AI is reshaping credit markets, with debt behind capital expenditure plans moving off balance sheet into opaque structures. Matthias Dettwiler and Greggory Price unravel the thread of interconnections and assess the underlying credit risks.
In June, Apollo and Blackstone spearheaded the largest private-credit deal on record, roughly USD 35 billion, none of which technically reached Anthropic, the company the debt ultimately rests on.
Instead, the money funded a special-purpose vehicle that bought around a million of Google’s custom AI chips and leased them to Anthropic, whose rental payments are designed to service the debt. Meanwhile, Broadcom supports the senior notes and Google backstops the site leases. In effect, Anthropic (which at the time of writing does not carry a credit rating) added roughly a gigawatt of compute without adding a dollar of debt to its own balance sheet.1 More recently, Nvidia enlisted six Wall Street firms – Apollo and Blackstone among them – to mobilize more than USD 500 billion in much the same way.
Some might argue that it was only a matter of time before the creative financial engineers got stuck into the AI trade. Whether this exercise portends more seismic and systemic issues is hard to say at this stage. What is clear, however, is that the June deal offers a rare, unguarded window into the financial realities of a fast-growing industry that stands to disrupt every corner of the global economy.
Forgive the barrage of data points, but some of the capital expenditure (capex) figures are worth surfacing given the sheer scale involved – even if many of the specifics are already dating as you read this. Consensus now puts hyperscaler capex near USD 940 billion in 2026 and USD 1.3 trillion in 2027, and those figures have been revised upward more than once.2
Figure 1: Hyperscaler Capex
Figure 1: Hyperscaler Capex
The spending splurge extends well beyond raw technology into construction, real estate, utilities and the power supply chain. Crusoe, a US neocloud operator, puts roughly half of the cost of a data center in the compute itself, with power, labor and materials making up the rest, for an all-in figure near USD 60 million per megawatt.
Naturally, massive expenditure like this demands massive borrowing. Net investment-grade issuance from the hyperscaler group has reached about USD 274 billion so far this year, against USD 130 billion in 2025 and USD 40 billion in 2022, the year before ChatGPT launched LLMs into economic orbit.3 In addition to the increase in leverage, the supply technical itself has been a material headwind for spreads. Datacenter specific issuance has topped USD 169 billion.
Figure 2: Hyperscaler 10y basket (bp)
www.ai: Who owes what to whom?
www.ai: Who owes what to whom?
Comparisons between AI and the internet bubble of the late 1990s and early 2000s have come thick and fast. Parallels are always useful as much for the differences they reveal as for the similarities they uncover. Circular financing itself is not a new phenomenon, but the scale of this cycle is unmatched and combined with the uncertainty around the useful lives of the assets being financed creates a new dynamic.
As the Anthropic deal highlights, the same firms often appear as customer, supplier, guarantor and creditor at different points in the capital and relationship stack. Still, the labs themselves are the primary point of weakness in the system given their lack of profitability and intense competition at the model layer.
OpenAI and Anthropic are some of the largest users of compute and the most likely to push the frontier, yet by traditional standards they are weak credits. Indeed, OpenAI reportedly burned USD 3.7 billion in the first quarter of 2026 alone and has at times carried commitments near USD 1.4 trillion4,5, and on its own guidance it will not be free-cash-flow neutral until 2030 at the earliest.6 Meanwhile, Anthropic’s lack of a credit rating is precisely why its debt had to be wrapped in a Broadcom guarantee to qualify as investment grade.
Concentration runs through the hardware as well. A narrow group of designers, foundries and packagers can reprice the whole stack if supply stays tight, and a single guarantor now stands behind tens of billions of chip-lease debt. The growing use of Google’s TPUs, Amazon’s Trainium and other in-house silicon keeps market shares in flux, which may improve the economics of compute over time but can also erode the residual value on which these leases depend.
The practical implication for investors is that AI exposure should be analyzed not only at the issuer level but at the network level. For credit where investors have limited upside in returns, yields need to compensate for the compounding risks in the system.
