Global artificial intelligence-related debt issuance has fallen dramatically, dropping from $113 billion in June to just $23 billion in September, according to Morgan Stanley data cited by the Financial Times. The nearly 80% decline marks a sharp reversal for an industry that spent much of 2026 raising record amounts of capital.
September's volume was less than half of August's total, and U.S. investment-grade AI-related bond issuance stopped entirely during the month. That pause followed roughly $306 billion in borrowing by major technology companies between January and August. For the first nine months of 2026, AI-related borrowing still reached approximately $466 billion, reflecting how aggressively companies funded expansion earlier in the year.
Morgan Stanley attributed much of the September weakness to the earlier borrowing surge, but higher interest rates and growing scrutiny of data center economics have also complicated new financing. Investors are increasingly questioning how quickly billions invested in AI infrastructure can translate into revenue, particularly as borrowing costs rise and construction delays threaten project economics.
Oracle offers a concrete example of the pressure. About $18 billion in loans backing Project Jupiter, a New Mexico data center campus associated with Oracle's OpenAI infrastructure commitments, were quoted at just 89–91 cents on the dollar in September. Banks including Santander and Jefferies reportedly struggled to distribute the loans as investors reassessed construction delays, Oracle's expanding debt burden, and uncertainty surrounding the project's power infrastructure.
The financing difficulties followed S&P's July downgrade of Oracle to BBB-, the lowest investment-grade rating. Trading below face value does not mean Oracle has defaulted; it indicates that investors are demanding a discount to assume the exposure. The situation reflects a broader problem for AI infrastructure stocks, where high revenue expectations must eventually cover financing expenses and capital expenditure.
The stress is emerging even as AI companies sign enormous long-term contracts. In April 2026, Anthropic announced a commitment to spend more than $100 billion on Amazon Web Services over ten years, securing access to as much as five gigawatts of computing capacity. Microsoft separately disclosed that OpenAI had contracted to purchase an additional $250 billion in Azure services.
These agreements generally establish future spending obligations, capacity reservations and commercial terms. Payments may occur as infrastructure becomes available, services are delivered, or contractual milestones are reached. A hypothetical $100 billion ten-year contract would average $10 billion annually if spending were evenly distributed, though actual payment schedules can differ substantially.
Cloud providers and infrastructure operators often finance equipment and construction before collecting most customer payments. CoreWeave, for example, carries about $104 billion in contracted revenue backlog but also reported $640 million in quarterly net interest expense, illustrating how financing costs remain substantial even when contracted demand is strong. A signed contract does not eliminate financial risk: contract value is not revenue, revenue is not cash flow, and future commitments are not guaranteed profits.
If an AI customer cannot pay, consequences depend on the contract. Some agreements contain minimum spending requirements or take-or-pay provisions, meaning customers may owe money even when they use less computing capacity than expected. Others include termination rights, performance conditions or negotiated remedies. Infrastructure operators could face revenue shortfalls while still owing lenders and equipment suppliers, and GPU collateral may become central to restructuring negotiations.
