AI的钱从哪来,谁在付Who Pays for the AI Bill
黑石总裁在台上说,AI 建设需要更多股权融资。翻译一下:债务那边装不下了,风险得找人分摊。
Gray 给的理由很诚实:债权人只有利息和本金,涨了也分不到,所以风险集中在他们身上。这话说得像风险管理,其实是融资方在给下一轮的钱找出口。股权融资意味着什么?意味着把风险从资产负债表挪到别人的养老金、保险和储蓄账户里。谁买这些股权,谁就在为「算力需求远高于供给」这句判断兜底。报道里没有出现任何一个这样的出资人的名字。
Roman 的数字更直白:未来五年 AI 资本开支约 10 万亿美元,今年已经发了约 5500 亿美元的 AI 债务,而全球公司债市场总共约 20 万亿。他说的原话是「某个时点,它会太多」。高盛那份报告给的数据中心资本开支是 8.5 万亿。这些数字放在一起,不是在描述一个产业,是在描述一次规模空前的举债。而举债的抵押物,是还没被验证的需求。
Gray 说黑石投资组合公司过去一年在 AI 模型上的支出涨了 20 倍,理由是编码、客户互动、内容创作带来了财务收益。20 倍是个很漂亮的数字,但报道没说这 20 倍支出换回了多少收入。支出涨 20 倍和收益涨 20 倍是两件事,前者可以靠融资完成,后者不行。
真正值得盯的是那句「结构化的交易,以限制他们承担的债务量」。结构化交易这个词在 2008 年之前也出现过,它的功能是把风险切碎、重新打包、卖给出价最高的人,然后让原始借款人看起来干净。Gray 说需要「创造力」,创造力在这里不是技术词,是金融工程词。
这套机制里,暴力不在 direct 层,没有人挨打。它在 structural 层:资本把未来五年的预期收益提前变现,代价由买股权的人、被涨价挤压的普通用户、以及最后接盘的那一批人承担。而 cultural 层负责让这一切听起来像进步——「AI 是未来」这个叙事,就是让 10 万亿的赌注显得像常识的那层糖衣。
峰会在香港开,讲的是全球资本怎么给美国的数据中心输血。谁被邀请上台讲话,谁的数字被引用,谁的需求被说成「远高于供给」——解释权在这里完成分配。至于那些电费、水费、土地被重新定价的地方居民,他们不在这个 panel 上。
Gray 说选址、许可、芯片、电力、资本都越来越难。前四样是物理约束,最后一样是这次会议要解决的问题。物理约束不会因为融资结构创新而消失,只会换个账户记。
Blackstone's president told a Hong Kong audience that AI buildout needs more equity. Translation: debt can no longer absorb it, so the risk needs somewhere else to go.
Gray's reasoning is honest enough. Lenders get interest and principal, nothing more, so the concentration risk sits on them. That sounds like risk management. It is actually a seller looking for the next buyer. Equity means moving the risk off the balance sheet and into someone's pension, insurance, savings. Whoever buys that equity is underwriting the claim that demand for compute is running well ahead of supply. The report names none of these people.
Roman's numbers are blunter: about $10 trillion of AI capex over five years, roughly $550 billion of AI debt issued this year, against a global corporate bond market of about $20 trillion. His words: at some point it's going to be too much. Goldman's data center figure is $8.5 trillion. Put together, this is not a description of an industry. It is a description of a debt event. The collateral is demand that has not been proven yet.
Gray says AI spending across Blackstone's portfolio companies rose 20-fold in a year, citing coding, customer engagement, content creation. Twenty-fold is a handsome number. The report does not say how much revenue that spending returned. Spending can be financed. Revenue cannot.
Watch the phrase about structured transactions to limit debt. That term had a career before 2008. Its function is to slice risk, repackage it, sell it to whoever bids, and leave the original borrower looking clean. Gray calls it creativity. Creativity here is a finance word, not a technology word.
No one is hit in this story, so the violence is not direct. It is structural: capital cashes in five years of expected returns in advance, and the bill lands on equity buyers, on users facing higher prices, on whoever holds it last. The cultural layer does the smoothing —「AI is the future」is the sugar that makes a $10 trillion bet look like common sense.
The summit was in Hong Kong; the subject was how global capital feeds American data centers. Who gets the panel, whose numbers get cited, whose demand gets called ahead of supply — that is where the interpretive power is handed out. The people whose power bills, water and land get repriced are not on the panel.
Gray lists sites, permits, chips, power and capital as the hard parts. The first four are physical. The last one is what the summit exists to solve. Physical constraints do not vanish when you restructure the financing. They just move to a different account.