It is now therefore imperative to figure out who depends on whose compute, power, data and financing, what support is explicit rather than assumed, and how quickly contracts reset if the economics shift. Strong growth and tight supply improve the near-term underwriting backdrop, but the quality of the debt still depends on where the investor sits in that network and which risks the documents actually transfer. Mismatches in contractual maturities across the stack can also amplify the potential impacts of this circularity.
There may never be a single point of failure in the system. However, the clearest credit risks ultimately lead back to the research labs. Circular financing arrangements may distort reported revenues, while failed equity investments in customers and partners would weigh on valuations; however, most of the system’s new leverage does not depend directly on either. Debt issued by increasingly leveraged cloud operators and associated data centers would be particularly exposed if a research lab failed to scale profitably. The second-order effects would depend heavily on the catalyst and could range from an orderly adjustment to more severe contagion.
Opacity is becoming a major challenge, though. By our calculations, once off-balance-sheet commitments are included, aggregate hyperscaler obligations rise from about USD 800 billion to roughly USD 3.5 trillion. Structures like the Anthropic vehicle effectively move the hardware, as well as the debt, off the operating company entirely but replace them with lease debt and other commitments found elsewhere in company disclosures. Still, the private nature of the labs leaves more debt unaccounted for.
Figure 3: Hyperscaler obligations – more than meets the eye
Public and private markets were already coming closer together, but recent events are arguably smudging the ledger beyond recognition. To adapt effectively, credit portfolios need to diversify by market and by common counterparty as much as they do by sector or rating. And transparency will be key to underwriting assessments as balance sheets continue to show only half the story.
From supply to demand
From supply to demand
Looking beyond the corporate fundamentals, all of this financing rests on the assumption that demand for compute holds at or above supply through the life of the debt. And recent talk of firms reselling capacity has raised the prospect of a glut, with more capacity supposedly due to come online.
So far, however, the evidence points the other way. Resale contracts are short-dated and priced at a premium, which looks like buyers paying up for scarce access rather than dumping surplus, and Meta’s continued spending on frontier models is hard to square with a company that thinks the race is over. Some idle capacity is simply the gap between training runs. The balance can shift as new sites come online, but for now the shortage of high-end compute is real.
Figure 4: Token pricing indices
Creditors should take note because scarce computing capacity keeps data centers busy, strengthens the position of operators when contracts are renewed and makes access to sites with sufficient power more valuable. It would be naive to underwrite today’s pricing forever, but the current imbalance creates a firmer near term base than the headline capacity numbers imply.
The issue we keep turning over in our minds is whether demand, pricing and contractual protection are enough to carry an issuer through to the point where its cash flow improves. In the near-term, proof of adequate returns on capital and favorable unit economics can be enough to be that bridge.
Oracle is a useful yardstick here. S&P cut it to BBB- with a stable outlook after conceding it had underestimated the capital its AI ambitions would need. Its remaining performance obligations have jumped to USD 638 billion from USD 138 billion a year earlier, providing real visibility into future revenue. However, much of this amount depends solely on OpenAI. This backlog also carries customer prepayments and bring-your-own-hardware terms that cut what Oracle must fund itself, and management has been clear it means to hold its investment-grade rating. Although it could be a while before cash flow turns, contracted growth, customer funding and scarce capacity do give the underwriting case real support.
Other risks are lingering in the credit markets downstream from the success of AI. In high-yield software, many balance sheets were designed for a lower-rate environment. With public enterprise values falling, concerns around the sustainability of these capital structures have increased. European spreads have largely retraced their widening from January, while US high-yield technology is approximately 40 basis points wider than the index after having been nearly 100 basis points wider. The recovery in spreads does not mean the business-model questions have been resolved; it means that investors are again being asked to distinguish between temporary volatility and a lasting impairment of earnings.
A similar tension is visible in investment grade. For US software companies such as Salesforce and Adobe, where the perceived impact of AI may be larger, market capitalizations and bond spreads have seen a spike in volatility but cash flow remains robust and may do so for some time.
