China's AI price war, data ceiling, and leaderboard games 🇨🇳
| by Kai · The Strategist | 11-min read |
This week's stories converge on a new phase in the China AI race, where cheap models, scarce Chinese-language data, and unreliable leaderboards are shaping the market as much as raw capability. The immediate effect is a pricing war that compresses margins across the stack, but the deeper pattern is that owning data and proving performance now matter more than token counts or benchmark rankings. Chinese labs are simultaneously pushing toward frontier scale, setting up a test of whether they can out-build the US or hit the data wall first.
China's cheap AI models rattle markets but analysts see a demand boom
Chinese open-weight AI models have rattled US investors, but analysts argue the resulting drop in model costs will expand overall AI demand. LLM inference prices have fallen from above US$2 per million tokens in early June to US$1.2 this week, according to Silicon Data's index.
- LLM inference prices per million tokens fell from above US$2 at the start of June to US$1.2 this week, per Silicon Data.
- Silicon Data said competition is up and prices are down, which is good for users and promotes faster AI adoption.
- Analysts say plummeting model costs will benefit the AI industry by boosting global demand for AI systems.
The headline framing is panic, but the underlying signal is a demand event. When Silicon Data's index shows inference prices falling from above $2 to $1.2 per million tokens since the start of June, the instinct on Wall Street is margin destruction. The analysts quoted in the report see it differently. Competition is up, prices are down, and that is precisely what accelerates enterprise adoption. The real story is not that Chinese open-weight models are deflating a bubble. It is that they have reset the economics of AI delivery, and the reset favors scale.
The mechanics of who gains and loses power are now discernible. Pure model providers, whether frontier labs or open-weight platforms, absorb the margin hit as they compete on price. But price per token is only half the equation. Cheaper inference lowers the cost of building agents and applications, which expands total token consumption. That volume shift rewards the infrastructure and downstream layers. Cambricon's 108 percent first-half revenue surge is an early marker: cheaper models mean more inference, and more inference means more domestic chips. Alibaba's Qwen 3.8 launch at 2.4 trillion parameters and ByteDance's push toward a 10 trillion parameter model suggest the frontier race continues even as unit economics compress, but the value capture migrates toward those who control compute supply or application distribution.
The second-order implication most coverage misses is that the binding constraint is no longer model cost. It is data and deployment. China faces a shortage of high-quality Chinese-language training texts, and a robotics startup briefly topped Nvidia on an AI benchmark before being removed. Both stories point to the same shift. As model weights commoditize, advantage migrates to proprietary data sets and physical-world AI, which is why embodied-AI data startup Kaiwang Data raised more than RMB100 million. Goldman's projection of US$13 billion in China AI revenue even as price competition sharpens is the tell: this is a volume business taking shape, not a margin business dying. The Arizona pitch to Taiwanese investors about a "TSMC effect" beyond chips is the same logic in another industry. Commoditized components do not erase value. They compound it downstream, and the strategists who position for that shift will own the next cycle.
| China |
Alibaba launches Qwen3.8 with 2.4T parameters as benchmarks undercut its claims
Alibaba launched Qwen3.8, its latest foundation model with 2.4 trillion parameters, claiming improvements in coding and workplace tasks and a second-place ranking behind Anthropic in one arena. Nikkei reports benchmark tests show Qwen3.8 Max falls short of Alibaba's earlier claims and trails domestic and foreign rivals. The API is already available and Alibaba plans to open-source Qwen3.8-Max next week.
- Qwen3.8 has 2.4 trillion parameters with improvements in coding and professional workplace tasks.
- In third-party Arena rankings, Qwen ranked second only to Anthropic's Claude series.
- Nikkei benchmark tests show Qwen3.8 Max trails various domestic and foreign rivals.
- Qwen3.8 is priced significantly below Moonshot AI's Kimi K3.
China faces AI data shortage as Chinese-language training texts run low
Chinese AI experts warn that a shortage of high-quality Chinese-language training data could become the next major bottleneck for the country's AI ambitions, potentially more difficult to solve than the US chip restrictions. Global supply of high-quality public human-generated text could be exhausted within six years, according to Epoch AI.
- Epoch AI estimates global high-quality public text could be fully exhausted within six years.
- Andrej Karpathy, who helped start OpenAI, cautioned that we could hit a data ceiling before 2030.
- Top American labs are spending heavily to mine offline human knowledge, sparking ethical debate.
Chinese robotics start-up briefly topped Nvidia on AI benchmark before being removed
Chinese physical AI start-up Spirit AI briefly topped the RoboArena robotics benchmark with its Spirit v1.6 model, beating Nvidia, before the benchmark's creators removed it and other models after finding evidence of benchmark manipulation. The episode highlights intensifying US-China competition in next-generation AI and the difficulty of evaluating autonomous systems.
- Spirit v1.6 by Spirit AI briefly overtook Nvidia in the RoboArena robotics rankings.
- The model was removed days later during a methodology overhaul by the benchmark's creators.
- Chinese start-up X Square Robot, which ranked fourth, was also dropped from the rankings.
- The benchmark's lead author cited 'benchmark manipulation' but did not name specific companies.
ByteDance pushes toward 10 trillion parameter model in race with Anthropic
ByteDance is training an AI model with as many as 10 trillion parameters, a size that could rival Anthropic's most advanced systems, according to people familiar with the effort. The Chinese company is still in early pre-training and represents the most ambitious push among Chinese labs to overtake US peers.
- ByteDance's model could have as many as 10 trillion parameters, three times larger than Moonshot's Kimi K3.
- Pre-training typically takes three to six months before fine-tuning, and the exact model size is not yet set.
- Anthropic's Mythos 5 is estimated at about 8 trillion parameters and Fable 5 at about 5 trillion.
- Chinese models from Moonshot and Alibaba have recently lagged only behind Fable 5 on benchmarks.
| Quick hits |
Watch whether ByteDance's reported 10-trillion-parameter model actually ships, since it would test whether Chinese labs can out-build the US frontier or hit the Chinese-language data wall first.
The Asia AI Brief