The next AI bottleneck isn't only GPUs—it's everything that keeps them busy.
AI 的下一個瓶頸不只是 GPU,而是讓 GPU 持續有效運作的整套配套。
The next AI bottleneck isn't only GPUs—it's everything that keeps them busy.
AI 的下一個瓶頸不只是 GPU,而是讓 GPU 持續有效運作的整套配套。
This week, NVIDIA showed how agentic AI changes the stack: Vera CPUs delivering up to 1.6x more agents per watt vs. the fastest x86, while Vera Rubin NVL72 lands on Oracle Cloud with plans for multi-gigawatt Superclusters. At the same time, reports say NVIDIA is looking at retrofitting telecom buildings into small inference sites—because power, permits, and local pushback now shape where silicon can actually run.
本週,NVIDIA 展示了代理式 AI 如何改變整體技術架構:相較於最快的 x86 處理器,Vera CPU 每瓦可支援的 AI 代理數量最高達到 1.6 倍;同時,Vera Rubin NVL72 登上 Oracle Cloud,並規劃打造數 GW 規模的超級叢集。另一方面,據報 NVIDIA 正考慮將既有電信機房改裝成小型推論據點,因為供電、許可審批與地方反對聲浪,如今都會影響晶片究竟能在哪裡真正投入運作。
The pattern is clear. Training still needs dense GPU clusters. Inference and agents need balanced CPU–GPU systems, flexible form factors, and creative use of existing power. Bigger racks alone won't win if watts sit idle between tool calls.
趨勢已經很清楚:模型訓練仍需要高密度的 GPU 叢集;推論與 AI 代理則需要均衡的 CPU 與 GPU 系統、靈活的硬體配置,以及更有創意地運用既有電力。如果系統在工具呼叫之間閒置、未能有效利用電力,光靠更大的機櫃並不會勝出。
At NexCore.TW Corporation, we help teams design AI/GPU server infrastructure that matches the workload—not just the headline chip. Happy to compare notes if you're planning the next wave of capacity.
在 NexCore.TW Corporation,我們協助團隊設計符合實際工作負載的 AI/GPU 伺服器基礎架構,而不只是追逐最受矚目的晶片。如果你正在規劃下一波算力擴充,歡迎一起交流。