传统保守资金的转向：寿险“浮存金”与AI基础设施的隐秘联结
在社交平台 X 上，一则关于人寿保险公司大举买入被包装为 A+ 级的“垃圾”数据中心债券的帖子，勾起了人们对电影《大空头》（The Big Short）开篇的痛苦回忆。尽管流传的“美国政府承诺在危机时为违约兜底”的说法并不属实，但寿险资金深度参与 AI 基础设施建设的现状却是客观存在的事实。要理解这一切是如何发生的，必须先回到保险公司的盈利模式。保险公司的主要收入来源之一是浮存金（Float：保险公司在收取保费与实际赔付之间的时间差内所持有的闲置资金）。由于人寿保险和养老金具有自然的“长尾”属性，资金可以沉淀数十年，因此投资收益往往能占到寿险公司总收入的 50% 左右。
在历史上，这类资金的投资策略极其保守，主要流向国债、高质量企业债、抵押贷款和实体房地产。然而，随着近年来公共债券市场的收益率不断被压缩，为了满足长期回报目标，这些保守资金开始将目光投向了过去几年中唯一展现出惊人资本吸纳能力的领域——AI 基础设施。数据中心由于具备长期租约、稳定的科技巨头（Hyperscalers）租户以及可预测的现金流，成为了完美契合保险资金需求的“类地产”资产。然而，这也使得原本规避风险的养老金和寿险资金，不可避免地暴露在 AI 行业的周期波动之中。
Original English Source
I came across a post on X that gave me flashbacks to the beginning of the movie Big Short. Life insurance companies buy garbage data center bonds mixed with high-quality assets keeping them A+ rated. Life insurance companies are protected by the US government, so in the event of a crash, they'll buy back defaults. This is not 2008, this is 2026. Get the hell on with it. Did it give me the heebie-jeebies? Yes, it absolutely did. And I had no idea if it was true. So, I went looking. Are data center bonds actually rated A+? Are life insurance companies really funding AI infrastructure? And what does this mean for the money sitting behind your life insurance, or your parents' pensions, or your own retirement account? Is the capital that has always been highly conservative being used to finance AI chips, power grid, and data centers? Are we really facing the 2008 crisis? And are we really buying garbage bonds? Let's find out. Let's start by refreshing our memory on how insurance companies make money. There are two primary ways: selling insurance with an added margin, pretty straightforward, and investing the money that they collected from people who bought the insurance. Life insurance companies in particular are heavily dependent on the investment portion of the income. And often around 50% of their revenue comes from investment earnings rather than insurance sales. The money that insurance company collects every month is otherwise known as the float. So, the float gets invested, and the longer it sits in investments and earns returns, the better. Which is why products like life insurance or pensions with a naturally long tail rely on investments so much. It would be stupid not to. Insurance companies investment strategy has always been very conservative. They invest in government bonds, corporate bonds, mortgage, real estate, and for decades that conservative strategy was more than enough. Because the yields on those safe assets were also high enough. But over the past several years, the earnings on public bonds compressed. And insurance companies found themselves needing to hit the same long-term return targets with less income from the safe investments. And what has been showing unprecedented amount of capital for the past few years? AI.机制一：私有信贷与特设目的载体（SPV）的表外融资戏法
AI 数据中心动辄数百亿美元的建设资金，正通过第一种关键机制——私有信贷（Private Credit：非公开交易的定制化贷款）和特设目的载体（Special Purpose Vehicle: 独立于母公司的防火墙法人实体）流向科技巨头。以 2025 年 10 月 Meta 在路易斯安那州筹建的巨型 AI 数据中心为例，Meta 并没有直接举债 300 亿美元，因为这会被计入资本开支并严重消耗其自身的借贷额度，阻碍其未来的连续扩张。
为了实现表外融资，Meta 成立了一个独立的 SPV 来持有该数据中心。私有信贷巨头蓝猫资本（Blue Owl Capital）出资认购了该 SPV 80% 的股权，Meta 持有余下的 20%。随后，该 SPV 向以 PIMCO（出资 180 亿美元）和黑石（BlackRock，出资 30 亿美元）为首的私有债权人团进行大额私募借贷。SPV 建成数据中心后，Meta 与其签订长期租赁协议并支付租金，SPV 再用这笔租金向 PIMCO 和黑石偿还利息。在此过程中，资金完全绕过了公开交易市场。
在这个链条中，寿险资金成为了影子债权人。例如，Meta 此前与阿波罗全球管理（Apollo）达成过一项几乎相同的 290 亿美元融资交易，而资金正是通过阿波罗旗下的寿险公司 Athene 提供的——这正是普通人购买人寿保险所缴纳的保费。通过将本属高风险的初创数据中心资产与 Meta 这样具有双 A 级信用评级的租户绑定，信用评级机构便给这些 SPV 债券评出了 A+ 的高分。然而，一旦 AI 需求未达预期导致 SPV 破产，Meta 作为母公司在法律上并不承担债务兜底责任，最终买单的将是提供资金的寿险政策持有者。
Original English Source
