如今大规模的人工智能建设大多由不公开交易的工具融资。因此，投资者现在面临的问题不仅是人工智能是否会有回报，还包括如何判断它是否开始出轨。
Much of today’s massive AI build-out is being financed with instruments that do not trade. So, the question for investors now is not only whether AI will pay off, but also how they will know if it is starting to come off the rails.
过去一年里，Meta Platforms、Alphabet和Amazon等超大规模云服务商已通过公共债务市场筹集资金，以支持其创纪录的人工智能基础设施计划，但该行业的大部分资金正通过私人信贷基金、基础设施载体和合资企业，在闭门交易中进行。
While hyperscalers like Meta Platforms, Alphabet and Amazon have tapped public debt markets over the past year to fund their record-breaking AI infrastructure plans, much of the sector’s funding is taking place behind closed doors through private credit funds, infrastructure vehicles and joint ventures.
由于没有市场价格可供参考，这些资产通常由其管理人每季度使用预期现金流模型和与类似交易的比较来估值。这是正常做法，如果做得好，应能提供一个良好的初步近似值。但这只是近似值，而且往往比市场反应更慢。
With no market price to go on, these holdings are valued by their managers, usually quarterly, using models of expected cash flows and comparisons with similar deals. That is normal practice and, if done well, should provide a good first approximation. But it is only an approximation, and they tend to move more slowly than markets.
这一过程可能人为压低波动率，形成所谓的“平滑估值”，使私人信贷看起来在承担较低风险的同时获得更高回报。但现实中，风险通常仍然存在，只是没有被准确衡量。
This process can keep volatility artificially low, creating so-called “smoothed valuations,” which can make private credit look like it’s generating higher returns with lower risk. But, in reality, the risk is usually still there; it is just not being accurately measured.
这一问题日益重要，因为人工智能私人贷款的规模已不再是边缘性的。经济合作与发展组织（OECD）估计，2025年人工智能相关交易占全球私人信贷交易的34%，是五年平均水平的两倍。
This issue is becoming increasingly relevant because the scale of AI private lending is no longer marginal. The Organisation for Economic Co-operation and Development (OECD) reckons AI-related deals made up 34%of global private credit transactions in 2025, twice the five-year average.
英国央行还估计，2026年至2028年约7000亿美元的数据中心支出将由私人信贷融资。对于全球金融稳定委员会（FSB）在2024年底估计的1.5万亿至2万亿美元的市场规模而言，这是一笔巨大的资金。简而言之，人工智能正迅速成为该资产类别最大的单一主题。
The Bank of England also estimates that roughly $700 billion of data-centre spending over 2026-28 will be financed by private credit. That is a vast sum for a market that the Financial Stability Board (FSB), a global watchdog, estimated at $1.5 trillion to $2 trillion at the end of 2024. In short, AI is fast becoming the asset class’s biggest single theme.
按模型估值 私人贷款人在此领域使用的估值模型拥有巨大权力，因此必须稳健，尽管像所有模型一样，它们永远不会完美。
Mark to model Valuation models used by private lenders in this space have enormous power, and thus need to be robust, even if, like all models, they will never be perfect.
但要让估值模型发挥作用，其假设必须稳健，而人工智能债务交易中的一些假设可能很脆弱。
But for a valuation model to be useful, its assumptions need to be robust, and some of those in AI debt deals may be fragile.
首先是抵押品问题。人工智能数据中心的实体结构并不代表总资产成本或预期价值的大部分。据摩根大通资产管理公司称，建筑物通常仅占成本的约30%，芯片和设备则占剩余的70%。贷款人越来越接受这些芯片作为抵押品。
For one, there’s the issue of collateral. An AI data centre’s physical structure does not represent the majority of the cost or expected value of the total asset. The building typically accounts for only around 30% of the cost, according to JPMorgan Asset Management. The chips and equipment represent the remaining 70%.
然而，芯片价值远非稳定。据数据提供商Silicon Data称，自2025年初以来，主要云服务商出租英伟达H100芯片的每小时费率在各地区已下降25%至50%，而较小的专业提供商的费率更低。对于资产大部分价值的模型，如果采用过时的芯片价值，肯定不会有用。杞人忧天？
Lenders increasingly accept those chips as security. Yet chip values are far from stable. The hourly rate big cloud providers charge to rent Nvidia’s H100, the workhorse AI chip of the most recent cycle, has fallen by 25% to 50% across regions since early 2025, according to data provider Silicon Data, and rates at smaller specialist providers are lower still. A model that assumes outdated values for the bulk of an asset’s value certainly won’t be useful.