Savvy investors will be wondering how incumbents might fare and, for us, three business-model defenses stand out as key gauges. First is proprietary data that customers must keep relying on. Second is regulatory and domain complexity that slows enterprise-wide automation. Third is trust in services where an outage or a bad payment costs far more than the service itself. Each buys time to adapt rather than immunity.
The new capital cycle
The new capital cycle
None of what we are seeing today is entirely unprecedented. The telecom build-out of the 2000s and the US shale boom of the 2010s were both financed heavily from outside, and both taught lenders hard lessons about supply.
A more direct comparison is arguably the vendor-financed leasing GE Capital pioneered for aircraft in the 1960s, whereby the maker of the asset helps fund it and keeps a claim on its residual value.
Today’s AI version clearly has similar elements, but a completely different starting point. Hyperscalers carry stronger balance sheets, real cash flows and diversified businesses, and the demand they are funding is visible in cloud revenue, contracted backlogs and a persistent shortage of compute. That is a much firmer footing for creditors than the raw issuance figures alone would suggest.
Stepping back in this way allows for some perspective. AI is loading leverage onto parts of the economy, concentrating dependence on a handful of suppliers and speeding the churn between winners and losers, all at once. For creditors the backdrop is constructive as much as cautionary. Tight supply supports pricing, and take-or-pay contracts, prepayments, guarantees and amortization can all firm up the downside. And while they do not erase duration, counterparty or technology risk, they can make the debt worth owning where the structure matches the useful life of the asset and the lender is paid for what remains uncertain.
Issuers to avoid are the ones funding long-lived infrastructure with short-lived technology and thin protection, overexposure to a given counterparty at such an early stage, or leaning on AI to defend revenue without the balance-sheet room to adapt. There are also considerable regulatory and construction risks that can emerge and are not typically as acute in non-AI underwriting.
The task is now to find where in the web strong growth and scarce capacity provide enough support to justify the financing risk.
Within this framework, we look for issuers that would not be adversely affected by a closing of the financing window and/or an imminent break in model scaling laws. We do see pockets of value in highly strategic data center debt with low energy risk and reputable construction histories and higher-yielding bonds tied to compute that can outperform if pricing stays tighter than the market expects. While the hyperscalers may continue to suffer from supply headwinds, we believe that consensus estimates for capex have likely caught up to reality and should alleviate some of the pressure going forward, but the absolute levels of spreads remain tight.
1 Apollo and Blackstone complete USD 35 billion private-credit deal to fund Anthropic’s compute expansion, Financial Times, June 2026.
2 Bloomberg Consensus Estimates, August 2026
3 Bloomberg Barclays indices, July 2026.
4 OpenAI burned USD 3.7 billion in first quarter of 2026, The Information reports, Reuters, 16 June 2026.
5 OpenAI data center pivot underscores Wall Street IPO concerns, CNBC, 22 March 2026 (NB. the USD 1.4 trillion commitments figure originates from Sam Altman directly).
6 OpenAI resets spending expectations, tells investors compute target is around USD 600 billion by 2030, CNBC, 20 February 2026.
7 Oracle Corp. downgraded to BBB- on elevated AI-related capital spending; outlook stable, S&P Global Ratings, 9 July 2026.
The Red Thread
Disruption 2.0
Disruption 2.0
Code: S 09/26 M-006903
About the authors
Matthias Dettwiler
Head of Active Fixed Income, Head of Investments UBS Asset Management Switzerland AG
Matthias Dettwiler is Head of Active Fixed Income at UBS Asset Management. Prior to this, Matthias was Head of Index Fixed Income from 2012. Matthias joined UBS in 1995 on a three-year banking apprenticeship and joined fixed income portfolio management in 2000. Matthias holds the ‘Certified International Investment Analyst’ designation and the ‘Swiss Federal Diploma for Expert in Finance and Investment’.
Greggory Price
Technology Credit Research Analyst Fixed Income
Greggory Price is a senior credit research analyst at UBS Asset Management, covering global investment grade and high yield technology issuers. He has 17 years of industry experience, including prior roles as a senior credit analyst at Barclays Investment Bank in New York and Lombard Odier in Geneva. Greggory is a CFA charterholder.
