Insights · 10 November 2026 · 2026年11月10日
Viewpoint · 觀點

The next jump in AI capacity may not come from buying more chips.

AI 算力的下一次躍進,可能不是來自買進更多晶片,而是來自更有效地運用電力、機櫃與晶片。

The next jump in AI capacity may not come from buying more chips. It may come from using your power, racks, and silicon more effectively.

AI 算力的下一次躍進,可能不是來自買進更多晶片,而是來自更有效地運用電力、機櫃與晶片。

At NVIDIA AI Day Seoul, NVIDIA showed that its DSX MaxLPS software, which reallocates power across racks in real time, helped cloud partner Nscale fit about 37% more GPUs and generate 50% more tokens on GB300 racks within the same power budget.

在 NVIDIA AI Day 首爾活動上,NVIDIA 展示了 DSX MaxLPS 軟體。這套軟體會即時在各機櫃之間重新分配電力,讓雲端合作夥伴 Nscale 的 GB300 機櫃在相同電力預算下,多裝了約 37% 的 GPU,產出的 token 也多了 50%。

Google also introduced TPU 8i Max and TPU 8tx. 8i Max packages two TPU 8i chips together, doubling compute, SRAM, and HBM bandwidth to address the memory wall. 8tx adds hardware-level security for confidential inference.

Google 也推出 TPU 8i Max 與 TPU 8tx。8i Max 將兩顆 TPU 8i 晶片封裝在一起,運算力、SRAM 與 HBM 頻寬都加倍,用來應對記憶體瓶頸;8tx 則提供硬體層級的安全防護,用於機密推論。

Separately, NVIDIA, SK Telecom, and Rebellions signed an MOU to connect NVIDIA GPUs with Rebellions NPUs via NVLink Fusion in SK Telecom’s Seoul data centers.

另外,NVIDIA、SK Telecom 與 Rebellions 簽署合作備忘錄,計畫在 SK Telecom 的首爾資料中心,透過 NVLink Fusion 將 NVIDIA GPU 與 Rebellions 的 NPU 串接起來。

My takeaway for NexCore.TW Corporation: the constraint is shifting from how many GPUs you can get, to how much useful work you can extract per watt, per rack, and per dollar. Power, cooling, memory, and the full system design matter as much as the server spec sheet.

我對 NexCore.TW Corporation 的啟示是:限制正從「能拿到多少 GPU」,轉向「每瓦、每機櫃、每一塊錢能產出多少有效工作」。電力、散熱、記憶體與整體系統設計,重要性已不亞於伺服器規格表。

Where is your AI capacity really constrained today?

你的 AI 算力,現在真正卡在哪裡?

Originally shared on LinkedIn