IT之家 9 月 24 日消息,据路透社报道,一项最新研究表明,美国人工智能基础设施建设对经济产出的消耗,可能超过电力、铁路、州际高速公路以及互联网等历史上大型基础设施建设项目,而且围绕 AI 产业形成的融资结构正变得越来越复杂,可能带来平台性金融风险。
哥伦比亚大学商学院金融与房地产学教授斯泰恩 · 范 · 纽文伯格(Stijn Van Nieuwerburgh)在为本周布鲁金斯学会(Brookings Institution)一场会议准备的论文中写道,最初由亚马逊、Meta 以及 Alphabet 旗下谷歌等公司利用自身积累的现金开展的 AI 投资,如今已经演变成一项大规模扩张计划。预计到 2032 年,美国每年将有约 3.6% 的国内生产总值(GDP)投入这一领域,累计规模超过 10 万亿美元(IT之家注:现汇率约合 67.23 万亿元人民币),同时融资方式也变得越来越复杂。
这一估算高于 19 世纪末美国铁路建设时期每年约占 GDP 2.2% 的投入。当时,铁路是美国历史上成本最高的通用工艺基础设施建设项目。相比之下,20 世纪 50 年代开始建设的美国州际高速公路平台,以及 20 世纪 90 年代中期启动的电信基础设施扩张,每年的投入规模均略高于 GDP 的 1%。
范 · 纽文伯格指出,铁路和电信基础设施扩张都曾引发显著的资产泡沫以及随后浮现的崩溃,而如今 AI 基础设施建设的庞大规模、尚未得到验证的收入来源,以及围绕 AI 产业逐步形成的复杂融资结构,都意味着这一轮扩张可能面临较大的下行风险。
“这实在太复杂了。”范 · 纽文伯格在向记者介绍相关研究时表示。当下,AI 企业、大型科技超大规模云服务商、银行、私人信贷机构、房地产公司以及众多参与基础设施建设的其他机构之间,正在形成越来越复杂的融资安排。
他保守估计,将来 7 年美国将新增约 183 吉瓦的数据中心容量,而当下已经安装的数据中心容量约为 57 吉瓦。
数据中心建设以及 AI 带来的潜在风险,已经成为美国政治和经济讨论的核心议题之一。一些地方政府越来越不愿意接纳大型数据中心,担心其对当地电力、水资源等基础设施造成压力;与此同时,美联储官员也在考虑,这一轮数据中心建设热潮是否正在加剧通胀压力。一些 AI 企业高管则已经提出,放缓基础设施建设和 AI 开发速度或许更加安全。
“存在实质性的下行风险”
范 · 纽文伯格在论文中指出,当下正在开展的 AI 基础设施投资规模,已经超过主要企业依靠自身现金流所能承担的范围。随着产业越来越依赖外部融资,总体杠杆率不断提高,风险也被重新分散到整个经济体系中,同时整个投资体系开始依赖尚未得到验证的将来收入来源。
这篇论文将于当地时间周五正式提交。
范 · 纽文伯格在接受记者采访时表示,围绕 AI 产业形成的大量特殊目的实体(SPV)以及相关融资安排,其不透明程度让人联想到美国次贷危机时期的情况。
当时,复杂的住房抵押贷款融资结构最终浮现大规模难题,并以足以冲击全球金融体系的速度恶化,最终引发了 2007 年至 2009 年的美国经济衰退。
不过,他在论文中强调,当下的情况并不意味着金融危机已经迫在眉睫。
“这些发展并不意味着金融困境即将浮现。AI 应用的强劲增长、较高的基础设施利用率以及模型能力的持续提高,都可能支撑预期中的基础设施投资,并带来稳定的现金流。”他写道。
但他同时指出,不确定的市面需要、快速的工艺变化、执行过程中的瓶颈以及较高的杠杆率叠加在一起,如果市面预期发生调整,就可能带来“实质性的下行风险”。
作为一个例子,范 · 纽文伯格估算,到 2032 年,AI 行业每年需要创造约 3.7 万亿美元(现汇率约合 24.87 万亿元人民币)的收入,才能达成当下预期的投资回报。
他表示,当下对 OpenAI 和 Anthropic 年合计收入的估算约为 1,000 亿美元(现汇率约合 6,722.65 亿元人民币),这意味着相关收入需要以每年约 80% 的速度增长,才能达到上述目标。
本文来源:https://www.ithome.com/1/006/987.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT之家 September 24 -- According to Reuters, a new study suggests that the U.S. artificial intelligence infrastructure build-out may absorb more economic output than historically major infrastructure projects such as electricity, railroads, interstate highways and the internet, and the financing structures forming around the AI industry are becoming increasingly complex and may pose platform-level financial risks.
