英偉達推收入分成，首批鎖定 21 萬張卡把簡體中文原文的重點轉成繁體中文讀者更容易閱讀的脈絡。核心在於 Nvidia is extending from GPU sales into financing, revenue sharing and idle-capacity backstops for AI cloud providers.
變化在哪裡
目前可核驗的事實包括：AI Computing Partner Program, Grace Blackwell GB300 deployments, revenue share or equity warrants, buyback or leaseback support for underused capacity, Sharon AI planning up to 40,000 GB300 GPUs, Firmus planning a 360 MW DSX AI factory in Batam with up to 170,000 GPUs, and roughly 210,000 GPUs in the first wave。這些細節讓新聞不只停留在發布會口號。
為什麼重要
The structure can accelerate smaller AI-cloud builders while also tying Nvidia to downstream utilization risk. AI 新聞已經不能只看模型分數，還要看成本、供應、監管與真實工作流程。
接下來看什麼
The stress test comes when GPU supply is no longer scarce and idle-capacity commitments become more expensive. 下一步要看的不是更多宣傳，而是第三方測試、客戶續用與長時間穩定性。
參考來源：NVIDIA blog, Tech Times and DigiTimes、CocoLoop。