Financing the AI Buildout: What It Means for Corporate Bond Markets

Credit Quality: A Strong Starting Point but Needs Watching   

The higher quality borrowers have the resources to manage the AI investment cycle. Their legacy businesses can accommodate rising leverage. However, these assumptions warrant monitoring, particularly where achieving financial targets depends on realizing returns from new AI investments.  

Investment returns on the AI buildout are only starting to emerge, though anecdotal reports of revenue and productivity gains are frequent. Most of the buildout has been contracted, though revenue recognition does not occur until the assets are delivering compute capacity, often to a model builder like Anthropic or OpenAI. Hence, we get the eye-popping figures like Oracle’s contracted-but-not-yet-recognized revenue of $664 billion as of the first quarter of 2027. This future revenue (and conversion to hyperscaler earnings) relies on model builders successfully selling AI inferencing software and services to companies and individuals. The amount of compute capacity needed remains uncertain. Data illustrating the economics of the buildout — such as return on investment, capacity utilization, and price elasticity — is often either not disclosed or embedded within results of larger business units.  

That being said, the hyperscalers can lean on legacy business lines — most of which generate substantial cash flow — to fund the buildout. Google has advertising and cloud; Amazon has retail, cloud, and advertising; Meta has advertising; Microsoft has cloud and software; Oracle has cloud and software; and SpaceX has satellites and rocketry. Results for the six companies continue to be strong with high rates of growth providing cushion for additional debt (Exhibit 3).  

Most hyperscalers also start from strong balance sheet position, with very low leverage. However, cohort capital expenditure is estimated at roughly $863 billion in 2026 and projected to rise approximately 50% to $1.3 trillion in 2027. This spending will comfortably exceed forecasted cash flow of $975 billion², a highly unusual dynamic for these companies that will likely lead to continued debt financing. Furthermore, we expect companies to continue employing alternatives to public debt financing, which can help manage capital market activity as well as add complexity to credit analysis. These alternatives include use of off-balance sheet lease financing and contingent obligations. For example, as of the second quarter of 2026, combined off-balance-sheet lease commitments from the hyperscalers are roughly $1.1 trillion, more than twice as much as reported long-term debt. Lease terms, including cancellation provisions, are often not disclosed. The ultimate timing and value of contingent obligations are uncertain.

A notable example is Meta’s financing of data centers in Louisiana and Texas. While Meta is only a minority owner of these projects, it leases the facilities and provides residual value guarantees (RVGs). RVGs only become recognized liabilities if Meta decides not to renew its lease. The RVG is dependent on the data center’s economic value at that time. Meta reports total potential data center financing exposure of $68 billion, of which only $4.7 billion was on-balance-sheet exposure in the form of minority ownership interests, while the remaining $63.3 billion represents lease and potential RVG obligations.  

Four developments we are monitoring: 

1. Midterm election outcomes. The pace of the AI buildout has been largely unrestricted, limited only by physical and financial capital as well as certain local regulatory constraints. With Democrats gaining control of Congress in the midterm elections, regulatory scrutiny of data center construction could increase, potentially slowing the pace of development. We expect projects already underway to continue but moratoriums in some states to lead to delays or cancellations. Importantly, a moratorium is temporary. However, the timing and conditions for lifting these restrictions remain uncertain as policymakers address environmental and power-grid impacts. In addition, some projects may shift to states with more hospitable regulatory environments. This has the potential to move the pace of development toward the traditional sweet spot for IG credit of “not too hot, not too cold” as near-term demand for credit moderates. However, project delays could also defer revenues and increase project costs, reinforcing the importance of existing cash flows to cushion the impact of delays. And a more measured pace of construction could also reduce the incentives for more aggressive use of balance sheets. Finally, it may also allow the AI adoption patterns to deepen and broaden, better demonstrating the economic benefits.  

2. Hyperscaler assumptions on the useful lives of their AI assets will have a direct bearing on reported profitability and cash flow. The expected useful lives and terminal values of AI infrastructure remain open questions. Servers and networking are depreciated over roughly five or six years by the hyperscalers. Excluding SpaceX, which only became public this year, the other five hyperscalers mostly lengthened these useful lives between 2022 and 2025. Assumptions of longer economic lives lowers annual depreciation and raises reported earnings. For example, in 2025, Meta lengthened most server and network asset lives to 5.5 years, exceeding the previous range of 4 to 5 years and adding $2.6 billion to net income. An exception was Amazon, which reduced expected useful life on a subset of equipment from 6 years to 5 years, citing the pace of technological development. This reduced net income by $1.0 billion.
   
