The biggest gossip in the AI community over the past two weeks: on OpenRouter, an anonymous model called “Ox Alpha” launched and, in just 6 days, consumed 232,000,000,000,000 tokens—2.3× the second-place DeepSeek-V4-Flash, directly ending DeepSeek’s 56-day run at the top of the leaderboard.
On August 26, Zhipu officially claimed it—Ox Alpha is GLM-5.3 Flash, with 320B parameters and 18B activations. Its AA IQ score is 57, matching Claude Opus 4.8. Pricing is 0.8 yuan for input and 2.8 yuan for output per million tokens—aiming to be cheap and generous.
But the thing most related to investors isn’t the model itself—it’s a line in the Zhipu AI technical documentation: “This time, the inference calls a domestic chip cluster with over 100,000 cards. ‘Hardware efficiency and per-token cost have reached a level comparable to mainstream NVIDIA GPUs.’” SemiAnalysis directly commented: “CUDA’s moat is being tested again.”
The official chip supplier hasn’t named anyone; the clues point to Huawei Ascend, Moore Threads, and Hygon. Before the post, Moore Threads’ stock price had already risen by more than 3%.
Three cards—now they’re all in place.
On the same day across the ocean, Nvidia turned in its homework: quarterly revenue of $96.2 billion, doubling year over year. Q3 guidance is $108 billion, and this guidance does not include any China data-center revenue (sales to H-Series H200 are less than 1%). The single most weighty line from Jensen Huang’s conference call: AI has started generating "profitable tokens"—compute power is shifting from technical investment to productive assets that directly create revenue. Translated into plain language: buying cards is buying a printing press.
Nvidia proves that compute power equals money. Zhipu proves domestic chips can handle top-tier large models. And Moore Threads proves that they’ve already started collecting payments.
On Moore Threads’ side: first-half revenue was RMB 1.736 billion, up +147% year over year—more than 2025’s full-year target in just half a year. In March, it signed a RMB 660 million contract for the Kuayo cluster (roughly equal to 44% of 2025 full-year revenue; the remaining payments should be collected before year-end). The Kuayo cluster has already landed on core procurement lists from internet firms, telecom operators, and local computing platforms. It supports training and inference of 600+ large models, with 95% linear scaling efficiency on a 100k-card setup. Guohai Securities expects net profit attributable to shareholders to return to positive in 2027 and reach RMB 1.5 billion in 2028.
Everyone in the朋友圈 is an insider: Zhipu has taken its path from running ChatGLM on thousand-card setups in 2024 all the way through to Day-0 adaptation for GLM-5.3. DeepSeek, MiniMax, China Mobile, ZhiYuan, Pony AI, and Qwen3.5 are all on the list.
One last line: Domestic GPUs have been talking about a "replacement logic" for two years. Moore Threads has already finished half the journey—the other half is profit. Don’t just watch the day-to-day percentage moves— the real heavy hitters are the tail payments to be collected by year-end, in the next earnings report, and in every Day-0 adaptation.
This train of domestic compute power has just pulled out of the station, and Moore Threads is already sitting in the front car. The rest—let time prove it.