关于AI基础设施的主流叙事,是一场井然有序的接力赛:先是GPU短缺,接着内存制约了吞吐量,然后CPU承压,最后存储也开始吃紧。
定价数据表明,这场接力确实存在,但节奏缓慢。有些组件在价格飙升之前先经历了暴跌。每一个瓶颈都会冻结下一个组件的供应链,但滞后时间长达数年,而且每一波浪潮都锁定了更高的成本基线。
2023年初的初始冲击来自GPU。ChatGPT发布后,买家将资本集中用于采购GPU,使得按需租用的Nvidia H100价格一度超过每小时9美元。¹
这种执念让传统计算领域陷入资源枯竭。2023年服务器出货量下降22%,跌至2018年水平以下,因为买家推迟了更新周期,以便为GPU腾出资金。² 内存制造商本已因疫情后产能过剩而遭受重创——该行业损失超过200亿美元,被迫削减晶圆产能高达40%——如今又失去了本可消化其库存的服务器需求。³
十八个月后,压力转移到了内存上。为了摆脱低迷并追逐AI利润,制造商将洁净室和光刻设备转向生产高带宽内存(HBM)。
HBM每GB容量消耗的晶圆产能大约是标准DDR5的三倍,因此每生产一比特HBM,就意味着减少约三比特的传统供应。企业级固态硬盘(SSD)合约价在单个季度内上涨了80%,而美光报告称,动态随机存取存储器(DRAM)价格环比涨幅在60%出头的区间。⁴
到2025年底,智能体开始挤压服务器CPU的空间。训练集群的CPU与GPU配比为一比八。智能体工作流则颠覆了这一比例:自主系统将计算周期用于编译代码、调用工具和管理状态,将配比推向1:1。⁵ 英特尔报告称,服务器CPU平均售价(ASPs)同比上涨27%,而出货量却在下降,并指出Xeon需求存在数十亿美元的缺口。⁶
到2026年,短缺已蔓延至大容量存储领域。由于每TB 150美元的高速闪存价格过高,云架构师们沿着技术阶梯向下退,转向传统、较慢的机械硬盘(HDD)来存放海量训练数据湖。西部数据(Western Digital)和希捷(Seagate)已确认其2026年全年近线存储产能已全部售罄。7
走出服务器机箱之外,情况同样一目了然。增长最快的成本是数据中心的水泥地板和加固墙体。
数据中心如今每吉瓦(GW)造价高达200亿美元以上($20b/GW),其中电气系统消耗了预算的一半。8 建筑成本已翻了两番,达到每平方英尺1,033美元,9 且不含土地成本。在室外,发电机升压(GSU)变压器的平均交货周期接近三年,而GE Vernova和西门子能源(Siemens Energy)的涡轮机产能已售罄至2029年,订单簿甚至排到了2031年。10
正如西门子的Barry Powell对这两难困境的观察:
“做也是死,不做也是死:如果你建得不够多,就会因失去市场份额而受到指责。如果你建得太多,又会因固定成本过高而受到指责。我们明白,某个时点可能会出现泡沫,而我们正在争分夺秒地尽快收回投资。”11
这就是物理硬件世界中的牛鞭效应。12 当一条价值链面临长达数年的制造延迟时,下游需求的突然冲击会向上游放大为大规模、滞后的过度反应。在一个瓶颈处缓解压力,会以可预测的延迟将其推向下一个组件。当这波浪潮最终退去时,交货周期长的资本品恰恰是暴露在产能过剩风险之下的环节。
超过20亿美元的国内变压器扩产项目、下一代300层NAND晶圆厂以及新的涡轮机生产线将于2027年和2028年投产。如果终端用户的软件收入无法跟上每吉瓦200亿美元设施的建设节奏,资本支出将面临经典的“鞭梢效应”冲击。
-
Silicon Data H100与B200 GPU租赁价格指数(2026)及NVIDIA硬件披露。 ↩︎
-
Omdia数据中心服务器追踪报告,2023年12月。2023年服务器出货量同比下降22%,降至不足1100万台,比2018年水平低5%。 ↩︎
-
三星电子2023财年业绩(半导体部门营业亏损14.88万亿韩元)、SK海力士2023财年业绩(营业亏损7.73万亿韩元)、美光科技2023财年10-K年报(净亏损58.3亿美元)以及西部数据2023财年业绩(亏损17亿美元)。↩︎
-
TrendForce,《AI智能体热潮引发企业级SSD供应紧张》,2026年6月11日;美光科技,截至2026年5月28日的季度10-Q报告,经营业绩部分。↩︎
-
英特尔公司(陈立武)与AMD(苏姿丰)2026年第一季度财报披露。↩︎
-
英特尔公司截至2026年3月28日的季度SEC 10-Q报告。↩︎
-
西部数据与希捷科技2025年第四季度/2026年第一季度财报电话会议。↩︎
-
iRecruit与数据中心建设成本基准报告(2026年)。高密度AI设施成本分析(每兆瓦2000万美元,即每吉瓦200亿美元)。↩︎
-
美国人口普查局与RSMeans专业关键任务设施建设成本指数。↩︎
-
Wood Mackenzie、《POWER》杂志及CNBC对GE Vernova燃气轮机积压订单的披露。↩︎
-
Brooke Sutherland,《AI焦虑促使供应商为数据中心繁荣破灭做准备》,《彭博工业力量》,2026年8月21日。↩︎
-
Tomasz Tunguz,《牛鞭效应与基础概率——影响初创企业的两大力量》,tomtunguz.com,2020年4月。另见《如何利用代理指标创造竞争优势》。↩︎
The popular narrative of AI infrastructure is a tidy relay race : first GPUs were scarce, then memory choked throughput, then CPUs took the strain, & finally storage started to bite.
