Global AI-related debt issuance fell to $23 billion in September from a June peak of $113 billion. The abrupt change can look like the capital markets have rejected the data-centre buildout. Morgan Stanley’s interpretation is closer to a pause after unusually front-loaded fundraising, with investors applying a higher selection threshold to the next projects.[1]
That selection phase matters because the financing structure has changed. Cash-rich technology companies, data-centre developers, chip buyers and private infrastructure vehicles are all asking bondholders and lenders to fund assets whose economic life, utilisation and residual value may differ from the debt maturity.
The financing cycle at a glance
| Measure | Value | Context |
|---|---|---|
| June 2026 issuance peak | $113 billion | Morgan Stanley monthly global AI-related financing series reported by the Financial Times.Source 1 |
| September 2026 issuance | $23 billion | Less than half August and roughly one-fifth of June.Source 1 |
| US investment-grade AI issuance in September | $0 | After approximately $306 billion raised from January through August in the cited market segment.Source 1 |
| Global 2026 issuance by end-September | About $466 billion | Financial Times account of Morgan Stanley data.Source 1 |
| 2025 hyperscaler investment-grade debt | More than $100 billion | Morgan Stanley’s published credit analysis.Source 2 |
| Morgan Stanley 2026 hyperscaler net-supply view | $130–150 billion | A narrower hyperscaler investment-grade estimate, not the same universe as global AI-related debt.Source 2 |
| June forecast for global 2026 AI debt | Nearly $570 billion | Reuters reported Morgan Stanley’s broader global estimate.Source 3 |
| Reported SpaceX financing talks | $40 billion | Approximately $10B bank loans and $30B investment-grade debt to buy Nvidia chips; talks may not become a completed deal.Source 4 |
A slowdown can mean the easy money was already raised
Companies raised heavily before September, so a weak month partly reflects timing. It also gives investors room to distinguish issuers with established cash flow and strong balance sheets from projects relying on distant utilisation assumptions. A bond market can remain open while demanding more collateral, covenants, pricing or sponsor equity.
Morgan Stanley’s credit view is deliberately mixed. Near-term risk remains manageable for the largest hyperscalers because they start with strong financial positions. At the same time, elevated capital expenditure compresses free cash flow, raises leverage and makes credit spreads more sensitive to execution and monetisation timelines.[2]
What bond investors are selecting for
| Question | Stronger evidence | Warning signal |
|---|---|---|
| Who repays the debt? | Existing diversified cash flow or a contracted creditworthy customer | A project company dependent on one speculative tenant |
| What is being financed? | Reusable site, power and network assets with clear ownership | Rapidly depreciating chips carrying most of the value |
| When does revenue begin? | Permits, grid connection, construction and customer schedule aligned | Debt service starts long before useful capacity |
| What happens if demand shifts? | Alternative tenants, modular deployment and transferable equipment | A purpose-built asset with no economic second use |
| Who absorbs overruns? | Sponsor equity, contingency and enforceable completion support | Leverage expands whenever construction or power costs rise |
| What is the social licence? | Published water, energy and community agreements | Local opposition or delayed infrastructure with no resolved path |
Hardware life and debt life may not match
A data centre can operate for decades, but accelerators can lose relative value quickly as newer generations deliver more capability per watt. Long-maturity debt secured mainly by rapidly superseded hardware carries a different recovery profile from debt tied to land, power and a flexible building. Financing documents should state which asset and cash flow actually support repayment.
The proposed SpaceX package makes the issue concrete: reported financing would fund Nvidia chips through a mix of bank loans and bonds. The project may have strategic value, but the lender still needs a view on chip delivery, utilisation, customer revenue, technological obsolescence and whether the hardware can be redeployed if the original plan changes.[4]
Selection is healthier than pretending every megawatt has the same economics
A financing pause can slow weak projects and improve terms for strong ones. It can also produce bottlenecks if sound projects lose access because the entire sector is treated as one trade. The appropriate response is greater project-level disclosure: total installed cost, power date, customer contracts, utilisation break-even, sponsor equity and downside recovery.
The financing question is increasingly about timelines, counterparties and physical assets. September’s lower issuance can reflect the heavy fundraising already completed as well as more selective lending. Borrower-level evidence—customer contracts, construction schedules, sponsor support and downside recovery—will distinguish those explanations more clearly than one monthly total.
Sources & further reading
- AI borrowing slows as investors grow wary of debt binge Financial Times · 10 Oct 2026 · Accessed 11 Oct 2026
- AI Dispersion in Credit Morgan Stanley Investment Management · 1 Apr 2026 · Accessed 11 Oct 2026
- Global AI debt issuance to top $500 billion in 2026, Morgan Stanley says Reuters via Investing.com · 10 Jun 2026 · Accessed 11 Oct 2026
- SpaceX seeks $40 billion financing to buy Nvidia chips, sources say Reuters · 6 Oct 2026 · Accessed 11 Oct 2026