THIS SPRING McKinsey made what seemed to be an extraordinarily bullish forecast of capital spending on the chips, data centres and energy to produce artificial intelligence (AI): $5.2trn worldwide in the next five years. Less than six months on, the consultancy is considering upping that estimate. Announcements in America suggest investing in generative-AI infrastructure is reaching fever pitch.
今年春天，麦肯锡发布了一份看似极度乐观的预测：未来五年，全球用于生产人工智能（AI）的芯片、数据中心和能源的资本支出将达到5.2万亿美元。不到六个月后，这家咨询公司正考虑上调这一数字。美国的一系列公告表明，对生成式人工智能基础设施的投资已接近狂热。
Such expenditure, accentuated by staggering data-centre deals unveiled by firms such as OpenAI, Nvidia and Oracle, is aimed at increasing the computing power its protagonists believe is needed to supply generative AI. But demand—especially the revenue-yielding sort—does not yet match the hype. Though consumers’ use of chatbots is rising, McKinsey has found that the success rate of AI pilot projects in firms it has canvassed is less than 15%, says Pankaj Sachdeva, a partner at the firm. He predicts an era of “lumpiness” between supply and demand that could last for years.
这笔支出因OpenAI、英伟达和甲骨文等公司公布的一系列惊人的数据中心交易而更加显眼，其目的在于增加算力——其推动者认为，这是供给生成式AI所必需的。然而，需求，尤其是能带来收入的需求，尚未与炒作相匹配。尽管消费者对聊天机器人的使用在上升，但麦肯锡合伙人潘卡吉·萨赫德瓦（Pankaj Sachdeva）表示，该公司调研的企业中，AI试点项目的成功率不足15%。他预测，供需之间将出现“块状”失衡的时代，可能持续数年。
The strength of demand for generative AI may be the most critical factor determining whether or not this infrastructure boom ends in a bust. But three novel aspects of the data-centre building frenzy add to the uncertainty: the centres’ remote locations, the non-public firms financing them and the weak credit quality of some borrowers. This trifecta reminds some sceptics of the last great infrastructure debacle: the telecoms boom of the late 1990s. Yet plenty of others are holding their noses and diving in.
生成式人工智能的需求强度，或许是决定这场基础设施狂欢最终会否以泡沫破裂收场的最关键变量。然而，数据中心建设热潮中出现了三个前所未有的新变数：选址愈发偏远、出资方多为非上市公司、部分借款人信用资质薄弱。这三重特征令一些怀疑者联想到上世纪90年代末电信泡沫那场基础设施灾难。尽管如此，仍有大量投资者屏住呼吸、纵身跃入。
Geography is the most tangible novelty. The new AI data centres are springing up in the middle of nowhere rather than in established clusters close to big sources of demand and interconnection hubs, such as northern Virginia. OpenAI and its partners, Oracle and SoftBank, have begun the first phase of Stargate (pictured below), a $500bn AI-infrastructure project announced by President Donald Trump in January, in a part of Texas with lots of wind and solar energy—and empty space. North Dakota and New Mexico have similar attractions.
最显而易见的新变化在于选址。新建的人工智能数据中心不再扎堆于北弗吉尼亚这类靠近主要需求枢纽和互联节点的成熟集群，而是拔地而起于荒无人烟之地。OpenAI及其合作伙伴甲骨文与软银，已在今年1月由唐纳德·特朗普总统宣布的5000亿美元“星际之门”AI基础设施项目中启动首期工程，地点选在了得克萨斯州一处风能、太阳能丰富且空地广袤的区域。北达科他州和新墨西哥州也具备类似的吸引力。
Such fresh locations solve a power problem: many existing clusters lack enough surplus energy for training the latest large language models (LLMs) developed by labs such as OpenAI. But isolation introduces risks for property investors that may not be adequately reflected in the returns, says Gautam Bhandari of I Squared Capital, an infrastructure-focused private-equity firm.
