AI 算力AI COMPUTE

训练与推理的算力核心 — 价值高度集中,但护城河正被定制 ASIC 蚕食The compute core for training and inference — value is highly concentrated, but the moat is being eroded by custom ASICs

整个 AI 故事的隐藏主角 — GPU 是消耗品,电力是天花板。2030 前最稀缺的不是芯片而是接入电网的电,以及把电送进机柜的设备链The hidden protagonist of the entire AI story — GPUs are consumables, power is the ceiling. Before 2030, the scarcest resource is not chips but grid-connected electricity and the equipment chain that delivers it to the rack

GPU 功耗从风冷飙到液冷拐点 — 单机柜上百千瓦,空气压不住了。散热从'附属件'变成'决定能不能上架'的硬约束GPU power density has crossed the inflection point from air to liquid cooling — hundreds of kilowatts per rack exceed air capacity. Thermal management has shifted from an accessory to a hard constraint on deployability

算力越大,机器之间'对话'越贵 — 数据中心内部的光模块/光互连用量随 GPU 数量超线性增长,是最直接的横向受益但常被忽视The more compute, the more expensive machine-to-machine communication — optical transceiver and interconnect consumption inside data centers grows super-linearly with GPU count, a direct and often overlooked beneficiary

几万张 GPU 要当成'一台机器'用,靠的是超高速交换网络 — 交换机/交换芯片/NVLink 是 scale-out 的隐形主干,直接吃 GPU 集群规模红利Tens of thousands of GPUs must function as a single machine, enabled by ultra-high-speed switching networks — switches, switch ASICs, and NVLink are the invisible backbone of scale-out, directly capturing GPU cluster scale dividends

把芯片+内存+网络+散热装成可交付的 AI 工厂 — 集成商吃规模但毛利薄Assembling chips, memory, networking, and cooling into a deliverable AI factory — integrators capture volume but at thin margins