
就在业内高管和分析师纷纷质疑人工智能是否正在催生又一个泡沫之际,信贷投资者正将数十亿美元资金投入这个新技术领域。
据知情人士本周透露,摩根大通(JPMorgan Chase & Co.)和三菱日联金融集团(Mitsubishi UFJ Financial Group)正牵头承销一笔超过220亿美元的贷款,用于支持Vantage Data Centers建设一个庞大的数据中心园区。据彭博社本月报道,Facebook母公司Meta Platforms Inc.将从太平洋投资管理公司(Pacific Investment Management Co.)和Blue Owl Capital Inc.获得290亿美元资金,用于在路易斯安那州乡村地区建设一个大型数据中心。
未来还将涌现更多此类交易。仅OpenAI一家公司就预计,其未来开发和运营人工智能服务所需的基础设施投入将达数万亿美元。
与此同时,业内关键人物也承认,人工智能投资者很可能将面临阵痛。OpenAI首席执行官萨姆·奥尔特曼本周表示,他看到当前人工智能投资热潮与上世纪90年代末的互联网泡沫存在相似之处。在谈及初创企业估值时,他直言:“总会有人在这里遭受损失。”此外,麻省理工学院(Massachusetts Institute of Technology)的一项研究计划发布报告指出,在企业界,95%的生成式人工智能项目未能带来任何利润。
这些情况足以让信贷市场的观察者们感到不安。
花旗集团(Citigroup)美国投资级信贷策略主管丹尼尔·索里德表示:“信贷投资者自然会回想起2000年代初的情景,当时电信公司可谓过度建设、过度举债,最终我们看到这些资产出现了重大减值。因此,从中期来看,人工智能热潮无疑会引发关于可持续性的疑虑。”
在早期,用于训练和驱动最先进人工智能模型的基础设施建设资金,主要由人工智能公司自身承担,其中包括Alphabet Inc.旗下谷歌(Google)和Meta Platforms Inc.等科技巨头。不过,近期的资金来源正越来越多地转向债券投资者和私人信贷机构。
据彭博情报(Bloomberg Intelligence)近期分析,这类融资风险敞口的形式和规模多样,风险等级也各不相同。许多大型科技公司——即所谓的AI超大规模服务商——一直通过发行优质企业债来为新建基础设施融资。由于这些债务有现有现金流作为担保,因此被认为相对安全。
如今,大部分债务融资正来自私人信贷市场。
瑞银集团(UBS)信贷策略主管马修·米什表示:“过去三个季度,人工智能领域的私人信贷融资规模每季度约为500亿美元,这是保守估计。即便不计入Meta和Vantage的巨额交易,私人信贷市场的资金供给也已是公共市场的两到三倍。”
与此同时,许多新的计算中心正通过商业地产抵押贷款支持证券(CMBS)融资,这类证券并非与企业主体挂钩,而是与园区产生的付款绑定。据摩根大通本月估算,人工智能基础设施支持的CMBS已较2024年全年总额增长30%,达到156亿美元。
索里德与花旗的一位同事在8月8日发布了一份报告,重点分析了公用事业公司面临的特殊风险。这些公司为建设满足高能耗数据中心所需的电力基础设施而大幅举债。索里德及其同事与其他分析师一样,都对当前如此巨额的投入感到担忧,因为人工智能项目尚未证明其具备长期创造营收的能力。
标普全球评级(S&P Global Ratings)私人市场分析全球主管露丝·杨表示:“数据中心项目的融资周期往往长达20至30年,而我们甚至无法确定五年后这项技术会是什么样子。我们会保守评估未来现金流,因为缺乏历史参考。”
瑞银集团指出,压力已经开始显现,表现之一是面向科技领域的私人信贷机构的实物支付(PIK)贷款正在增加。根据瑞银的数据,第二季度,商业发展公司(BDCs)的PIK收入占比升至6%,创下自2020年以来的最高水平。
但这股资金洪流短期内似乎难以停歇。
穆迪全球项目与基础设施融资团队高级副总裁约翰·梅迪纳表示:“直接贷款机构不断在筹集资本,而这些资金必须找到去处。他们把这些资本需求巨大的AI超大规模服务商视作下一个长期基础设施投资标的。”(*)
译者:刘进龙
审校:汪皓
就在业内高管和分析师纷纷质疑人工智能是否正在催生又一个泡沫之际,信贷投资者正将数十亿美元资金投入这个新技术领域。