Stijn Van Nieuwerburgh, a professor of finance and real estate at Columbia Business School, wrote in a paper prepared for a Brookings Institution conference this week that the AI investment initially carried out by companies such as Amazon, Meta and Alphabet's Google using their own accumulated cash has now evolved into a massive expansion plan. By 2032, the United States is expected to devote about 3.6% of its gross domestic product (GDP) to this sector each year, with a cumulative scale of more than $10 trillion (IT之家 note: at the current exchange rate, equivalent to about 67.23 trillion yuan), while the financing methods have become increasingly complex.
This estimate is higher than the roughly 2.2% of GDP invested each year during U.S. railroad construction in the late 19th century. At the time, railroads were the most expensive general-purpose infrastructure project in U.S. history. By comparison, the U.S. interstate highway system, whose construction began in the 1950s, and the telecommunications infrastructure expansion launched in the mid-1990s each involved annual spending slightly above 1% of GDP.
Van Nieuwerburgh noted that the expansion of both railroad and telecommunications infrastructure triggered significant asset bubbles followed by crashes, and that the enormous scale of today's AI infrastructure build-out, its unproven revenue sources and the increasingly complex financing structures taking shape around the AI industry mean this expansion may face substantial downside risk.
"This is just too complex," Van Nieuwerburgh said as he briefed reporters on the research. AI companies, major technology hyperscale cloud providers, banks, private credit institutions, real estate firms and many other organizations involved in infrastructure construction are forming increasingly complex financing arrangements.
He conservatively estimates that the United States will add about 183 gigawatts (GW) of data center capacity over the next seven years, while installed data center capacity is currently about 57 GW.
Data center construction and the potential risks posed by AI have become core issues in U.S. political and economic debate. Local governments are becoming increasingly reluctant to accept large data centers, worrying about the pressure they put on local electricity and water resources; meanwhile, Federal Reserve officials are considering whether the data center construction boom is adding inflationary pressure. Some AI company executives have already suggested that slowing the pace of infrastructure construction and AI development may be safer.
"Substantial downside risks exist"
Van Nieuwerburgh wrote in the paper that the scale of AI infrastructure investment currently under way already exceeds what major companies can support using their own cash flows. As the industry becomes increasingly dependent on external financing, overall leverage continues to rise, risk is being redistributed across the whole economic system, and the entire investment system has begun to rely on unverified future revenue sources.
The paper will be formally submitted Friday local time.
Van Nieuwerburgh told reporters that the large number of special-purpose vehicles (SPVs) and related financing arrangements emerging around the AI industry are so opaque that they recall the situation during the U.S. subprime crisis.
At that time, complex residential mortgage-backed financing structures eventually exposed massive problems and deteriorated at a speed capable of shocking the global financial system, ultimately triggering the 2007 to 2009 U.S. economic recession.
However, he emphasized in the paper that the current situation does not mean a financial crisis is imminent.
"These developments do not mean that financial distress is about to emerge. Strong growth in AI adoption, high infrastructure utilization rates and continuing improvements in model capabilities could all support expected infrastructure investment and deliver stable cash flows," he wrote.
But he also pointed out that if market expectations adjust, the combination of uncertain market demand, rapid technological changes, execution bottlenecks and high leverage could bring "substantial downside risk."
As an example, Van Nieuwerburgh estimated that by 2032 the AI industry will need to generate about $3.7 trillion in annual revenue (at the current exchange rate, equivalent to about 24.87 trillion yuan) to deliver the investment returns currently expected.
He said the current estimates for the combined annual revenue of OpenAI and Anthropic are about $100 billion (at the current exchange rate, equivalent to about 672.265 billion yuan), which would require revenue to grow at about 80% per year to reach that target.
