一张 GPU 跑一小时，电费大概一毛到三毛多美元，租金要两美元往上。所以算力生意的账都在合同里，价格锁多久，卡闲下来谁兜底。最后用 Abilene 把整个系列串一遍。

风冷输在物理上，同样带走 100 kW 的热，要的空气体积大约是水的 3500 倍。机房费不费水，要看哪个环路是开着的，闭环的机柜后面照样可能接着一座天天蒸发的冷却塔。

NVLink 管机柜里面，InfiniBand 和以太网管机柜之间，再往外是园区之间的光纤。训练跑多快，看最慢的那张卡和最挤的那根线，所以网络拓扑得跟楼板图一起画。

断电之后，先是电容顶几毫秒，再是电池，最后柴油发电机接手。越靠前的一棒反应越快，存的能量越少。真出事多半是在交接那一下，比如所有电源共用的那一根母线。

从 345 kV 的高压线降到芯片上的 1 V，电压一路降，电流一路涨，发热跟着电流的平方涨。800 V 直流进机柜，其实就是把输电网高压低流的老办法搬进机房。

并网要排长队，按 SemiAnalysis 的估算，排队名单里大部分还没有实际的地块撑着。电费在 GPU 的账里只占个零头，所以大家宁可烧效率低的燃气机组，也要早一个月开机。

prefill 拼算力，decode 拼内存带宽，context 每长一倍，一张卡能同时服务的人就少一半。模型怎么服务用户，决定了机房要堆多少内存和带宽。

一个 token 从机房里出来，背后有三根管子一直在跑。电进来，热出去，数据在几万张 GPU 之间来回传。机房能干多少活，看三根里最细的那根，现在最细的是电。

Running one GPU for an hour costs roughly ten to forty cents of electricity and rents for two dollars and up. So the money in compute lives in the contracts: how long the price is locked, and who backstops the idle hours. Then Abilene ties the whole series together.

Air cooling lost on physics: moving 100 kW of heat takes roughly 3,500 times more air than water by volume. And whether a data center drinks water depends on which loop is open, since a sealed rack loop can still feed an evaporating tower.

NVLink inside the rack, InfiniBand or Ethernet between racks, fiber between campuses. A training run moves at the pace of its slowest card and most crowded link, so the network topology gets drawn together with the floor plan.

When the grid drops, capacitors catch the first milliseconds, batteries carry the next minutes and generators take the long haul: the earlier the runner, the faster it reacts and the less energy it holds. In the failures the course covers, the baton drops at the handoffs, like the one bus every power path shares.

Power arrives at hundreds of kilovolts and leaves at about one volt at the chip. At every step down the voltage falls, the current climbs and resistive heat grows with its square. 800 V DC in the rack is the grid's old high-voltage trick moved indoors.

Grid connections are slow, and by SemiAnalysis's count most of the queue isn't backed by real sites. Electricity is a rounding error next to the GPUs, so builders burn gas in inefficient turbines just to switch on a month earlier.

Prefill is limited by compute, decode by memory bandwidth, and every doubling of context halves how many users a GPU can serve at once. How a model serves people decides how much memory and bandwidth the building has to hold.

Every token that leaves a data center rides on three pipes: power coming in, heat going out, and data moving between tens of thousands of GPUs. The building can only do as much work as its narrowest pipe allows, and today that pipe is power.