# AI infrastructure is moving onto credit markets—and into harder-to-see obligations > Source: > Published: 2026-08-14 11:05:13.736101+00:00 # AI infrastructure is moving onto credit markets—and into harder-to-see obligations - Nvidia said financing platforms being developed with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR could mobilize more than $500 billion of third-party capital. That is a target for potential financing, not capital already raised. [[1]](https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital)[[2]](https://www.axios.com/2026/08/11/nvidia-chip-securitization-wall-street) - Oracle disclosed $260 billion of additional lease commitments, substantially related to data centers, that had not yet appeared on its balance sheet as of May 31, 2026. [[3]](https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm) - Meta reported $182.88 billion of future lease obligations for data centers, colocation and network infrastructure, plus $237.67 billion of non-cancelable contractual commitments, as of March 31, 2026. [[4]](https://www.sec.gov/Archives/edgar/data/1326801/000162828026028526/meta-20260331.htm) - Morgan Stanley estimates that AI-related debt issuance could reach about $500 billion in 2026, while its broader research framework projects a $1.5 trillion financing gap for global data-center spending through 2028. Those are estimates, not official market totals. [[5]](https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/ai-debt-surge-carolyn-campbell-vishwas-patkar)[[6]](https://www.morganstanley.com/content/dam/msdotcom/en/assets/pdfs/Research_Bridging-Data-Center-Gap.pdf) - The main risks are demand, refinancing, technology obsolescence, power constraints and concentration among lenders and tenants—not evidence of an immediate systemic credit event. [[7]](https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026)[[8]](https://www.bis.org/publ/qtrpdf/r_qt2603u.htm) The financing behind the artificial-intelligence infrastructure boom is becoming as complex as the hardware it supports. Companies are raising bonds, signing long-term capacity leases, borrowing against data-center cash flows and turning to private-credit funds and special-purpose vehicles to build facilities and acquire GPUs. [[7]](https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026)[[8]](https://www.bis.org/publ/qtrpdf/r_qt2603u.htm) That expansion gives developers and AI companies access to more capital. It also makes total obligations harder to see in any single balance sheet or debt statistic. A lease signed by a cloud company, a loan issued to a data-center operator and a securitization backed by tenant payments can all fund the same physical campus while placing the risks with different investors. [[7]](https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026)[[9]](https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/ai-financing-credit-markets-data-centers-debt-lindsay-tyler-anish-shah) ## The headline figures are targets, commitments and debt—and they are different things Nvidia said in August that it had partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish independent financing platforms that could mobilize more than $500 billion of third-party capital for AI infrastructure over time. Nvidia said the investors would make independent financing decisions. [1] The announcement does not mean Nvidia has raised $500 billion, nor does it identify a single pool of debt with one maturity schedule or one set of borrowers. [[1]](https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital) [[2]](https://www.axios.com/2026/08/11/nvidia-chip-securitization-wall-street)The industry now uses several overlapping forms of finance. Investment-grade bonds can fund a hyperscaler directly. Project-finance loans can fund a campus against contracted lease payments. Asset-backed or commercial-mortgage securitizations can package those cash flows for institutional investors. Private-credit funds can provide construction loans, equipment finance or bridge facilities before a project reaches the public bond market. [[7]](https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026)[[9]](https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/ai-financing-credit-markets-data-centers-debt-lindsay-tyler-anish-shah) Morgan Stanley reported in June that AI-related debt issuance was approaching $250 billion in 2026 and said it expected the figure to double for the full year. The firm’s estimate includes multiple forms of financing and should not be treated as a standardized industry series. [[5]](https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/ai-debt-surge-carolyn-campbell-vishwas-patkar) A separate Morgan Stanley research report projects roughly $2.9 trillion of global data-center capital spending through 2028, including chips, servers and facilities, with approximately $1.5 trillion to be supplied by external capital after estimated hyperscaler cash flows. It projects an $800 billion private-credit opportunity, $150 billion of data-center ABS and CMBS