AI 需求正在以超出可用算力的速度上涨，尽管从业者在“tokenmaxxing”中疯狂消耗模型 token 来炫耀 AI 采用速度。3 月 Anthropic 在高峰时段开始限流并调整订阅方案，4 月其服务每天约有 30 分钟中断，OpenAI 在4 月突然下线了 Sora 视频生成，GitHub（微软旗下）也在4 月 20 日停止接受其编程机器人的新订阅。
需求压力转化为基础设施资本竞赛：4 月20 日 Anthropic 与亚马逊签署1000亿美元合作协议以锁定最多5 吉瓦服务器算力（约五分之一计划于今年底上线），4 月24 日又披露谷歌将投入40亿美元；与此同时，五大云巨头的合并资本开支已超过7500亿美元并增长三倍。尽管如此，甲骨文、亚马逊和Alphabet 今年已累计举债逾1000亿美元，但美国去年仍有1560亿美元数据中心项目因地方反对与诉讼被阻滞或延误，十余州在讨论20 兆瓦上限等限制，Meta 与 Microsoft 也分别推出10%与约7%裁员以腾出现金流。
硬件端瓶颈已全面加深：NVIDIA H100 芯片租赁价格自11 月起已上涨约30%，亚马逊称 Trainium2 几乎售罄且 Trainium4 已大量预订，SK Hynix、三星与美光三家主流供应商的2026年 HBM 供应几乎告罄，且未来三年需求预计仍将持续超出供给。摩根士丹利估计，agentic AI 每1个GPU需1个CPU，而聊天式系统仅1比12，推动 Intel 市值在六个月内翻倍以上；供应商资本支出仅增长一半且仅为云厂商的不足三分之一，台积电先进制程产能已满负荷并将2026年资本支出增至550亿美元（同比+34%，预计2027年650亿美元），而马斯克筹建的“Terafab”预计最早也要到2028年才可能小规模投产。
AI demand is increasing faster than available compute, even as teams compete in “tokenmaxxing” by burning through model tokens. In March Anthropic began throttling heavy users and later changed subscription terms, while in April it logged about 30 minutes of daily outages, OpenAI abruptly shut down Sora, and GitHub, owned by Microsoft, stopped accepting new subscriptions for its coding bot on April 20.
Demand pressure has become an infrastructure capital race: on April 20 Anthropic announced a $100 billion partnership with Amazon to secure up to five gigawatts of server capacity, with about one-fifth expected by year-end, and on April 24 it said Google would contribute $40 billion. The five hyperscalers now have combined capex above $750 billion (tripled), Alphabet, Amazon and Oracle have already raised over $100 billion in debt, yet local opposition still blocks rollout as $156 billion in U.S. projects were delayed or blocked last year and more than ten states consider limits like 20 MW caps, while Meta and Microsoft announced 10% and about 7% workforce reductions.
The hardware bottleneck has spread across the stack: Nvidia’s H100 rental price is up about 30% since November, Amazon said Trainium2 access is nearly sold out and much Trainium4 capacity is pre-booked, and HBM suppliers SK Hynix, Samsung and Micron say most 2026 supply is sold out with demand projected to exceed supply for at least three years. Morgan Stanley estimates agentic AI needs one CPU per GPU versus one-to-12 for chatbot systems, reviving CPU demand and helping Intel’s market cap more than double in six months; yet hardware suppliers have lifted capex only halfway while cloud giants tripled theirs, TSMC is at full run-rate with 2026 capex raised to $55 billion (+34%) and expected $65 billion in 2027, and Elon Musk’s planned Terafab is still years from meaningful scale.
Source: Silicon ceiling
Dateline: The Economist May 2nd 2026