The pricing data says the relay is real but slow. Some components plunged before they surged. Each bottleneck freezes the next component’s supply chain, but the lag runs in years, & every wave locks in a higher baseline cost.
The initial shock in early 2023 belonged to the GPU. When ChatGPT launched, buyers concentrated capital on procuring GPUs, sending on-demand Nvidia H100 rental rates past $9 an hour.1
That monomania starved conventional computing. Server unit shipments fell 22% in 2023 to below 2018 levels as buyers deferred refresh cycles to fund GPUs.2 Memory makers, already reeling from a post-pandemic glut that cost the industry more than $20b & forced wafer cuts of up to 40%, lost the server demand that would have absorbed their inventory.3
Eighteen months later, the pressure migrated to memory. To escape that slump & chase AI margins, manufacturers converted cleanrooms & lithography tools toward High Bandwidth Memory (HBM).
HBM consumes roughly three times the wafer capacity per gigabyte of standard DDR5, so every bit of HBM output removes about three bits of conventional supply. Enterprise solid state drive (SSD) contract prices rose 80% in a single quarter, while Micron reported dynamic random-access memory (DRAM) prices climbing in the low-60s percentage range quarter over quarter.4
By late 2025, agents squeezed server CPUs. Training clusters ran one central processing unit (CPU) to eight GPUs. Agentic workflows invert that : autonomous systems spend their cycles compiling code, calling tools, & managing state, pushing the ratio toward 1:1.5 Intel reported server CPU average selling prices (ASPs) rose 27% year over year against falling unit volumes, citing billions in unmet Xeon demand.6
By 2026, the shortage reached bulk storage. Priced out of high-speed flash at $150 per terabyte, cloud architects retreated down the technology ladder into traditional, slower hard disk drives (HDDs) for bulk training data lakes. Western Digital & Seagate confirmed their entire 2026 nearline production is sold out.7
Beyond the server chassis, the story is unambiguous. The fastest-rising cost is the data center’s concrete floor & the reinforced walls.
Data centers now cost upwards of $20b per gigawatt ($20b/GW), with electrical systems consuming half the budget.8 Construction costs have tripled to $1,033 per square foot, 9 excluding land. Outside, generator step-up (GSU) transformers average nearly three year lead times, while GE Vernova & Siemens Energy have sold out turbine production through 2029, with order books stretching to 2031.10
As Barry Powell of Siemens observed of the double bind :
“You’re damned if you do, damned if you don’t: If you don’t build enough, then you’re going to get dinged for losing some market share. And if you build too much, you’re going to get dinged for fixed costs. We understand at some point there could be a bubble and we’re racing to pay back the investments as quickly as possible.”11
This is the Bullwhip Effect in physical hardware.12 When a value chain suffers from multi-year manufacturing latency, sudden demand shocks downstream amplify into massive, lagged overreactions upstream. Relieving pressure at one bottleneck pushes it into the next component with a predictable delay. When the wave finally breaks, long-lead-time capital goods are the ones left exposed to overcapacity.
Over $2b in domestic transformer expansions, next-generation 300-layer NAND fabs, & new turbine production lines will deliver in 2027 & 2028. If end-user software revenues do not keep pace with $20b per gigawatt facilities, capital expenditure will face a classic crack of the whip.
-
Silicon Data H100 & B200 GPU Rental Price Indices (2026) & NVIDIA hardware disclosures. ↩︎
-
Omdia Data Center Server Tracker, December 2023. 2023 server shipments fell 22% year over year to under 11m units, 5% below 2018 levels. ↩︎
-
Samsung Electronics FY2023 results (14.88tn won semiconductor division operating loss), SK hynix FY2023 results (7.73tn won operating loss), Micron Technology FY2023 Form 10-K ($5.83b net loss), & Western Digital FY2023 results ($1.7b loss). ↩︎
-
TrendForce, “AI Agent Boom Triggers Enterprise SSD Supply Crunch,” June 11, 2026 ; Micron Technology, Form 10-Q for the quarterly period ended May 28, 2026, Results of Operations. ↩︎
-
Intel Corporation (Lip-Bu Tan) & AMD (Lisa Su) Q1 2026 Earnings Disclosures. ↩︎
-
Intel Corporation SEC Form 10-Q for the quarterly period ended March 28, 2026. ↩︎
-
Western Digital & Seagate Technology Q4 2025 / Q1 2026 Earnings Calls. ↩︎
-
iRecruit & Data Center Construction Cost Benchmark Reports (2026). High-density AI facility cost analysis ($20m/MW or $20b/GW). ↩︎
-
U.S. Census Bureau & RSMeans Construction Cost Indices for specialized mission-critical facilities. ↩︎
-
Wood Mackenzie, POWER Magazine, & CNBC GE Vernova turbine backlog disclosures. ↩︎
-
Brooke Sutherland, “AI Jitters Have Suppliers Preparing for Data Center Boom to Go Bust,” Bloomberg Industrial Strength, August 21, 2026. ↩︎
-
Tomasz Tunguz, “Bullwhip and Base Rates — The Two Major Forces Impacting Startups,” tomtunguz.com, April 2020. See also “How to Create Competitive Advantage with Proxy Metrics”. ↩︎