issuance, and $200 billion of incremental investment-grade technology bonds. These are forecasts, not committed funding. [[6]](https://www.morganstanley.com/content/dam/msdotcom/en/assets/pdfs/Research_Bridging-Data-Center-Gap.pdf) ## Lease commitments are becoming a parallel financing channel Oracle’s 2026 annual report illustrates how quickly obligations can grow outside conventional debt measures. As of May 31, Oracle had $260 billion of additional lease commitments, substantially all related to data-center arrangements, expected to begin between fiscal 2027 and fiscal 2029. The commitments generally run for 15 to 19 years and were not included on Oracle’s consolidated balance sheet at that date. [[3]](https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm) Oracle also warned that if customer demand or customer payment performance falls short, it could be locked into multiyear data-center commitments and related financing without receiving corresponding revenue. The filing identified a guarantee of up to $3.3 billion connected to one of the leases. [[3]](https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm) Meta’s filings show a different combination of obligations. The company reported approximately $182.88 billion of lease obligations for data centers, colocation and network infrastructure that had not yet commenced as of March 31, 2026. It also reported $237.67 billion of non-cancelable contractual commitments, mostly connected to cloud capacity, servers, network infrastructure and data centers. [[4]](https://www.sec.gov/Archives/edgar/data/1326801/000162828026028526/meta-20260331.htm) Meta separately disclosed a $12.31 billion initial lease commitment tied to a data-center venture scheduled to begin in 2029, as well as residual-value guarantees with an aggregate threshold of about $28 billion. Meta said those guarantees were not probable and therefore no liability had been recorded. [[4]](https://www.sec.gov/Archives/edgar/data/1326801/000162828026028526/meta-20260331.htm) These commitments are not equivalent to funded debt. They may be supported by growing revenue, transferred to another tenant or reduced under contractual conditions. But they create fixed or semi-fixed claims on future cash flow. If a project is delayed, a customer defaults or computing demand shifts to newer technology, the obligation can remain after the expected revenue has weakened. ## Private credit and structured finance move risk away from the largest balance sheets The attraction of these structures is straightforward. A hyperscaler or AI company can commit to use capacity without paying the full construction cost upfront. The developer can borrow against that contracted revenue. A bank or private-credit fund can later refinance the project or sell an asset-backed security to insurance companies, pension funds or other institutional investors. [[7]](https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026)[[9]](https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/ai-financing-credit-markets-data-centers-debt-lindsay-tyler-anish-shah) The Bank for International Settlements described a similar pattern in 2026: hyperscalers are using off-balance-sheet arrangements, often with private-credit partners, while debt is serviced by lease cash flows and held by private funds and other institutional investors. Some deals include guarantees or contractual support from highly rated technology companies. [[8]](https://www.bis.org/publ/qtrpdf/r_qt2603u.htm) The Bank of England said AI companies’ use of leveraged finance, structured finance and private credit was growing rapidly in the first half of 2026. It also pointed to securitized data-center structures, special-purpose vehicles and bespoke arrangements that broaden the footprint of AI debt across the credit system. The bank said AI-related debt remained a small share of total outstanding debt entering 2026. [[7]](https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026) The structures are not inherently fragile. A long-term lease with a financially strong tenant, a powered site in a constrained market and equipment with durable resale value can support relatively conservative lending. S&P Global estimated about $50 billion of data-center securitization financing between 2021 and September 2025, and roughly $50 billion of data-center projects funded by private markets between 2021 and May 2025. [[10]](https://www.spglobal.com/en/research-insights/special-reports/look-forward/data-center-frontiers/data-center-risk-if-ai-promises-fade) The concern is aggregation. Investors may evaluate each loan as a contractually protected infrastructure asset while underestimating how many borrowers, tenants, lenders and projects ultimately depend on the same assumptions about AI adoption, power availability and GPU utilization. That is a risk concentration problem, not proof that the structures replicate the mortgage market before 2008. ## The weak points are concentrated in the cash-flow assumptions The first risk is demand. Data centers built