On the morning of October 27, Zhipu — the "world's first large-model stock" — surged as much as 12 percent in Hong Kong trading, with fellow Chinese model maker MiniMax following with gains above 5 percent. The catalysts were twofold and foreign: Anthropic's IPO filing and OpenAI's half-price GPT-6.1 Terra launch put a fresh bid under the entire model layer, reminding global investors that listed exposure to frontier models now exists — and that the deepest open-weight ecosystem sits in China.
The Token Ledger Tells the Story
OpenRouter's weekly data for October 19 to 25 makes the shift measurable: global model calls reached 188 trillion tokens, up 14.63 percent week over week, and Chinese models accounted for 93.54 trillion of them — up 18.14 percent and marking a 26th consecutive week in which Chinese models out-consumed American ones. DeepSeek's Flash 1017 alone processed 33.6 trillion tokens in the week, topping the global usage chart. Usage, not benchmarks, is where ecosystem gravity shows up first.
Zhipu's Fundamentals Catch Up With Its Ticker
Zhipu's rally is not purely narrative. The company reported first-half 2026 revenue of 953.9 million yuan, up 399.7 percent year over year, with MaaS open-platform and API revenue of 825.2 million yuan — up an astonishing 2,735.7 percent and now 86.5 percent of the total. Its full-business annual recurring revenue stands at 1.8 billion dollars, and management raised the 2026 year-end ARR guidance from 2.4 billion to 3 billion dollars. Goldman Sachs upgraded the stock to Buy and lifted its own year-end ARR forecast to 3.2 billion dollars. Amazon Bedrock has integrated Zhipu's GLM-5.3 with a revenue-sharing arrangement, and management says similar cloud partnerships at home and abroad will start contributing recognized revenue from October.
Zhipu H1 revenue: 953.9 million yuan, up 399.7 percent; API revenue up 2,735.7 percent.
ARR: 1.8 billion dollars; year-end guidance raised to 3 billion dollars; Goldman at 3.2 billion.
OpenRouter: Chinese models at 93.54 trillion weekly tokens, 26 straight weeks ahead of the U.S.
DeepSeek Flash 1017: 33.6 trillion tokens in one week, No.1 globally.
DeepSeek: The 15 Billion Dollar Elephant
DeepSeek closed its latest round at roughly 100 billion yuan — about 15 billion dollars — the largest single AI financing in Chinese history, after raising its target twice from an initial 50 billion yuan. CATL and Tencent led a roster stacked with industrial capital, bringing the year's cumulative raises to roughly 150 billion yuan at a pre-round valuation near 74 billion dollars, against about 1 billion dollars of annualized revenue. CITIC Securities is advising on a STAR Market IPO targeted for early next year — which, if the timeline holds, would make DeepSeek the first frontier-lab listing anywhere, potentially beating Anthropic's ANT to the bell and certainly outrunning OpenAI and Moonshot.
The same week carried a pointed juxtaposition: the U.S. Department of Homeland Security issued a federal exclusion order barring not just DeepSeek products but third-party services built on its models — and DeepSeek answered the ban by fully open-sourcing DeepInfer, its production inference engine, supporting both NVIDIA and Huawei Ascend hardware. The economic daily commentary framed it bluntly: you can lock institutions out of an account, but not out of open source.
The Open-Weight Wave Behind the Rally
The funding story sits on top of a relentless open-weight drumbeat. Moonshot's Kimi K3.1 Pro — a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters and a million-token context — promised full open weights by November 4, with pricing cut 35 and 32 percent to 13 and 68 yuan per million tokens. MiniMax published the weights of M3.1 Flash (428 billion total parameters, 23 billion active) on Hugging Face and ModelScope. Tencent open-sourced Hy-SimulMT-30B-A3B, a speech-to-speech simultaneous interpretation model with roughly 3-second latency and real-time voice cloning under Apache 2.0. Nanbeige released its 33B agentic coding model, and Hugging Face with academic partners shipped Carbon-X, a 14.8-billion-parameter open DNA foundation model. Capital, compute and open weights are now reinforcing one another — the flywheel American labs built on closed APIs is being rebuilt in public, at commodity prices, in Shanghai, Beijing and Shenzhen.
Sources: Sina Finance, NetEase, China Economic Net, AI225 Daily, AI Native Foundation, AIGC.NEWS, threerouter
Top comments (0)