And if you think about AI, insurance companies clearly wouldn't be investing in pure tech portfolio because, first of all, OpenAI and Anthropic are not even public yet. Secondly, because Google Nvidia is not really part of a typical investment portfolio for an insurance company. But what is something that touches AI and typical assets such as real estate that insurance would invest in? The data centers. Data centers have something that makes them very attractive to the kind of investor an insurer already is. They're a long-term lease with a hyperscaler tenant. They've got contracted cash flows and multi-decade assets. That is, if they don't fall apart. And what is happening now is that AI data centers are being wired into the conservative investment portfolio of insurance companies through A+ rated bonds. How do they get that rating, you may ask? They get it through the mechanism that we explained in this video that went viral called the special purpose vehicle. So, how exactly is this happening? Let's unpack that. Mechanism number one, private credit. In October 2025, Meta needed to build a massive AI data center in Louisiana. To build the data center, they needed money. And here's what they did. Instead of borrowing $30 billion directly, Meta created a separate legal entity, or an SPV, that would own the data center. Why? Don't they already have $30 billion to pay up front? I mean, with all due respect, it's Meta. Well, a few reasons. If Meta fully owned it, the $30 billion would count as capital expenditure. And if they took that loan directly, it would use up Meta's borrowing capacity. And this is a problem because if Meta decides to spin up another data center 2 months from now, and then another one, and another one, they would have already had a massive debt on their books, which would prevent them from borrowing more. So, they spin up an SPV, a totally legal thing to do. Blue Owl Capital, a private credit firm, puts up 80% of the ownership of that SPV. Meta keeps the remaining 20%. The SPV needs cash to build the data center, so it borrows it privately from a small group of lenders. PIMCO puts in 18 billion, BlackRock 3 billion, and other private lenders fill in the rest. The 30 billion never touch a public exchange. It is negotiated deal by deal between the SPV and each lender privately. This is the private in private credit. Then, Meta signs a long-term lease with its own SPV. Meta pays rent to use the data center, and that rent is the money that SPV uses to pay interest back to PIMCO, BlackRock, and other lenders. You may ask, "But what does insurance have to do with this?" Stay with me. PIMCO and BlackRock are asset managers. They manage money on behalf of their clients. So do life insurers. Life insurance companies are active participants in private credit. And just like PIMCO and BlackRock, they lend money to SPVs. Meta, in fact, had an almost identical deal done with Apollo. Apollo acted as a lender. They gave them 29 billion dollars through Athene. Athene is Apollo's life insurance company. And where does Athene get the money from? From people who pay for life insurance. And this is how a data center that is a risky asset. It is risky because of a lot of things. It depends on the grid, on the water supply, on the chip demand, on the availability of chips, on a multitude of geopolitical factors, becomes an A-rated bond. Think about it. If you take a random data center and put it under an A+ bond, you end up with a B-rated bond. But if you take a data center under an SPV and bundle it with a double A tenant like Meta, you end up with an A+ bond. And if the data center falls apart, by fall apart I'm referring to events like the AI bubble burst or we realize that we have massive overcapacity or we don't have enough grid to support the data centers we built or the government comes up with laws against data centers, the SPV can file bankruptcy. Meta is not going to cover the lost money because it's not on Meta's books and it's not Meta that owes money to the lenders. BlackRock and PIMCO will take a hit but they manage money that people gave them as investments and investments come with a certain level of risk anyway. But when it's your life insurance money, it might very much affect your well, life.机制二与三：资产证券化与物理建造，养老金深度介入AI版图