for AI require large upfront spending, but their economics depend on customers paying for compute over many years. If model training becomes more efficient, inference prices fall or customers consolidate workloads, operators may struggle to renew leases at the rates used to justify construction debt. The Bank of England identified the pace of AI debt growth and its links to the wider financial system as areas requiring monitoring. [[7]](https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026) The second risk is refinancing. Construction loans often mature before a campus has a long operating record. A project may need to refinance into bonds or securitization markets at a time when interest rates, credit spreads or investor appetite have changed. The Dallas Fed noted that private-credit financing is more likely to carry floating rates, while data-center owners generally seek long-term fixed-rate financing to match the life of their assets. [[11]](https://www.dallasfed.org/research/economics/2026/0210-searls-aifinancing) The third is technology risk. GPU and networking equipment can produce revenue for years, but their value depends on performance, power efficiency and compatibility with newer systems. CoreWeave’s March 2026 filing disclosed a 16-year lease with contractual rent payments ranging from $18.7 billion to $19.6 billion for a site with access to 525 megawatts of power. It also estimated $900 million to $1.6 billion of additional equipment obligations for leased facilities. [12] Those commitments can be economically sensible if utilization remains high; they can become restrictive if the equipment or capacity is displaced before the lease ends. The final risk is opacity. Public companies disclose major commitments, but private funds, project vehicles and bilateral contracts do not provide one standardized measure of total exposure. The BIS said these arrangements can blur the boundary between corporate and project finance and between public and private credit. [8] That makes lender concentration, cross-guarantees and correlated exposures difficult to assess from public data alone. ## What the financing boom does—and does not—establish The new financing structures show that institutional capital is willing to fund AI infrastructure beyond the cash generated by the largest technology companies. They do not establish that every proposed campus will be built, that every financing platform will reach its stated target or that AI revenue will cover the obligations already contracted. Nvidia’s $500 billion announcement is a sign of growing financial intermediation around compute; it is not a balance-sheet transfer of that amount to Nvidia or its partners. [[1]](https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital)[[2]](https://www.axios.com/2026/08/11/nvidia-chip-securitization-wall-street) For analysts and lenders, the useful measure is therefore broader than reported debt. It includes future lease payments, purchase commitments, guarantees, residual-value exposure, tenant concentration, power contracts, construction debt and the dates when each project must be refinanced. Oracle, Meta and CoreWeave provide unusually large disclosed examples, but the same pattern extends through operators, cloud providers and private vehicles. [[3]](https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm)[[4]](https://www.sec.gov/Archives/edgar/data/1326801/000162828026028526/meta-20260331.htm)[[12]](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000222/crwv-20260331.htm) The buildout can continue without a broad credit event if demand grows into the capacity and projects refinance smoothly. The harder scenario would be a synchronized slowdown in AI spending that reaches projects before their leases, power contracts and debt maturities have adjusted. The relevant exposure is being created now, across contracts that will come due over many years. ## Companies mentioned ## Further sources [[1] Nvidia announcement on partnerships with Apollo, BlackRock, Blackstone, Brookfi… ↗](https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital) [[2] Axios coverage of Nvidia’s planned financing effort and the distinction between… ↗](https://www.axios.com/2026/08/11/nvidia-chip-securitization-wall-street) [[3] Oracle fiscal 2026 Form 10-K, including $260 billion of additional lease commit… ↗](https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm) [[4] Meta first-quarter 2026 Form 10-Q, including $182.88 billion of future lease ob… ↗](https://www.sec.gov/Archives/edgar/data/1326801/000162828026028526/meta-20260331.htm) [[5] Morgan Stanley’s June 18, 2026 discussion of AI debt issuance, which said issua… ↗](https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/ai-debt-surge-carolyn-campbell-vishwas-patkar) [[6] Morgan Stanley research report estimating roughly $2.9 trillion of global data-… ↗](https://www.morganstanley.com/content/dam/msdotcom/en/assets/pdfs/Research_Bridging-Data-Center-Gap.pdf)+6 more The stories that matter, in one email. Free — unsubscribe anytime.