这些新地点解决了一个能源难题：许多既有集群缺乏足够剩余电力，来训练OpenAI等实验室开发的最新大型语言模型（LLM）。然而，专注于基础设施的私募股权公司I Squared Capital的Gautam Bhandari指出，偏远带来的隔离风险可能并未充分体现在地产投资者的回报中。
Data centres are usually financed over decades, but those on the cutting edge of AI may become obsolete far quicker, Mr Bhandari says. That is partly due to technology: Nvidia, the dominant maker of AI-related graphics processing units (GPUs), relentlessly improves the efficiency of its chips, which may require regular data-centre upgrades, such as new cooling systems. With land so readily available, a rival data-centre builder with a better, cheaper design can easily set up elsewhere. That raises the likelihood of stranded assets.
班达里表示，数据中心通常按几十年的周期融资，但处于AI前沿的设施可能更快就会被淘汰。部分原因在于技术：AI图形处理器（GPU）的主导制造商英伟达不断大幅提升芯片效率，这可能要求数据中心定期升级，例如更换冷却系统。由于土地供应充足，拥有更优、更便宜设计的竞争者可以轻松另起炉灶，从而提高了资产被废弃的概率。
The boom’s sources of finance are also relatively new. Until recently, big suppliers of capital for data centres in prime locations were stockmarket investors, via real-estate investment trusts (REITs). They were most comfortable when a data centre’s power consumption was far less than 100 megawatts. But in the AI era, appetites are measured in gigawatts (GW), and the costs can run to $50bn per GW.
此次热潮的资金来源也颇为新颖。此前，黄金地段数据中心的主要资金供给方是股票市场投资者，通过房地产投资信托（REITs）投入。他们通常只愿接纳功耗远低于100兆瓦的项目。但在AI时代，胃口已跃升至吉瓦（GW）级别，每吉瓦成本可高达500亿美元。
As demand for capital has rocketed, REITS have been constrained by their own borrowing capacity, says David Guarino of Green Street, a commercial-property research firm. Their place is being taken by private-credit firms (some of which have acquired ex-REITs) and sovereign-wealth funds, as well as banks. These are sophisticated lenders with deep pools of capital, used to the sort of project finance involved in AI infrastructure lending. But their participation shifts risk from equity to debt markets, putting the banking system more squarely in the line of fire if defaults rise.
商业地产研究机构Green Street的戴维·瓜里诺指出，随着资本需求飙升，REITs因自身借贷能力受限而力不从心。取而代之的，是私人信贷公司（其中一些已收购原REIT资产）、主权财富基金以及银行。这些机构资金雄厚、经验丰富，熟稔AI基建项目融资。然而，它们的参与把风险从股权市场转移至债务市场，一旦违约率上升，银行体系将更直接地暴露在火线之下。
And the risk of default is raised by the dubious creditworthiness of some firms at the heart of the building boom. This was not a big concern when cash-rich cloud giants, such as Amazon, Microsoft and Google, were recipients of much of the finance. They are “the best tenants in the world”, says Mr Guarino. But more recently, AI labs like OpenAI and “neocloud” firms that rent out GPUs have entered the fray, increasing the quantity—but decreasing the credit quality—of those involved. The more of them there are, the more they face competition, pressure on returns and uncertainty about their long-term viability.
此外，建设热潮核心企业的信用资质参差不齐，也抬高了违约风险。此前，资金大多流向亚马逊、微软、谷歌等现金流充沛的云巨头，这并不算大问题——它们被瓜里诺称为“全球最优质的租户”。但近来，OpenAI等AI实验室以及出租GPU的“新云”公司纷纷入场，虽扩大了参与主体数量，却拉低了整体信用质量。参与者越多，彼此竞争越激烈，回报承压，长期生存能力也更不确定。
It is not just lenders who worry about these neophytes. Utilities, conservative by nature, may also think twice about signing long-term energy contracts with them. “You do not know which of these players will be around in five, ten or 15 years’ time,” says Mr Sachdeva. In response, he says, insurance policies, securitisations and the like are being designed to mitigate the risks. Likewise, tech giants such as Nvidia are pitching in with a web of vendor financings and crossinvestments that could also reassure counterparties. But if the worst happens, such incestuousness will increase the vulnerability of the AI ecosystem as a whole.
