

The real fragility is in the valuation gap, not the technology gap

The Star 50's price-to-earnings multiple above 150 leaves little room for disappointment. Some fund managers caution that China's investor base remains more retail-heavy and short-term oriented than that of US markets, which can make price moves more violent. At the same time, the sharp divergence between AI and chip stocks on one side and consumer internet giants on the other shows how dependent the rally is on Beijing's industrial-policy priorities rather than broad economic recovery.

What the AI Price War Means for Buyers and Investors

• Enterprise buyers should reprice AI contracts against current market prices: the Silicon Data index shows leading US model prices have fallen about 25% since mid-July, meaning pre-July per-token rates are likely stale.
• RFP reviewers should ask vendors to quote cost per completed task, not just token price. Artificial Analysis found OpenAI's GPT-5.6 Luna at maximum effort costs roughly twice DeepSeek's V4 Flash for similar performance, while Anthropic's Opus 5 at medium effort approaches Moonshot's Kimi K3 at maximum effort.
• For investors, compare the Star 50's multiple above 150 with the Nasdaq 100's near 35. This year's Shanghai IPO cohort has averaged a P/E of 268, leaving little margin for disappointment.
• Watch Alibaba and Tencent as a reality check on the Chinese AI trade: the stocks are down 17% and 26% this year, showing that Beijing's high-tech pivot is not lifting consumer-focused internet names.
• Companies with mid-range AI workloads can test Chinese-origin models as a benchmark: DoorDash and Airbnb have already adopted Chinese-built models, and US labs are responding with price cuts.