我理解中美国 AI 对电力的需求
美国 AI 对电力的核心需求,我个人看法是大规模、稳定、连续、可接入的电力供应。
AI 数据中心的耗电强度远高于传统数据中心。以前的数据中心主要跑云计算、存储、网页、APP 服务,用电已经很大,现在 AI 数据中心跑的是 GPU 集群,训练模型、推理服务、视频生成、智能体应用,都会让服务器长时间处在高负载状态。GPU 越多,功率越高,散热越难,��个机房对电力的要求也越高。
一个大型 AI 数据中心的用电规模可以达到几百兆瓦,未来头部项目甚至可能接近 GW 级别。这个概念很重要,因为这已经接近一座小城市的用电量。AI 公司扩张算力,表面上是在买 GPU、建机房、铺服务器,底层其实是在消耗更多电力。
所以 AI 对电力的要求很特殊。它要的是稳定电、连续电、能长期锁定的电。数据中心需要 24 小时运行,不能只在白天运行,也不能只在电价便宜的时候运行。训练模型需要稳定算力,推理服务也要随时响应用户请求,电力供应一旦不稳定,就会影响服务器运行、客户服务、云业务和整体交付能力。
从能源类型来看,核电最符合 AI 数据中心的长期需求。核电稳定、低碳、容量因子高,可以持续供电,也适合大型科技公司签长期购电协议。微软、亚马逊、Meta 这些公司愿意长期签订核电资源,本质上就是在提前锁定未来的稳定算力底座。
但核电的问题也很现实,建设周期太长。新核电站从审批、建设到并网,需要很多年。所以短期内,美国 AI 用电缺口很难完全靠新核电解决。能在未来几年更快补上的电源,大概率还是天然气。天然气电厂建设周期相对更短,调峰能力更强,能够更快满足数据中心的用电增长。
所以,美国 AI 用电的结构大概率会是核电提供长期稳定基荷,天然气承担中短期补电和调峰,风电、太阳能、储能作为补充,帮助降低综合电价和满足绿色能源目标。
美国现在电力真正紧张的地方,还在电网侧。很多地区并非完全没有发电能力,问题出在电送不到数据中心。输电线路不够,变压器不够,变电站不够,并网容量不够,���批和施工周期也很长。AI 公司可以很快采购 GPU、建设机房、部署服务器,但电网扩容没法按季度完成。
因此我个人判断未来几年,美国 AI 用电需求会推动三类资源的发展:
第一类是稳定电源,包括核电、天然气和部分水电。AI 数据中心需要长期稳定供电,稳定电源的价值会越来越高。
第二类是电网基础设施,包括输电线路、变压器、开关设备、变电站和电力工程。电能不能送到数据中心,会直接决定数据中心能不能落地。
第三类是数据中心内部电力系统,包括 UPS、配电系统、液冷、散热和能源管理。GPU 功率越高,机房内部的供电和散热压力越大,这些基础设施的重要性也会继续上升。
因此我觉得未来几年,美国 AI 电力主线会长期推高稳定电源、电网设备、电力工程和数据中心基础设施的价值。AI 竞争越激烈,电力资源的战略价值越高。
#Bitget ���了就是VIP!Crypto、美股、CFD,全球先机一站布局
Just some random notes about $AVGO earnings transcript
- Revenue target reiterated ($100B+ 2027, pretty sure markets wanted that to be raised this earning, hence the drop)
Remember $NVDA Jensen comments about $MRVL $1T company around networking/connectivity/interconnects?
- “So as the TPUs continue to accelerate, there’ll be pressure overall on margins.
But the connectivity side, the AI networking side of the business has very rich margins”
“Demand for … networking is simply insatiable”
Also very positive read through as well for the $LITE and the other players. But for TPU margins it goes down at scale, which is understandable.
- “they are placing orders in fairly huge demand, which basically gives us a lot more visibility.. runs all the way to 2028 right now”
positive read through on overall AI demand since it’s 2026 now… and orders are out in 2028
- The initial order for 1 gigawatt, which includes XPUs and our networking has been received and will start Delivery in the second half of 2027. for our other two customers, we expect shipments to begin late 2026 and accelerate into 2027.
$META custom AI program h2 2027 timelines
- “Our revenue, our content per gigawatt will increase. you start putting a lot, you start putting embedding CPU cores into the same XPUs and making those chips basically multi die with lots of hvm.”
Just for the GW modelers.
- “For OpenAI we have delivered silicon and we are on track for production late 2026”
OpenAI custom program timeline
- “If you ask about 27 or 28 that will continue to grow. We expect in fact 28 to be a substantial growth from what we are forecasting in 27.”
More about the demand ramp, go brrr
- “Google, that we expect a diversity of sources from them”
Mediatek (2454) primary beneficary, maybe $MRVL. Already expected though Google doesn’t sole source so they don’t get bottlenecked.
There’s quite a lot of AI demand visibility way until 2028, which is bullish on the AI sector as a whole. Regardless, Broadcom ends the week +0% lol.
TLDR: Strongly bullish AI demand, especially networking. Stocks don’t move in a straight line up, but demand curves 2026-> 2027 -> 2028.