除了直接进行私有信贷借贷，资本市场还通过资产证券化（Asset-Backed Securities: 资产支持证券，简称 ABS）和直接基础设施投资这两种渠道将大众资金引入 AI 建设。
在 ABS 机制中，以数据中心运营商 QTS 为例，若其在菲尼克斯拥有两栋已与科技巨头签订长期租约的数据中心，为了提前套现未来的租金流，QTS 会将这些建筑和租赁合同打包转入一个 SPV。该 SPV 随即在市场上向投资者发行 ABS 债券。投资者立即支付现金，并在未来逐月获取租金分配。这种将未来租金“折现”的证券化产品，其最大买家正是需要长期匹配资产负债的人寿保险公司。然而，普通房产的 ABS 基础是居民的刚性居住需求，而数据中心 ABS 的底层资产却高度依赖于 AI 行业的景气度。
与此同时，以加拿大养老基金投资公司（CPP Investments）为代表的国家级养老金则走得更远，开始直接扮演“物理建造者”和开发商的角色：
- 2025 年 7 月：CPP 联合德意志银行向一家 SPV 联合提供了 2.25 亿加元，用于在安大略省剑桥市建设数据中心。
- 2025 年 12 月：CPP 与工业地产巨头嘉民集团（Goodman Group）达成合资，在欧洲（法兰克福、阿姆斯特丹和巴黎）共同开发价值达 140 亿美元的数据中心项目。
- 2026 年 7 月：CPP 宣布向数据中心运营商 EdgeConneX 承诺投入约 20 亿美元，用于欧美地区的 AI 基础设施扩建。
在短短一年内，这个代表着数千万加拿大劳动者退休金的 G7 国家最大养老基金之一，就向 AI 数据中心领域砸下了超 100 亿美元。放眼整个美国，寿险公司持有的 4 万亿美元债券资产中，约有 20%（即 8000 亿美元）被投资于此类难以快速变现的私有信贷、ABS 以及基础设施信贷中。
Original English Source
Mechanism number two, securitized products. Picture a major data center operator, QTS. QTS owns two big data center buildings in Phoenix, Arizona. QTS signs a long-term lease with a hyperscaler tenant who pays rent every month. Now, instead of waiting for years to collect that rent, QTS wants the cash now. And so they create an SPV. They transfer the Phoenix buildings and the release contracts into the SPV. The SPV's only job is to hold those buildings and collect the rent. Then the SPV sells bonds to investors. Investors pay cash right away. SPV gets cash instantly and investors get rent payments that the SPV will collect over the years into the future. This is called an asset-backed security or an ABS. ABS can be backed by all sorts of future payments. Be it credit card bills or equipment leases or auto loans, rent payments. And now, increasingly data center rent. It's like if you borrow money from your parents and buy a house with the money that you borrowed for cash, and you pay your parents back by collecting rent from tenants in the house that you bought. The problem is that houses are rented by people, and data centers in many ways rely on the hype around AI. And who do you think invests in asset-backed security assets? That's right. Life insurance companies. Mechanism number three, infrastructure. For decades, infrastructure as an asset referred to boring but essential things that countries couldn't live without. Things like roads, airports, ports, water plants, factories, nuclear stations. It was the kind of assets that keep producing cash whether the economy is booming or in a recession. Because people needed to live. Life insurance companies love this stuff exactly because it is slow, predictable, and lasts 30-plus years. You know who else likes to invest in infrastructure? Pension funds. And through infrastructure, life insurance and pension funds also invest money in AI buildout. Let's follow one specific pension fund to see exactly how this happens. The Canada Pension Plan, or using the word that every Canadian knows very well starting age 18, the CPP. For every working Canadian who is employed full-time and pays taxes, there is a line on every pay slip we get that says CPP. The Canada Pension Plan's Reserve Fund is managed by CPP Investments, and this arm manages retirement investments of basically every working Canadian. So, in July 2025, CPP put up 225 million Canadian dollars to build a data center in Cambridge, Ontario. Once again, the CPP and Deutsche Bank lent money to an SPV to build a data