AI Infrastructure: From Zero to $100B and Beyond: How the Emergent Sector Is Reshaping the Non-IG Market

Relevant Item:
Excel Download: AI Infrastructure Financing Summary

Morgan Stanley last year estimated that the AI infrastructure buildout would cost approximately $3 trillion from 2025 to 2028, with approximately half of that total, or $1.5 trillion, requiring external financing. For context, $1.5 trillion is roughly equivalent to the total size of each of the U.S. high-yield market or the U.S. leveraged loan market. McKinsey estimates that the total capital need for the sector will hit $7 trillion by 2030, much of it going toward new data centers and related infrastructure.

Most of this financing has been and will continue to be raised in the investment-grade market, as the largest players in the space are investment-grade hyperscalers such as Amazon, Meta and Oracle. Recent estimates compiled by the Dallas Fed suggest that in 2026 we may see as much as $300 billion of AI-related investment-grade issuance. However, AI infrastructure issuers have been getting more active in the below-investment-grade market as well.

In fact, we estimate that high-yield and unrated AI infrastructure issuers have raised more than $107 billion in committed and funded debt as of early May 2026. Approximately $68.7 billion of this total was raised by neoclouds, while the remaining $38.7 billion was raised by AI-native data center companies. An Excel file with the full list of instruments included in this total is available on the Sector Insights analysis page and for download HERE.

On the basis of current publicly disclosed capacity targets and our capital cost estimates, discussed further in this piece, we believe that the seven largest non-investment-grade AI infrastructure issuers may need to raise as much as $435 billion of asset-level and corporate-level financing to meet their publicly disclosed capacity targets over the next four years, pushing the total potential amount well past half a trillion dollars. Importantly, not all of this amount will be funded in the debt market – these issuers can and do raise equity, and they are likely to recycle capital returned from their existing contracts. Nevertheless, we believe that the sector will remain one of the most prolific issuers of below-investment-grade debt.

The AI infrastructure value chain may appear complex, with any discussion of it fast becoming technical and jargon-laden (CoreWeave’s inaugural offering memorandum dedicated five pages to a technical glossary, for example). However, at a high level, the value chain is actually fairly straightforward.

In short, in order to deliver compute to the end user (whether for inference or training load), the following links in the chain are necessary:

• Utilities, including power supply, either from the grid via an interconnection, or directly from a power plant under a power purchase agreement, or PPA (also known as “behind-the-meter”), a fiber/data connection and, in many cases, a dedicated water connection to support liquid cooling.
• Powered shell, or the data center building envelope, which needs to be connected to a power supply, and is often equipped to accommodate the specialized compute equipment (e.g., with specialized cooling systems), and
• Compute infrastructure, or the graphics processing unit, or GPU, racks and related equipment, such as memory, switches and storage, which does all the “work.”

The last two parts of the value chain are the most capital intensive ones. Morgan Stanley has estimated that a new powered shell costs upwards of $10 per watt (W) of electrical capacity, depending on the layout and location of the facility, while compute infrastructure costs are estimated to be approximately $23 per watt comprising $16.50 per watt for AI hardware and $6.50 per watt for non-AI equipment.

Octus’ analysis shows that both estimates may be somewhat understated. Construction budgets disclosed by issuers we have analyzed show total capex that ranges from $8 per watt to over $20 per watt of “critical IT load” capacity. Some of the variability is due to project scale, scope and location, as, for example, some of the more expensive projects include purpose-built behind-the-meter power infrastructure; however, a weighted average for the sample appears closer to $12 per watt.

The all-in cost of compute hardware may actually be higher than the $23 per watt estimate cited above. For example, as of March 31, CoreWeave had approximately $31.40 of gross depreciable property, plant and equipment, or PP&E, on its balance sheet, per watt of active power, while the ratio for Nebius was closer to $22 per watt of active power. CoreWeave’s active capacity is roughly 6 times larger than that of Nebius, pushing the implied weighted industry average capital cost closer to $30 per watt on the basis of disclosed details.

For comparison, according to a recent U.S. Energy Information Administration report, capital costs for widely used utility-scale power generation capacity generally range from under $1,000 (for natural gas combined cycle) to $4,000 (for offshore wind) per kilowatt, or between $1 and $4 per watt.

As a result of their capital intensity, powered shells and compute infrastructure are the areas that, to date, have required the most external capital, which has been raised as both debt and equity. While a lot of the large hyperscalers which have their own AI labs, such as Google, Microsoft and Meta, directly own and operate data centers and related hardware, the sheer size of the total capital needed to fill the aggregate compute need has led to an explosion of creative ownership and financing structures.

