As enterprises start "reducing the cost of computing power," Goldman Sachs warns that the AI capital expenditure of 5.3 trillion is approaching credit saturation!

The AI infrastructure investment frenzy is reshaping the global capital market landscape, and the hidden debt risks behind it cannot be ignored. Goldman Sachs Group, Inc.'s latest forecast shows that from 2025 to 2030, the capital expenditure of super-scale cloud computing enterprises in the AI and data center field will reach a cumulative total of $5.3 trillion, creating an unprecedented super-cycle of capital expenditure. Goldman Sachs Group, Inc. predicts that super-scale enterprises will need financing from various markets, and they may encounter constraints in the liquidity credit market. New York University's Honorary Professor Gary Marcus referred to the above statement by Goldman Sachs Group, Inc. as "a frightening sentence" when sharing the relevant analysis. He said: For me, the question now is not whether the super-scale model will collapse, but how serious the accompanying damage will be. Gary Marcus further warned: Super-scale cloud service providers cannot recoup their $5.3 trillion investment unless they extract it back from taxpayers through massive government subsidies. That's what they intend to do. Meanwhile, Morgan Stanley estimates that by 2028, the capital expenditure on global data center construction alone will be close to $2.9 trillion, with a significant proportion relying on debt financing. This means that once the market undergoes an adjustment, losses will no longer be limited to shareholders, but may spread through the credit market to the entire society. On the other side of this investment feast is the tightening purse strings of companies. Early adopters of AI such as Uber (UBER.US), Amazon.com, Inc. (AMZN.US), and Walmart Inc. (WMT.US) have set limits on the amount of AI usage for employees or pushed for cost reduction measures. After switching Anthropic's billing model to token-based billing, Carter Busse, Chief Information Officer of software company Workato, saw daily spending skyrocket by seven times, exclaiming: We have created a monster. The $5.3 trillion super-cycle, financing pressures spread to the bond market According to the analysis by analysts at Goldman Sachs Group, Inc., AI capital expenditure is continuing to rise at a faster rate than actual data center construction, indicating that future bottlenecks may shift from model demand to financing capabilities, power supply, and project execution. Morgan Stanley's estimate is more specific. It predicts that by 2028, the $2.9 trillion capital expenditure on global data center construction will come from the following sources: Super-scale cloud enterprises' own cash flow of approximately $1.4 trillion; Corporate bonds of about $200 billion; Asset securitization credit of about $150 billion; Private credit, asset mortgage financing, and joint debt of about $800 billion; Other capital of about $350 billion. This structure implies that AI infrastructure investment is largely driven by credit. AI media personality Rohan Paul pointed out on the X platform that as a few super-scale cloud companies cannot issue bonds to the public bond market without restriction, investors have begun to worry about issuer concentration risk. The complexity of data center financing further exacerbates this issue. It is not a single asset, but a combination of land, power access, network links, buildings, cooling systems, and AI servers, and the financing demand naturally spills over into multiple markets such as infrastructure funds, real estate funds, private credit, and corporate bonds. Once the market undergoes a systemic adjustment, the chain of transmission of losses will be far more complex than during the dot-com bubble. Companies brake, from "unrestrained use" to "AI financial responsibility" On the demand side, the high operating costs of AI are forcing companies to reassess the value of every query and automated workflow. Uber is the most representative case. This ride-hailing giant spent its entire annual AI budget for 2026 in just one quarter. After exhausting it in April, Uber announced a monthly token limit for employee usage of a single AI tool, set at $1,500. Uber's President and Chief Operating Officer Andrew MacDonald admitted: It is becoming increasingly difficult to prove that spending on AI tokens is reasonable, and to draw a clear causal relationship between spending data and actual product functionality improvements. Walmart Inc. has also set limits on the token usage of its internal AI assistants. Suresh Kumar, Global Chief Technology Officer of Walmart Inc., said that usage of the Code Puppy programming platform is "soaring," and now it is time to "take a step back and reassess." Behind this trend is a structural shift in billing models. Major AI labs such as Anthropic and OpenAI have switched some services from fixed subscriptions to token-based billing, making companies more sensitive to the cost of every prompting word and automated process. Costi Perricos, Global Generative AI Lead at Deloitte, said: The cost of computing power is starting to come into view for CFOs and boards. Consumers and companies have been told that AI is cheap or free, but in reality, it is not. Sam Altman, CEO of OpenAI, also admitted this month that costs have become a "significant issue" for customers this year, a topic that was almost never mentioned last year. The contradiction between cost reduction at the enterprise level and laboratory valuation The cost reduction actions at the enterprise level also cannot be ignored for the potential impact on the upstream of the AI industry chain. Both Anthropic and OpenAI plan to go public later this year with valuations close to a trillion dollars. However, the trend of reducing AI expenditures at companies is putting potential pressure on revenue growth expectations for these two companies. Major AI platforms have started to take measures to guide users to use cheaper non-cutting-edge models to maintain adoption rates. Kyle Daigle, Chief Operating Officer of GitHub, stated that Microsoft Corporation has communicated pricing changes with customers in advance, discussing "adaptability and use cases," and emphasizing that "not all tasks require cutting-edge models." Microsoft Corporation, Amazon.com, Inc., and Alphabet Inc. Class C have also launched tools to automatically route user requests to the most cost-effective fitting model. Some companies are turning to open-source models to run on local servers or personal devices in order to reduce costs paid to AI labs and cloud service providers. Patel of Cisco Systems, Inc. summed up the plight of many companies: Our engineers want more tokens, and we have to find a way to pay for them. This statement reflects the dilemma of the entire industry: the strategic value of AI has been widely accepted, but the commercial logic of continuing to pay for it has yet to be tested by the market. This article is reprinted from "Wall Street View". Editor: Jiang Yuanhua.