center. Five months later, December 2025, CPP signs a partnership with a real estate giant, Goodman Group, to build a portfolio of data centers across Europe worth up to $14 billion. The data centers are going to be located in Frankfurt, Amsterdam, and Paris. In this case, the CPP is also the lender and co-developer of data centers. They're not only lending money, they're also constructing data centers. Again, I want you to understand the scale. The Canadian pension fund physically builds data centers. July 2026, CPP commits roughly $2 billion to build out more AI infrastructure across Europe and North America. And they do it in partnership with private equity firms and a data center operator, EdgeConneX. Now, if you add it all up, in the span of a year, the largest pension fund of the second largest country on the planet, member of G7, and yes, I know we'll probably be kicked out soon given how our economy is doing, but nevertheless, committed well over $10 billion in three separate deals into AI data centers locally and in Europe through construction partnerships. Now, zoom out from pension and insurers in Canada and look at the southern neighbor, the US. Life insurance companies in America collectively hold about $4 trillion in bonds. Out of the $4 trillion, roughly 800 billion, meaning one out of every five dollars, is put into instruments that are difficult to sell. And private credit, ABS, and infrastructure are investing in AI data center build-out. What I'm getting at is the fact that the conservative pension and life insurance capital is now inevitably exposed to AI infrastructure. And AI build-out is being financed by the pools of capital that are supposed to be the safest money in the system.估值错配与快速折旧：为何数据中心不同于传统房地产？
将数千亿大众退休及养老资金引入这一领域的金融机构通常认为，即使 AI 需求降温，数据中心作为物理地产依旧具有极高的清算价值。但这种乐观预估忽略了三个核心系统性风险：
-
芯片折旧与“房地产假象”： 数据中心与普通商业地产的本质区别在于内部搭载的 GPU（图形处理器）。在普通的资产评估中，金融机构常常采用长达数十年折旧周期的地产模型来评估数据中心。然而，GPU 根本不是房地产。科技巨头（如微软、谷歌、亚马逊）通常在 5 至 6 年内将 GPU 资产折旧归零。实际上，由于 NVIDIA 每 18 个月就会迭代一次芯片微架构，每一代新芯片的问世都会导致上一代芯片在能效比和算力上急剧贬值，甚至失去市场竞争力。如果评估机构高估了这些硬件资产的有效寿命，将导致账面上出现数千亿美元的虚增利润和资产价值泡沫。一旦遭遇清算，实际可变现的现金额将远低于账面评估值。
-
严重的产能过剩风险： 芯片算力的疯狂迭代正在引发数据中心建设的局部过剩。一个典型的例子是 Elon Musk 的 Colossus 1 数据中心。在建成初期，该中心的庞大算力因 xAI 自身的训练调整而处于低负荷运转状态，所幸最终成功向 Anthropic 和谷歌进行了二次转租。但对于全行业而言，目前究竟有多少已建成的“空置算力”和未被充分利用的闲置容量，外界无从知晓。
-
地缘政治不确定性： 全球半导体供应链（美国、中国、台湾及中东）正处于前所未有的地缘政治博弈中心。一旦供应链发生硬件断供或能源危机，即使建造了宏伟的机房，拿不到最新制程芯片的数据中心也会瞬间沦为没有产出的物理空壳。
当评级机构继续套用评估写字楼的公式来为这些充斥着高淘汰率芯片的实体进行评级时，“大空头 2.0”的底层逻辑便已成型。
Original English Source
You may ask, "But what's wrong with this? We all know AI isn't going anywhere. It's just the beginning. Compute capacity is the new gold. What's so wrong with investing in AI assets?" It is risky because of three things. A lot of money that people earn throughout their lifetime gets used to fund illiquid debt, long-term loans, and hardware that depreciates fast. When a data center gets built, the loan behind it funds the building shell and all the power that goes in it. But that's normal real estate investment. What makes a data center different is the chips. When somebody funds a data center, they fund the GPUs sitting inside the building. Nevertheless, the loans are typically structured like real estate investments, but the GPUs are not real estate. Microsoft, Google, and Amazon depreciate their GPUs over 5 to 6 years, and Nvidia releases a new chip architecture every 18 months, and each new generation makes the previous one dramatically cheaper and less competitive. Therefore, if hyperscalers overstate how long these chips stay useful, it adds up to hundreds of billions in overstated profits. Even if a hyperscaler says that a GPU is worth a certain amount on paper, and at the same time a lender wants to get their money back and sell that