Debt Doesn’t Hallucinate, Part 2: Recent Developments in AI Debt Markets

Markets were volatile this week as investors weighed rising oil prices, higher Treasury yields and persistent inflation concerns against resilient economic activity and continued strength in AI spending. The 10-year Treasury yield briefly reached 5.34% on Thursday, its highest level since 2002, as elevated oil prices and concerns around government borrowing pressured global bond markets (Reuters, 10/1/2026). Yields eased from those highs later in the day, while August inflation data released on Wednesday came in softer than expected (CNBC, 9/30/2026). Through Thursday, the S&P 500/Nasdaq were -1.0%/-0.7%, respectively, for the week (AP, 10/1/2026).

Technology and AI remained a key focus. Higher yields pressured valuations across semiconductor and other growth stocks earlier in the week, while investors continued to debate the sustainability and returns on rapidly increasing AI infrastructure spending. At the same time, underlying AI demand remained strong, with Micron highlighting significant long-term customer commitments. Despite recent volatility, the S&P 500 remains up ~12% YTD. Our NPM Private Market Tracker*, which shows the average price performance of the 50 largest names in our internal NPM Price data, is +90% YTD. (AP, 10/1/2026; Reuters, 9/30/2026; NPM Data)

AI Debt Gets Its First Real Stress Test

We are writing this as a follow up to our 8/7 piece, “Debt Doesn’t Hallucinate.” Since our earlier piece on AI debt, the market has softened, with benchmark AI infrastructure bonds down ~$5 since that time.

• The Meta-backed “Beignet Investor” ($27.3bn; A+ rated) bonds currently trade at a price/yield of $92.1/7.5% vs. $97.8/6.9% on 8/7 and a yield at pricing in 10/25 of 6.6%.
• The Meta-backed “Sopaipilla Investor” ($12.5bn; A+/AA- rated) bonds currently trade at $99/7.7% vs. $104.7/7.1% at the time of our last piece and a yield of 7.5% at issue in 7/26 (PitchBook, 7/28/2026).
• In the senior unsecured HY market, data center bellwether CoreWeave 9.75% of 2031 bonds ($2.75bn; B1/B) currently trade at $87/13.4% vs. $92/11.9% at the time of our last piece. CRWV equity is -2% over the same time period.

And while the market softening over the past month has been notable, another emerging story is the changing terms on which investors are willing to provide financing. From our perspective, capital still remains abundant, but lenders are demanding higher yields, stronger contractual protections and greater visibility into who ultimately bears construction, power, technology and residual-value risk. According to the Bank of England, citing Morgan Stanley estimates, ~$450bn of AI-related debt is expected to be issued globally in 2026 (roughly double the 2025 level) (Reuters, 9/30/2026), and this increase has not come without greater selectivity from investors. In this weekly, we wanted to highlight what we view as 4 key developments in the AI debt financing market over the last month.

Development #1: Debt Market is Increasingly Focused on Execution and Supply Chain Risk

The clearest catalyst has been Project Jupiter, the 2.45GW Stargate-linked data center in NM being developed by STACK Infrastructure (Blue Owl) for Oracle to provide compute for OpenAI. On 9/24, it was reported that Oracle had sent a force majeure notice to Blue Owl, seeking to delay payments on the data center due to permitting issues related to a pipeline needed to bring natural gas to the site (Bloomberg, 9/24/2026). Oracle and Blue Owl equities each fell ~4% on the news (Investing.com, 9/24/2026), while loans related to the project had already been quoted around 89-91 cents on the dollar the prior week (Financial Times, 9/18/2026). Blue Owl has said the notice does not change the financial commitments, and Oracle has said the project remains on its planned schedule (Data Center Dynamics, 9/24/2026). To us, the problem is notable because demand for the capacity itself is not the central concern. Rather, investors are confronting a much more basic infrastructure issue: enormous AI contracts have limited value if the data center cannot obtain electricity or other supply chain items on schedule.

The episode appears to be modifying how the market thinks about AI infrastructure. During the first phase of the “boom,” the involvement of a hyperscaler or major AI laboratory meant markets treated these projects as almost riskless: long-term contracts backed by bulletproof customers and enormous expected utilization. Project Jupiter has reminded investors that the credit chain is considerably more complicated.

Lenders increasingly want to understand exactly when lease payments begin, who absorbs construction delays, whether power has actually been secured, what constitutes force majeure, who funds cost overruns and whether debt service continues if the facility is completed later than expected. In other words, execution and energization risk are becoming nearly as important as tenant credit risk. We have seen similar parallels in other industries, such as the offshore contract drillers in the mid 2010s, where seemingly “bulletproof” contracts are questioned by the market for reasons perhaps as simple as the fact that they had never been legally tested.

