Compute Scheduling and Embodied AI Lead Today's Two Storylines

Highlights

• Alibaba Cloud's compute scheduling patent wins China Patent Gold Award, said to handle 100x peak-to-trough gaps
• Ant Lingbo embodied challenge names top 15; real-machine showdown set for Nov 14 in Shanghai
• Apple Intelligence adds iPad and MacBook scenario filings, expanding its China footprint

Compute: Scheduling Moves to the Fore

On Nov 6, the 26th China Patent Awards were handed out in Wuxi. Alibaba Cloud's submission, "A method and device for processing user requests," won the China Patent Gold Award, making it the only cloud vendor to take gold this year. The patent targets demand tides in AI hyperscale compute clusters: compute demand can vary 100-fold between peak and trough. Provisioning for peak leaves capacity idle; provisioning for average leaves peak demand short. The patented approach dynamically senses load, intelligently matches requests with heterogeneous compute, scales up at peak and contracts at trough, and enforces multi-tenant isolation. Alibaba Cloud says it has incubated products including Lingjun Zhenwu super-node instances, the Agent Sandbox, and AI inference suites. MiniMax used Agent Sandbox to spin up 5,000 sandboxes within 10 seconds and cut total cost of ownership by 25%, while Moonshot's Kimi achieved elastic scaling of tens of thousands of sandboxes per minute and cut startup time by 50%. (Source: Leiphone)

The value of this thread is not the award itself but that it turns "compute scheduling" from a back-end engineering problem into an industry issue open to public debate. Beyond model capability, whoever can raise resource utilization between peak and trough gains extra room on inference cost.

Embodied AI: From Leaderboard to Real Machines

The first Ant Lingbo Embodied Large-Model Challenge announced its finalists on Nov 6: 15 teams and 59 contestants advanced from 1,855 teams and 3,252 contestants worldwide. The event covered 13 countries and 648 universities and research institutions. GPU input for the preliminary round was raised from a planned 200 cards to 350, providing more than 340,000 card-hours of compute across two stages. From Nov 14 to 15, the top 15 will hold a 48-hour offline real-machine hackathon in Zhangjiang, Shanghai, using the unified Robbyant Forge platform and Songling COBOT MAGIC devices, completing the full workflow from task understanding and data collection to model tuning and real-machine deployment based on the LingBot-VLA 2.0 open-source pre-trained model. Ant Lingbo says its open-source models have accumulated more than 33,000 GitHub stars. (Source: Leiphone)

The 48-hour real-machine stage exposes engineering problems more than the simulation or offline evaluation of the preliminaries: model generalization, software-hardware coordination, on-site data collection — any weak link shows up directly in task success rates. It is also a window into whether embodied large models are actually usable.

Devices and Subscriptions: AI Features Start to Add Up

According to the "Wangxin Shanghai" WeChat account, "Apple Intelligence" generative AI services added two new scenarios — Apple tablets and Apple laptops — and completed filings. The currently applicable scenarios are Apple phones, tablets, and laptops. Previously, on Sept 29, multiple China-market iPhones updated their filings and added large-model filing numbers. (Source: ITHome)

Samsung, meanwhile, is signaling charges: according to reports, Samsung is preparing to test a subscription service called Galaxy AI Premium at $4.99 per month or $47.99 per year. Subscribers get more AI feature quota, with some features moved into the paid tier, including phone-based AI video generation, Bixby Action cross-app automated task execution, and daily briefings. A free standard tier remains but also has usage or quota caps; once exhausted, users must buy credits or subscribe to the Max tier, which offers four times the standard quota. Samsung has not announced an official launch date. (Source: ITHome)

The two take different paths: Apple first addresses compliance filings and spreads features to more devices; Samsung is testing whether to shift generative AI compute costs to users. Whether the latter is accepted will be answered by test data in the coming months.

Two Japanese Labs: IPO and Full-Duplex

Nikkei Asia reported that Japanese AI lab Sakana AI could IPO as early as fall 2027 and has appointed Goldman Sachs as adviser, potentially becoming the first Japanese GenAI developer to go public. Founded in 2023, Sakana AI focuses on sovereign AI scenarios. In November 2025 it completed a Series B at a ¥400 billion valuation and is developing its in-house Sawara series, planning to release a smaller-scale version first in spring 2027. (Source: ITHome)

On the same day, another Japanese lab, Shizuku AI, unveiled the full-duplex model Shizuku-Duplex, calling it the second full-duplex model supporting natural Japanese conversation after GPT-Live, one that starts its thinking process without a complete query. The model splits into front-end interaction and back-end asynchronous processing, supplemented by context compression every five minutes, with 80ms as one interaction "frame," and supports audio, text, and action outputs. It ranked highest among six tested models for Japanese conversational naturalness and second only to GPT-Live in response quality. (Source: ITHome)

Both threads point to the same thing: Japan's GenAI ecosystem is extending from the model layer to the capital layer and the interaction layer at the same time, but scale remains small, and Sakana's IPO timetable is still at the "earliest" stage.