据知情人士本周透露,摩根大通(JPMorgan Chase & Co.)和三菱日联金融集团(Mitsubishi UFJ Financial Group)正牵头承销一笔超过220亿美元的贷款,用于支持Vantage Data Centers建设一个庞大的数据中心园区。据彭博社本月报道,Facebook母公司Meta Platforms Inc.将从太平洋投资管理公司(Pacific Investment Management Co.)和Blue Owl Capital Inc.获得290亿美元资金,用于在路易斯安那州乡村地区建设一个大型数据中心。
未来还将涌现更多此类交易。仅OpenAI一家公司就预计,其未来开发和运营人工智能服务所需的基础设施投入将达数万亿美元。
与此同时,业内关键人物也承认,人工智能投资者很可能将面临阵痛。OpenAI首席执行官萨姆·奥尔特曼本周表示,他看到当前人工智能投资热潮与上世纪90年代末的互联网泡沫存在相似之处。在谈及初创企业估值时,他直言:“总会有人在这里遭受损失。”此外,麻省理工学院(Massachusetts Institute of Technology)的一项研究计划发布报告指出,在企业界,95%的生成式人工智能项目未能带来任何利润。
这些情况足以让信贷市场的观察者们感到不安。
花旗集团(Citigroup)美国投资级信贷策略主管丹尼尔·索里德表示:“信贷投资者自然会回想起2000年代初的情景,当时电信公司可谓过度建设、过度举债,最终我们看到这些资产出现了重大减值。因此,从中期来看,人工智能热潮无疑会引发关于可持续性的疑虑。”
在早期,用于训练和驱动最先进人工智能模型的基础设施建设资金,主要由人工智能公司自身承担,其中包括Alphabet Inc.旗下谷歌(Google)和Meta Platforms Inc.等科技巨头。不过,近期的资金来源正越来越多地转向债券投资者和私人信贷机构。
据彭博情报(Bloomberg Intelligence)近期分析,这类融资风险敞口的形式和规模多样,风险等级也各不相同。许多大型科技公司——即所谓的AI超大规模服务商——一直通过发行优质企业债来为新建基础设施融资。由于这些债务有现有现金流作为担保,因此被认为相对安全。
如今,大部分债务融资正来自私人信贷市场。
瑞银集团(UBS)信贷策略主管马修·米什表示:“过去三个季度,人工智能领域的私人信贷融资规模每季度约为500亿美元,这是保守估计。即便不计入Meta和Vantage的巨额交易,私人信贷市场的资金供给也已是公共市场的两到三倍。”
与此同时,许多新的计算中心正通过商业地产抵押贷款支持证券(CMBS)融资,这类证券并非与企业主体挂钩,而是与园区产生的付款绑定。据摩根大通本月估算,人工智能基础设施支持的CMBS已较2024年全年总额增长30%,达到156亿美元。
索里德与花旗的一位同事在8月8日发布了一份报告,重点分析了公用事业公司面临的特殊风险。这些公司为建设满足高能耗数据中心所需的电力基础设施而大幅举债。索里德及其同事与其他分析师一样,都对当前如此巨额的投入感到担忧,因为人工智能项目尚未证明其具备长期创造营收的能力。
标普全球评级(S&P Global Ratings)私人市场分析全球主管露丝·杨表示:“数据中心项目的融资周期往往长达20至30年,而我们甚至无法确定五年后这项技术会是什么样子。我们会保守评估未来现金流,因为缺乏历史参考。”
瑞银集团指出,压力已经开始显现,表现之一是面向科技领域的私人信贷机构的实物支付(PIK)贷款正在增加。根据瑞银的数据,第二季度,商业发展公司(BDCs)的PIK收入占比升至6%,创下自2020年以来的最高水平。
但这股资金洪流短期内似乎难以停歇。
穆迪全球项目与基础设施融资团队高级副总裁约翰·梅迪纳表示:“直接贷款机构不断在筹集资本,而这些资金必须找到去处。他们把这些资本需求巨大的AI超大规模服务商视作下一个长期基础设施投资标的。”(*)
译者:刘进龙
审校:汪皓
Credit investors are pouring billions of dollars into artificial intelligence investments, just as industry executives and analysts are raising questions about whether the new technology is inflating another bubble.