hardware, the real cash is going to be far less than what they claimed before. On top of it, we have a very real risk of overproduction and extremely rapid development of data centers that doesn't match the demand. This already happened with Elon Musk's Colossus 1 data center that was sitting underused. They built it for xAI's own LLM training, but never used up the capacity. And thank God they were able to lease it to Anthropic and Google. And this is just one example. But the reality is we don't know how much unused capacity we already have. And this is why we might be seeing the big short 2.0. Rating agencies are applying real estate models to hardware that becomes obsolete the moment the next generation comes out. And if you combine it with geopolitical tensions between the US, China, Taiwan, the Middle East, there's absolutely no guarantee that the value that these data centers claim to carry holds up. These bonds or loans are priced like real estate. But inside the real estate, which is a data center, there is hardware that loses value the moment the next model comes out.安全网的虚像：国家养老金与寿险资产的系统性风险敞口
面对潜在的 AI 泡沫破裂，人们往往寄希望于金融系统的安全网，但在寿险和养老金领域，这存在巨大的认知误区。
首先，医疗保险并不属于此风险风暴的中心。虽然全美有约 66% 的人口依赖医疗保险，医院的现金流也极度依赖保险公司的日常拨付，但数据表明，健康保险公司（Health Insurers）与人寿保险公司的投资策略截然不同。医疗险公司由于需要应对高频的日常医疗赔付，其资产配置必须保持极高的流动性，主要集中于国债、地方政府市政债和高等级企业债，极少染指私募债与 SPV 等非流动性项目。因此，AI 泡沫的破裂不会直接传导为系统性的医疗支付危机。
然而，对于寿险公司和国家养老基金而言，安全网的缺失才是最核心的隐患。绝大多数购买保险的人潜意识里认为“保险公司也是有保险的”（意指有国家信用兜底）。但事实上，美国联邦政府并没有像针对商业银行存款（FDIC 机制）那样的联邦级机构来为寿险产品或商业养老金提供全额担保。
寿险公司的监管主要隶属于各州政府。当单一的中小型寿险公司出现流动性危机时，州政府的保障基金确实可以介入实施重组或救助。但这种地方性的“互助网”在设计之初，根本无法抵御系统性的行业海啸。当全美前几大寿险巨头因为配置了相同的、缺乏市场每日交易流动性（Illiquid）的私有信贷资产而同时遭遇资产重估贬值时，各州政府的担保机制将在瞬间崩盘。
综上所述，虽然目前尚无确凿证据表明崩溃迫在眉睫，但这向所有普通投资者敲响了警钟：曾经被视作最安全避风港的退休储蓄和长期保单，如今已被悄然捆绑在了高度复杂的 AI 泡沫战车之上。
Original English Source
So all of this got me wondering about things that are tied into insurance that would have a much more devastating impact if we were to go through a crash. And if you live in North America and the US more so than Canada, your entire life very much depends on the health insurance. Which brings me to my final question. If this debt is sitting in insurance company portfolios, what happens to the part of life that depends on insurance more than anything, i.e. health care? In 2026, around 66% of Americans get medical care through medical insurance, and the data confirms that. Which means that hospital cash flow is also fully dependent on the payments from insurance companies because patients don't pay hospitals directly. So, my question was, if life insurance companies are getting wired in the AI stack and national pension funds are getting involved in construction of data centers in Europe and there are constant talks and speculations about the bubble and it's very probable that we may run into an oversupply crisis that already happened with RAM memory after the crypto mining went bust. We highly recommend that you watch the episode about the RAM memory. We talked about non-obvious consequences of this crisis such as schools and students. Or if there is a stock market crash or if data centers start closing. Will that affect health care? Do health insurance companies invest in data centers and AI the same way life insurance and pension funds do? At first sight, the data looks slightly alarming. Health insurance companies invest income from premiums, too, and they also do bonds and mortgages and real estate and private market assets. And their money is also part of the same broader financial system that is funding AI infrastructure. But if you dig deeper, it becomes clear that the investment strategy of health insurers differs significantly from investment strategy of life insurance. Health insurers generally keep more of their portfolios