As enterprises begin to focus on "computing power cost reduction," Goldman Sachs warns that 5.3 trillion in AI capital expenditure is approaching credit saturation!

The AI infrastructure investment boom is reshaping the global capital markets landscape, with hidden debt risks that cannot be ignored.

According to the latest predictions from Goldman Sachs, from 2025 to 2030, the capital expenditures by hyperscale cloud computing firms in AI and data center fields will total $5.3 trillion, marking an unprecedented supercycle of capital spending.

Goldman Sachs expects that these hyperscale companies will have to seek financing from various markets because they may face constraints due to saturation in the liquid credit markets.

NYU Honorary Professor Gary Marcus, when sharing related analysis, referred to Goldman’s statement as a "frightening sentence," saying:

For me, the question is no longer whether the hyperscale model will break, but rather how severe the collateral damage will be.

Gary Marcus further warned:

It will be impossible for these hyperscale cloud service providers to recover their $5.3 trillion investment unless they extract it from taxpayers through massive government subsidies. That is exactly what they intend to do.

Meanwhile, Morgan Stanley estimates that the capital expenditure for global data center construction alone will reach nearly $2.9 trillion by 2028, with a relatively large proportion relying on debt financing. This means that, once the market undergoes an adjustment, losses will no longer be limited to shareholders but could spread to society at large via credit markets.

The flip side of this investment feast is the increasingly tighter corporate budgets. Early large-scale AI adopters such as Uber, Amazon, and Walmart have already imposed caps on employees’ AI usage or are pushing for cost reduction measures.

After Anthropic switched its billing model to token-based pricing, Workato’s Chief Information Officer Carter Busse saw the company’s daily spending soar sevenfold and couldn’t help but remark:

We’ve created a monster.

According to analysts at Goldman Sachs, AI capital spending is climbing at a faster pace than actual data center construction, which means future bottlenecks may shift from model demand to financing capacity, electricity supply, and project execution.

Morgan Stanley’s estimates are more detailed. They expect that by 2028, out of the $2.9 trillion in global data center construction capex, the funding sources will be as follows:

• About $1.4 trillion from hyperscale cloud firms’ own cash flows;
• About $200 billion in corporate bonds;
• About $150 billion in asset-backed credit;
• About $800 billion in private credit, asset-backed financing, and joint venture debt;
• About $350 billion in other capital.

This structure means that AI infrastructure investment is, to a significant extent, credit-driven.

AI social media commentator Rohan Paul pointed out on X that, since only a handful of hyperscale cloud firms cannot issue public bonds without limit, investors have begun to worry about issuer concentration risk.

The complexity of data center financing further exacerbates this problem.

It’s not a single asset, but a combination of land, electricity access, network links, construction, cooling systems, and AI servers, and thus the financing demand spills over into infrastructure funds, real estate funds, private credit, and corporate bonds across several markets.

Should there be a systemic market correction, the transmission chain of losses will be significantly more complex than during the dot-com bubble era.

On the demand side, the high operating costs of AI are forcing companies to reconsider the value of every query and automated workflow.

Uber is the most representative case. Wallstreetcn mentioned that this ride-hailing giant exhausted its entire 2026 annual AI budget in just one fiscal quarter.

After finishing it off as early as April, Uber announced a $1,500 monthly cap on employee usage (by tokens) for any single AI tool. Uber President and COO Andrew Macdonald admitted:

It’s getting harder to justify spending on AI tokens and difficult to draw a clear causal line between such spending and actual product feature improvements.

Walmart has also set caps on internal AI assistant token usage. Walmart Global CTO Suresh Kumar noted that usage of the company’s Code Puppy programming platform “skyrocketed,” and now it’s time to “take a step back and reassess.”

The trend is backed by a structural shift in billing models. Major AI labs like Anthropic and OpenAI have shifted some services from fixed subscriptions to token-based billing, making enterprises much more sensitive to the cost of every prompt and automated process.

Deloitte Global Generative AI Head Costi Perricos says:

The cost of computing power has started drawing the attention of CFOs and company boards. Consumers and enterprises have always been told that AI is cheap or free, but that’s just not the case.

OpenAI CEO Sam Altman also acknowledged this month that costs have become a “major challenge” for customers this year—a topic that was rarely mentioned last year.

Cost-reduction actions at the corporate level are also exerting significant pressure upstream in the AI industry chain.

Both Anthropic and OpenAI are planning IPOs later this year, with valuations close to $1 trillion. However, the trend of enterprises cutting AI expenditures is putting potential pressure on revenue growth expectations for these companies.

Major AI platforms have started to respond by steering users towards more affordable, non-frontier models to sustain adoption rates.

GitHub COO Kyle Daigle revealed that Microsoft has proactively communicated pricing changes to clients, discussing “adaptability and suitable scenarios,” and stressed that “not every task requires a frontier model.”

Microsoft, Amazon, and Google have also launched tools that automatically route user requests to the most cost-effective model for their use case.

Some firms have shifted to open-source models, running them on local servers or personal devices to cut payments to AI labs and cloud providers.

Cisco’s Patel expressed the predicament faced by many enterprises:

Our engineers want more tokens, and we have to figure out how to pay for them.

This statement reflects the quandary facing the entire industry: the strategic value of AI is broadly recognized, but the business case for continually paying for it still faces market scrutiny.