担忧这些新手的不仅仅是贷款方。天性保守的公用事业公司，也可能对与它们签订长期能源合同三思而行。“你无法判断这些企业五年、十年或十五年后是否还存在。”萨赫德瓦说。为此，据他介绍，业内正在设计保险、证券化等工具来缓释风险；同样，英伟达等科技巨头也通过错综复杂的卖方融资和交叉投资来安抚交易对手。然而，一旦最坏情况发生，这种“圈内互保”反而会加剧整个人工智能生态系统的脆弱性。
Such interlinkages are adding to concerns that an infrastructure bubble is forming, similar to the laying of fibre-optics and undersea cables in the early days of the internet. Andrew Odlyzko of the University of Minnesota, a historian of infrastructure manias from 19th-century railways onwards, used to downplay the economic impact of an AI bust. He believed that if a few tech giants were forced to write off their investments in data centres, that would only torch a few years’ worth of profits. Now, he says, he is “much more alarmed” because of the number of firms making big investment pledges.
这种盘根错节的关联，进一步加剧了人们对“基础设施泡沫”正在形成的担忧——情形与互联网早期铺设光纤和海底电缆时相似。明尼苏达大学的安德鲁·奥德莱兹科（Andrew Odlyzko）研究自19世纪铁路以来的基础设施狂热史，他曾对AI泡沫破裂的经济冲击不以为意，认为即便少数科技巨头被迫注销数据中心投资，也不过是烧掉几年的利润。如今他却表示，由于大批企业正许下巨额投资承诺，他“深感忧虑”。
He sees parallels with the late-1990s investment frenzy that culminated in the dotcom crash. Proposed deals such as Nvidia’s potential investment of $100bn in OpenAI if it buys up to 10GW of GPUs remind him of Nortel’s vendor-financing arrangements with buyers of its equipment during the telecoms bubble. Others, however, play up the differences. Nick Del Deo of MoffettNathanson, an equityresearch firm, says that in the telecoms boom cable was laid without customers in place. Today, data centres are built only when counterparties sign contracts for it—though the contracts’ details will be critical in determining whether returns justify the risk.
奥德莱兹科看到了与上世纪90年代末投资狂潮的相似之处，那场狂潮最终以互联网泡沫破裂收场。英伟达潜在向OpenAI投资1000亿美元、后者同步采购至多10吉瓦GPU的拟议交易，让他想起北电（Nortel）在电信泡沫期间向设备买家提供的卖方融资安排。然而，也有人强调两者差异。股票研究机构MoffettNathanson的尼克·德尔·迪奥（Nick Del Deo）指出，当年电信繁荣时，光缆是在并没有客户的情况下铺设的；如今，数据中心只有在交易对手签约后才会动工——不过，合同细节将至关重要，决定回报能否配得上所承担的风险。
For now, the potential rewards are so tantalising that money is pouring in, he says. Cheerleaders such as Sam Altman, OpenAI’s boss, argue that the risks of underbuilding are at least as serious as those of overbuilding, because of the long-term economic potential of generative AI. It may be that even if there is a surplus of capacity in the most advanced data centres, it can be absorbed by running, rather than training, LLMs, Mr Sachdeva says. But that comes back to the question of when demand for generative-AI chatbots and applications catches up with the ambitions of those supplying them. That is the most bewitching uncertainty of all.
德尔·迪奥说，目前潜在回报太过诱人，资金正蜂拥而入。OpenAI首席执行官萨姆·奥特曼（Sam Altman）等“啦啦队”认为，由于生成式AI的长期经济潜力，建设不足的风险至少与建设过剩一样严重。萨赫德瓦也提到，即便最先进的数据中心出现容量过剩，也可能通过“运行”而非“训练”大语言模型来消化。但这终究回到那个最迷人、也最不确定的问题：人们对生成式AI聊天机器人及应用的需求，究竟何时才能追上供给方的宏大野心？