JPMorgan Chase & Co. and Mitsubishi UFJ Financial Group are leading the sale of a more than $22 billion loan to support Vantage Data Centers’ plan to build a massive data-center campus, people with knowledge of the matter said this week. Meta Platforms Inc., the parent of Facebook, is getting $29 billion from Pacific Investment Management Co. and Blue Owl Capital Inc. for a massive data center in rural Louisiana, Bloomberg reported this month.
And plenty more of these deals are coming. OpenAI alone estimates it will need trillions of dollars over time to spend on the infrastructure required to develop and run artificial intelligence services.
At the same time, key players in the industry acknowledge there is probably pain ahead for AI investors. OpenAI Chief Executive Officer Sam Altman said this week that he sees parallels between the current investment frenzy in artificial intelligence and the dot-com bubble in the late 1990s. When discussing startup valuations he said, “someone’s gonna get burned there.” And a Massachusetts Institute of Technology initiative released a report indicating that 95% of generative AI projects in the corporate world have failed to yield any profit.
Altogether, it’s enough to make credit watchers nervous.
“It’s natural for credit investors to think back to the early 2000s when telecom companies arguably overbuilt and over borrowed and we saw some significant writedowns on those assets,” said Daniel Sorid, head of U.S. investment grade credit strategy at Citigroup. “So, the AI boom certainly raises questions in the medium term around sustainability.”
The early build-out of the infrastructure needed to train and power the most advanced AI models was largely funded by the AI companies themselves, including tech giants like Alphabet Inc.’s Google and Meta Platforms Inc. Recently, though, the money has been increasingly coming from bond investors and private credit lenders.
The exposure here comes in many shapes and sizes, with varying degrees of risk. Many large tech companies — the so-called AI hyperscalers — have been paying for new infrastructure with gold-plated corporate debt, which is likely safe due to the existing cash flows that secure the debt, according to recent analysis from Bloomberg Intelligence.
Much of the debt funding now is coming from private credit markets.
“Private credit funding of artificial intelligence is running at around $50 billion a quarter, at the low end, for the past three quarters. Even without factoring in the mega deals from Meta and Vantage, they are already providing two to three times what the public markets are providing,” said Matthew Mish, head of credit strategy at UBS.
And many new computing hubs are being funded through commercial mortgage-backed securities, tied not to a corporate entity, but to the payments generated by the complexes. The amount of CMBS backed by AI infrastructure is already up 30%, to $15.6 billion, from the full year total in 2024, JPMorgan Chase & Co. estimated this month.
Sorid and a colleague at Citi put out a report on Aug. 8 focusing on the particular risks for the utility firms that have boosted borrowing to build the electrical infrastructure needed to feed the power-hungry data centers. They and other analysts share a commonly held concern about spending so much money right now, before AI projects have shown their ability to generate revenue over the long term.
“Data center deals are 20 to 30 year tenor fundings for a technology that we don’t even know what they will look like in five years,” said Ruth Yang, global head of private market analytics at S&P Global Ratings. “We are conservative in our assessment of forward cash flows because we don’t know what they will look like, there’s no historical basis.”
The stress has begun to appear in the rise of payment-in-kind loans to tech-oriented private credit lenders, UBS Group noted. In the second quarter, PIK income in BDCs reached the highest level since 2020, climbing to 6%, according to UBS.
But the fire hose of money is unlikely to stop anytime soon.
“Direct lenders are constantly raising capital, and it has to go somewhere,” said John Medina, senior vice president in Moody’s Global Project and Infrastructure Finance Team. “They see these hyperscalers, with this massive capital need, as the next long-term infrastructure asset.”