人工智能面临算力供应紧张，token需求在加速增长：OpenRouter显示，1月到3月间每周token消耗增加了4倍，部分原因是编码工具使用上升。各公司已开始配给资源，Anthropic调整条款抑制高峰时段重度使用，Amazon称“产能约束”限制了增长，OpenAI财务负责人表示缺乏足够算力，已暂停视频生成模型，这已开始改变利润分配和使用激励。

计算扩张短期内难以迅速实现：在美国，地方反对新数据中心使建设放缓，而变压器、开关设备和燃气轮机短缺会导致项目延迟2到5年，使处理器成为最紧张瓶颈。硬件供应商在资源分配中影响很大，软件公司因而转向自研芯片，自研芯片成本大约只有买Nvidia芯片的一半，但能形成大规模可行方案的只有谷歌，而且它在十年多前就开始了；取代TSMC更难，Intel和Samsung仍落后；马斯克提出的Terafab对标TSMC估算高达5万亿美元到13万亿美元。

AI繁荣长期建立在查询成本持续下降的前提上：推理价格在一年内已下降五到十倍，而印度低价订阅等促销掩盖了企业为维持这些价格付出的高现金消耗。OpenAI和Anthropic预计未来几年将亏损数十亿美元，并在两家都准备IPO的背景下更想证明未来盈利；随着AI从技术领域向更广泛行业扩展，token需求可能按数量级增长，企业将不得不更关注“如何高效用AI”，而不只是“是否使用AI”。

Artificial intelligence is facing a compute supply squeeze as token demand accelerates; OpenRouter reports weekly token use quadrupled from January to March, partly from heavier use of coding tools, and firms are already rationing. Anthropic has tightened terms to curb peak-hour heavy use, Amazon cited capacity constraints on growth, and OpenAI’s finance chief said limited compute means the company is not pursuing every opportunity, with the model-maker already dropping its video-generation model, a shift that is beginning to alter profit allocation and usage incentives.

Supply expansion is not coming quickly: in the United States, local opposition is slowing data-center builds, while shortages of transformers, switchgear and gas turbines can delay projects by two to five years, making processors the tightest bottleneck. Hardware suppliers dominate access despite denying favoritism, so software firms are shifting to custom chips that can cost about half as much as Nvidia, but only Google has produced one at scale after a decade-plus effort, and replacing TSMC remains difficult as rivals Intel and Samsung lag; Musk’s proposed Terafab to rival TSMC is estimated at an extreme $5 trillion to $13 trillion.

The AI boom was built on the assumption that inference keeps getting cheaper, and inference prices have indeed fallen five- to ten-fold in a year, masking heavy cash burn behind promotions such as cut-price subscriptions in India. OpenAI and Anthropic are each expected to lose billions in coming years, and as AI use spreads beyond tech into broader sectors, token demand may grow by orders of magnitude, forcing firms to prioritise efficient use over mere adoption.

Source: Compute says no

Subtitle: most valuable company, remain scarce. The squeeze extends to other types of silicon, too, including memory chips and central processing units ( CPU s). Few of these constraints will ease any time soon. Supply chains take years to expand and hardware-makers are still investing more cautiously than the hyperscalers they supply. When hardware is expensive the size of your balance-sheet matters more than ever (see Business section). Whichever part of the supply chain you look at, only a handful of firms have the financial muscle and bargaining power to lock up the hardware they need. This year the five data-centre "hyperscalers"-Amazon, Google, Meta, Microsoft and Oracle-will together shell out more than $750bn on capital expenditure. Open AI and Anthropic have announced hundreds of billions of dollars in partnerships and investments. Nvidia is said to have bought most of the memory it will need in 2026 and part of 2027 well in advance. It has also invested across a range of tech firms to shore up its supply chain. The greatest profits will be found at chokepoints. The AI boom has especially benefited Nvidia and TSMC , the Taiwanese manufacturer that makes almost all of the most advanced chips. Chip manufacturers' pricing power has become as enormous as their transistors are tiny. Nvidia's gross margin is about 75%, up from 60% in 2019. TSMC 's gross margin is above 60%, roughly twice that of many other contract manufacturers.

Dateline: The Economist May 2nd 2026