AI 晶片與算力

模型背後是晶片、機房與電網。這個主題收集全站談算力的文章：NVIDIA 與其他晶片的動向、資料中心的建設與成本，以及算力供不應求對產品團隊的實際影響。

文章大致圍繞三個問題：算力從哪裡來、受什麼限制（電力、併網、供應鏈）；大型實驗室如何用自研晶片、多源供應與租用機房組出算力組合；硬體層面的變化（晶圓級推論、機密運算）對推理成本與安全的意義。

必讀精選抓出這條主線；時間線可以按年份回看整個算力競賽的推進。

Behind every model are chips, buildings, and the power grid. This hub gathers the compute coverage on this site: NVIDIA and the wider silicon picture, data center buildouts and their costs, and what compute scarcity means for teams shipping AI products.

The posts cluster around three questions: where compute comes from and what constrains it (power, grid interconnection, supply chains); how major labs combine custom silicon, multi-vendor supply, and rented data centers; and what hardware-level shifts such as wafer-scale inference and confidential computing mean for cost and security.

The must-read picks lay out that throughline; the timeline lets you follow the buildout year by year.