Development #2: GPU Lending is Becoming More Institutionalized

Another important development this month is the further institutionalization of GPU-backed lending. Rather than lending solely against a data center and its contracts, banks and private-credit funds are increasingly financing the computing equipment itself. On 9/16, Reuters reported that a consortium of banks is providing a ~$22bn chip loan to Blackstone and Alphabet’s AI cloud venture, following a $35bn chip loan (Apollo/Blackstone) in 6/26 that was among the largest private credit deals on record (Bloomberg, 6/5/2026). These structures essentially transform contracted compute capacity into an asset-backed credit product.

One market question for these deals has been residual value of chips. Unlike a bridge or toll road, GPUs can be superseded by new chip generations several times within the term of a loan. Lenders therefore increasingly want GPU debt to fully amortize within the term of the associated customer contract rather than depending upon an uncertain residual value. Additionally, Nvidia is reported to be working with insurers on products to protect lenders against losses on loans backed by AI chips (Financial Times, 9/29/2026), potentially allowing a much larger pool of institutional capital to finance GPU purchases.

Development #3: Off Balance Sheet Leverage Receiving More Attention

Hyperscalers and semiconductor companies increasingly support data center and compute financings through off balance sheet arrangements, which can include residual-value guarantees, lease commitments and other contractual arrangements vs. conventional corporate borrowing. These structures can lower project financing costs while keeping significant debt out of corporate leverage statistics. For instance, under GAAP a lease that has not yet commenced (and nine large technology companies have ~$1.2tn of them, per 24/7 Wall St, 8/22/2026) is kept off balance sheet until the lease commences. Total off balance sheet commitments for these companies are estimated at ~$3tn (Wall Street Journal, 8/17/2026).

Why does this matter? This matters because it leads to higher correlation in portfolios when multiple investments can ultimately depend upon the same underlying economic exposure. For example, an investor might own a hyperscaler’s bonds, project debt for a data center leased to that hyperscaler, securities issued by the data center developer and private credit exposure financing the GPUs inside the building without realizing they could all essentially be the same bet.

Ultimately it all comes down to the underlying earnings power of AI. For Microsoft, Alphabet, Amazon, Meta and the other hyperscalers, balance sheet leverage is currently very low (CNBC, 7/24/2026). The question, however, is whether hyperscaler AI capex, which is projected to exceed $1tn annually by 2027 (Bloomberg, 9/25/2026), will ultimately generate returns needed to support existing debt, new debt, and ~$3tn of off balance sheet commitments. Moreover, for neoclouds and other leveraged AI infrastructure companies, the market is frequently underwriting future EBITDA that does not yet exist. Those companies must raise capital to purchase GPUs and build capacity before they can deliver the contracted computing services that generate the revenue used to service the debt.

Development #4: Macro Environment has Become More Difficult

The U.S. 10-year Treasury yield increased ~50bps during September to around 5.3% from 4.8% entering the month (Reuters, 9/30/2026), raising the absolute cost of capital across the AI infrastructure stack. This matters disproportionately for projects requiring enormous upfront investment against cash flows expected over ten or twenty years. When investors can earn >5% on government securities, a complicated data-center project involving construction, power, technology and residual-value risk must offer a meaningfully higher return. We estimate the move in Treasuries alone accounted for about half (using change in spread * duration) the weakness in AI Infrastructure bonds we have observed since our last piece on the topic published Aug 7.

Nonetheless, the market continues to demonstrate a substantial capacity to finance AI. In September, the market digested ~$22bn of chip-backed financing for Blackstone and Alphabet (Reuters, 9/16/2026), $11bn of HY bonds for SoftBank to help finance its AI investment program (Reuters, 9/24/2026) and $2.3bn of HY bonds for CleanSpark Inc., which is building a data center for Meta (company press release, 9/18/2026). The Ba2 rated CleanSpark bonds priced on 9/18 and currently trade at $99/8.1%, and to us represent a good indicator of cost of capital for sr secured HY risk wrapped by a hyperscaler. The important signal is therefore not that financing has stopped, but that investors increasingly differentiate deals based on structure and ultimate credit support. What has changed is the price of uncertainty.

As we enter the fourth quarter we believe the variable to watch is credit dispersion more so than aggregate issuance. If Microsoft, Alphabet and Meta supported projects continue pricing easily while more speculative neoclouds and/or poorly structured projects do not, the credit market is simply imposing greater underwriting discipline. To us, a more consequential warning would be persistent distribution problems and higher costs of capital even for well-structured, strongly contracted hyperscaler-backed projects.

Private Market Performance by Sector