in liquid and conventional assets such as treasuries, municipal bonds, mortgage-backed securities, corporate bonds. They do have alternative investments, but the exposure to illiquid private credit is much more concentrated in the life insurance sector. So, the comforting answer here is that a hypothetical AI data center bust does not automatically become a health care crisis. Health insurers are not, at least today, carrying this particular bet. But this answer poses a different and more uncomfortable question. If the biggest exposure sits with life insurance and pension funds, what protects the people whose retirement income those companies are supposed to back? Most people assume that insurance is insured, but in reality, there is no federal body that steps in and guarantees the money when things go south. Life and health insurance companies are regulated primarily by individual states. And if one insurer fails, the state can step in and save the day. But that system is designed to deal with one insurer failing. When you have a bunch of major insurance companies weakened by the same market shock at the same time, especially when that shock sits in private assets that do not trade every day, it's a different story. So far, there is no evidence that this will happen. And to be clear, this is in no way any sort of prediction that AI market will crash or that data centers will fail or that life insurers will become insolvent. The reason we made this episode is for you to start understanding how the technology transformation affects your capital, your retirement money, your annuity money, your life insurance, so that you can be in control of what's happening. Capital that used to be concentrated in public bonds, mortgages, and traditional infrastructure is shifting to fund complex and technology-dependent AI infrastructure. Whether it is a smart, progressive, risky, but effective evolution or perhaps a mispricing is not something that anyone can know yet. But at the very least, you can stay informed. So let's go back to where we started. Life insurance companies buy garbage data center bonds mixed with high-quality assets keeping them A+ rated. What's actually true in this statement is that, yes, insurance companies are indeed piling billions into data center debt. Yes, the debt is really getting bundled and rated in ways that may not fully capture how fast the hardware depreciates. That's where the big short analogy is coming from. But the government bailout is not true. There is no federal promise to buy back the defaults and that is exactly the point that we were trying to bring forward. The point is that the safest capital in the world is now tied into private credit, securitized bonds, and infrastructure debt, all of which are betting on the data center demand. If AI demand keeps growing the way everyone expects it to, none of this matters. The leases will be paid, the rent is going to be paid, the chips are going to be produced and keep earning more. The bonds will mature on schedule and this can be a forgotten analysis about something that worked out just fine. But if it doesn't and demand slows down, if a cheaper chip architecture arrives faster than we anticipate, if even one major hyperscaler decides to walk away from the lease, the people finding out about this might be the people who put their money in those insurance companies. We built this episode because everybody who follows tech is asking the same question. Are we facing 2008 again? And the answer is nobody knows. So to sum it up, good luck to all of us. We hope you enjoyed the episode and we'll see you in the next one. Bye.📌 文中提及的人物和组织
公司/组织: Meta, BlackRock, PIMCO, Apollo, Goodman Group, Deutsche Bank, EdgeConneX, QTS
媒体/书籍: The Big Short