有人可以认为，这种担忧被夸大了：最大的人工智能项目背后的租户是世界上信用最好的公司之一。此外，私人信贷违约率仍然很低。
A fuss over nothing? You could make a case that this worry is overdone: The tenants behind the biggest AI projects are among the most creditworthy companies in the world.
而且，对于愿意持有贷款至到期的耐心投资者来说，私人信贷不频繁的估值或许是优点而非缺点。如果借款人继续付款，为什么市场每次波动就要下调贷款价值呢？
Moreover, private credit default rates remain low. And for a patient investor who will hold a loan to maturity, private credit’s infrequent valuations are arguably a feature, not a bug.
这种逻辑对封闭式、长期资本有效。但当投资者可以退出时，这种逻辑就不那么有效了。
If the borrower keeps paying, why mark the loan down every time markets wobble? That logic works for closed-end, long-horizon capital.
我们已经开始看到这种情况发生。蓝橡树资本（Blue Owl Capital OWL.N）是数据中心最大的贷款人之一，今年在面对大量赎回请求后，已两次将其私人信贷基金的赎回上限设定为5%。
It works less well when investors can leave. And we’re already starting to see that happen.
提供定期按公允价值赎回的半流动性基金，制造了金融稳定委员会（FSB）所称的“先行者激励”。假设一只基金将其贷款估值为每美元100美分，但今日买家只愿支付90美分。最先赎回的投资者按面值获得付款。亏损则落在留下来、直到估值追上现实的投资者身上。
Blue Owl Capital OWL.N, one of the largest lenders to data centres, twice capped redemptions from its private credit funds at 5% this year after heavy withdrawal requests. Semi-liquid funds that offer regular redemptions at stated asset value create what the FSB calls a “first-mover incentive”. Suppose a fund values its loans at 100 cents on the dollar, but a buyer today would pay only 90 cents. Investors who redeem first are paid at par.
不要成为最后知情者。这一问题现在更为重要，因为两种趋势正在汇聚。人工智能建设越来越依赖私人资本，而私人市场正在向普通储户开放，两者恰好同时发生。
The shortfall falls on those who stay once the marks catch up. Don’t be the last to know This issue matters more now because two trends are converging. The AI build-out is leaning ever harder on private capital at exactly the moment private markets are being opened up to ordinary savers.
在英国，最大的职场养老金提供商已根据大厦宫协议承诺，到2030年将其默认基金的10%投入私人市场。在美国，401(k)退休账户正在向另类投资开放。金融稳定委员会估计，2024年散户投资者已持有全球私人信贷约13%，较近十年前有所上升。
In Britain, the largest workplace pension providers have pledged under the Mansion House Accord to put 10% of their default funds into private markets by 2030. In the US, 401(k) retirement accounts are being opened up to alternatives. The FSB estimated retail investors already held around 13% of global private credit in 2024, up from almost a decade before.
大多数养老金储户并不主动选择投资。在美国，2024年，59%的由先锋集团管理的职场计划参与者将其整个账户投资于单一目标日期基金。在英国，根据金融行为监管局的数据，超过90%的职场养老金储户投资于其计划的默认策略。因此，人工智能最不透明债务的边际买家，越来越可能是那些永远不会阅读基金季度报告的储户。
Most pension savers don’t actively choose their investments. In the US, 59% of participants in Vanguard-administered workplace plans held their entire account in a single target-date fund in 2024. In Britain, more than 90% of workplace pension savers are invested in their scheme’s default strategy, according to the Financial Conduct Authority. So the marginal buyer of AI’s least transparent debt could increasingly be savers who will never read a fund’s quarterly report.
在过去的科技繁荣中，晚来的资金通过公共市场进入，如2000年日内交易者购买点com股票。如果这次繁荣以糟糕告终，晚来的资金可能通过养老金进入，价格由模型设定。与点com投资者不同，储户不会在损失发生时看到损失。他们将在估值追上现实，或试图退出时才发现。
In past tech booms, the late money came through public markets, from day traders buying dotcom stocks in 2000. If this one ends badly, the late money could arrive through pensions, at prices set by models. And unlike the dotcom investor, the saver will not see the loss as it happens. They will find out when the marks catch up, or when they try to leave.