Series
A close look at how computing, energy, and networks shape the foundations of AI competition.
Infrastructure
From final training runs to HBM, advanced packaging, networking, depreciation, utilisation, inference scheduling and customer total cost of ownership, this article maps where AI spending goes and why a lower unit price does not guarantee lower total expenditure.
From final training runs to HBM, advanced packaging, networking, depreciation, utilisation, inference scheduling and customer total cost of ownership, this article maps where AI spending goes and why a lower unit price does not guarantee lower total expenditure.
Read articleStarting from the difference between electricity and capacity, we track how power generation, transmission, grid connection, cooling and computing power scheduling jointly determine whether the model can continue to operate.
Read article机柜内每颗 GPU 有 1.8 TB/s 的互联带宽，跨出机柜的单端口速率是 800 Gb/s。从 scale-up、scale-out 到 scale-across，三层网络的落差决定了并行策略承受什么压力、以太网在 UEC 1.0 之后解决什么问题，以及跨数据中心训练能扩展到多远。
Read articleGoogle 一个月处理的 token 已是全人类公开文本存量的十倍以上。这个比较说明「数据不够」这四个字被混用了。数据约束分为存量、许可、生产、搬运四层：可训练文本按中位估计 2028 年见顶，但许可层的天花板已经先到；合成数据只在有验证器的地方成立；而在物理层，HDD 产能按日历年售罄，存力开始需要像电力一样提前锁定。
Read article