AI companies are borrowing increasingly large sums of money to build the computing facilities they require. Although sales are booming, questions remain over whether revenues will be able to keep pace with mounting infrastructure costs. The outcome matters beyond individual balance sheets, because this borrowing wave now underwrites data center construction and chip purchases across the industry.

The scale of that gap is visible in Anthropic's financial performance. According to Reuters, the company recorded $4.6 billion in revenue in 2025 alongside an operating loss of $8.06 billion.

Its finances have improved since then. CNBC reported that Anthropic's annualized revenue run rate reached $65 billion in July 2026, with projections pointing to $100 billion by year-end. 24/7 Wall St. also reported a positive adjusted operating result in the second quarter.

OpenAI, meanwhile, earned $6.7 billion in revenue in the second quarter but faced a loss of $12.3 billion on its operations. By September, its annual revenue was heading toward $70 billion, as reported by Axios.

Anthropic is considered to be closer to achieving profitability, although its adjusted results exclude certain costs — and rapid revenue growth does not necessarily translate into sustainable profits.

AI borrowing exposes growing financial risks

By early August 2026, Goldman Sachs estimated approximately $500 billion in AI-related debt issuance. CNBC reported that JPMorgan anticipates $4.1 trillion in total issuance through 2030. On those figures, what has already been borrowed represents only a fraction of what lenders expect over the remainder of the decade.

It is reported that Broadcom is seeking more than $50 billion to finance OpenAI chips, while SpaceX is pursuing an additional $40 billion. Oracle is also searching for funding for its data centers. Broadcom is additionally expected to provide Anthropic with $42 billion for its chip developments.

These arrangements distribute financial risk across lenders, suppliers, and AI developers. As a result, the financial stability of AI labs is becoming increasingly relevant to broader lending market, since strains at a single lab or its chip suppliers would reach the creditors financing them.

Borrowing money, however, is only part of the problem. Infrastructure costs do not end with construction. Bryan Musto estimates yearly depreciation costs of between $390 billion and $530 billion for the five largest hyperscalers by 2028. Depreciation reflects the declining value of servers and chips over their working lives, so the charge recurs annually even after facilities are built. His model shows that their annual revenues would need to fall within the range of $675 billion and $785 billion in order to cover depreciation and earn a 10% pretax return — excluding electricity and staffing.

Debt boom could reshape global AI competition

According to estimates by PwC, total global investment in data centers could reach $31.6 trillion by 2050. In a separate analysis, another paper examines potential scenarios under which AI-related debt could reach anywhere from $3 trillion to $7 trillion in 2029 — projected figures rather than actual borrowings.

The increasing dependence on borrowing is giving rise to concerns about financial stability. The Bank for International Settlements warns that a downturn in AI investment could impact the wider financial market. Small, heavily indebted firms would be the most affected, while tech giants with stronger finances would be able to gain additional ground from the situation.

Nevertheless, a debt crisis is by no means guaranteed. The far bigger question is whether AI businesses can generate enough revenue to cover their debts while simultaneously continuing to absorb the cost of expensive computing infrastructure. The next checkpoints are concrete: whether Anthropic's projected $100 billion run rate holds by year-end, whether the debt packages pursued by Broadcom, SpaceX, and Oracle reach completion, and how credit markets continue to price the growing volume of AI-linked bonds.