   More conservative (shorter) useful life assumptions help account for faster-than-expected technology obsolescence. In this scenario, assets would need to be replaced sooner than their depreciation schedules imply. Replacement capex would run ahead of depreciation, resulting in surging cash outflows and overstating sustainable earnings. Ultimately, a company could be forced to recognize an earnings hit in the form of accelerated depreciation resulting from shortened useful lives, losses on early retirement, or impairment charges. 

3. Observable measures of AI demand show continued increases in AI adoption, although it remains early in the development of the market. Adoption measures include the U.S. Census Bureau’s Business Trends and Outlook Survey (BTOS) and the St. Louis Fed Real-Time Population Survey. The October 22 BTOS shows almost 25% of respondents using AI, up from 17% in November 2025 (Exhibit 4). And the third-quarter 2026 St. Louis Fed Real-Time Population Survey reports an overall adoption rate of 61%, a modest decline from the second quarter but an increase from 45% two years ago (Exhibit 5). Both surveys are relatively new and use different methodologies. There is no guarantee they fully capture the economic uptake of AI. But they have become useful market signals. Finally, while there is debate over whether the AI buildout is a bubble, it may simply be a matter of timing. The fiber infrastructure buildout of the 1990s was considered an overbuild for many years but was eventually utilized as data consumption grew into the supply.  

4. Compute suppliers are increasingly providing financing and/or backstops to customers. While AI model builders (e.g., OpenAI, Anthropic) and neocloud providers (e.g., CoreWeave) are driving much of the demand for compute, they currently lack IG credit profiles. The cash-rich hyperscalers have an economic interest in supporting their growth and have been providing financing and/or investments. While sustaining demand, this also weakens the signals that would otherwise discipline capacity buildout. For example, Amazon has announced equity investments worth up to $50 billion in OpenAI and up to $25 billion in Anthropic. These investments include a phased component based on certain events occurring. Meanwhile, OpenAI and Anthropic have pledged to purchase compute from Amazon in amounts worth at least $138 billion and $100 billion, respectively, over multiple years. This activity is not limited to the hyperscalers, with chip financing arrangements expected to ultimately exceed data center financing. A very recent example is Broadcom and a consortium of capital providers announcing a $63 billion financing facility for accelerator chips and equipment. Anthropic will lease the hardware and Broadcom provides partial, contingent support up to a maximum potential liability of $43 billion. These types of supply financing agreements are becoming increasingly common. They balance stronger sales and earnings growth against leverage and earnings risk via equity investments.  

Portfolio Approach  

Credit selection remains paramount.

• As the IG bond market has evolved significantly over the past year, we are mindful of the growth in long-duration risk tied to the AI story. We prefer companies whose legacy businesses can support additional debt obligations, rather than those that depend on AI revenue to achieve financial targets. 

• We continue to monitor the hyperscalers’ quarterly results for evidence of durable AI monetization. In particular, consistent disclosure of AI-specific return on investment — not simply revenue growth — would improve our comfort levels. We are also looking for financial reporting from model builders Anthropic and OpenAI demonstrating that their services are funded primarily by customer revenues rather than support from compute suppliers. Such evidence could point to broader adoption and help alleviate concerns about circular financing. 

• Despite more attractive pricing on construction and project debt secured by data center assets, or on subordinated obligations such as leases, we believe the risks outweigh the potential returns. Deal terms and conditions are more complex and less tested compared to senior unsecured obligations. The time-worn industry phrase “cheap for a reason” comes to mind given the risk of a tenant walking away from a lease in a few years, or shorter than expected economic life for chips and servers in a given project. 

We remain selective on participating in new issues, as we expect credit spreads to remain vulnerable to negative AI sentiment. 

• The surge in hyperscaler issuance this year led to progressively higher concessions on new issues. For example, in February 2026, Alphabet was able to price a $20 billion issuance with spread concessions ranging between 7 and 13 basis points. This is modestly wider than the 0–5 basis-point range seen on typical IG financings. Yet in August 2026, at the peak of supply surge, Alphabet’s $25 billion deal required a much more generous concession of 25–27 basis points to clear the market.   

• Sizable issuance is expected through 2028, and we expect elevated new-issue concessions to persist or even increase. As the AI investment buildout slows, less issuance will be required, and markets could unwind the spread penalty on new issues. If this occurs, we could become more involved in hyperscaler debt, as these companies may return to their longer-term historical profile as infrequent borrowers with a bias toward structurally tighter credit spreads.