Since the AI infrastructure buildout began in earnest over the past three years, we have seen three types of ownership structures develop:

• Independent “AI native” data centers, such as Core Scientific, Applied Digital, Terawulf and Cipher Digital, build and then rent out the powered shell, but not the hardware, largely to hyperscalers and neoclouds. Most of these began life as crypto miners but pivoted to AI use cases as that market took off.
• Independent “neoclouds,” such as CoreWeave, Nebius, IREN, Fluidstack and Lambda, usually own GPU “racks” and rent this capacity to end customers for use in training AI models and AI inference tasks, but typically do not own their powered shells.
• Hyperscaler-backed joint ventures, such as Meta’s Hyperion or the various Oracle-Stargate entities, are functionally just powered shell owners, but they tend to operate much larger facilities than the independents and are typically sponsored by large hyperscalers, which use them as off-balance-sheet financing vehicles. In many cases, the hyperscalers provide de facto guarantees to these entities.

Companies that fit into the first two categories tend to be rated as high-yield borrowers at the corporate level, and as such, they are where we have focused our research efforts. Both the integrated hyperscalers and the JVs they sponsor have largely issued debt in the investment-grade market so far.

We would note that the industry is continuously evolving, with the lines between these categories likely to continue to blur. CoreWeave is partially “backward integrating” by funding some of its data centers via a joint venture with Core Scientific, while IREN is largely a data center operator that is in the process of “forward integrating” into GPU hardware. The most notable recent entry into the infrastructure link of the value chain is SpaceX’s xAI, which recently leased out compute capacity to Anthropic at one of its captive data centers.

The financing structures employed by AI infrastructure owners can be roughly divided into two types, asset-level and corporate-level financings. While the line between the two is sometimes blurry with many asset-level structures carrying corporate guarantees, for example, while some corporate bonds can look to unencumbered assets for recovery, the main difference between the two is in the principal underlying risk the creditor takes. In asset-level structures, the primary credit support is physical collateral and related contracts, while corporate debt relies largely on the residual equity value, or remainder interest, of a company’s asset-based financing and the implied value of future business.

There has been more variability in the asset-level structures. The below highlights the structures we have seen just among the companies we have analyzed:

• Deferred-draw term loans, or DDTLs, are typically backed by GPU clusters and long-term customer contracts, though more recently this structure has been employed by data center owners as well. CoreWeave is the pioneer of this structure and has, so far, raised approximately $27 billion in DDTL commitments from the institutional private and broadly syndicated loan markets, with its most recent $3.1 billion facility placed into the broadly syndicated loan market last week. The largest of these facilities actually achieved an investment-grade rating, driven largely by the credit profile of the contract counterparty, Meta. DDTL-type structures have also been employed by CoreWeave’s fellow neoclouds Fluidstack and NScale, as well as by some powered shell owners, including Galaxy and Core Scientific.
• Data center SPV bonds are issued by powered shell owners and are usually backed by a specific property and its associated lease. Borrowers using these structures include subsidiaries of TeraWulf, Cipher Compute (including Black Pearl), Applied Digital and Core Scientific, among others. These are often structured with limited recourse to the ultimate corporate parent and are issued into the high-yield market, though they look almost like a single-asset, single-tranche commercial mortgage-backed securities.
  • In many cases, the underlying lease is guaranteed or otherwise backed by an investment grade counterparty. This is true even in cases where the actual tenant is a non-IG neocloud, such as Fluidstack, whose leases with multiple data centers are backed by Google.
• Secured supplier financing serves to fund equipment purchases before the equipment is ready to be delivered into a more permanent financing structure with the facilities typically secured by such equipment, which is primarily GPUs. These are typically not syndicated into the institutional market but are usually included in the debt balance of the issuer.
• Trade financing, largely in the form of customer prepayments, is typically amortized over the life of the related contract and is typically reflected as deferred revenue, rather than debt, on the issuer’s balance sheet.

Away from the non-IG market, the various hyperscaler AI JVs have been big issuers of asset-level investment-grade paper, although not all of it is nominally secured. The pioneering deal in this space was Beignet Investor, a Blue Owl/Meta joint venture, which issued $27 billion of A+ rated bonds last year to fund a very large data center to be leased to Meta. Notably, Meta provided a quasi-guarantee to the borrower entity via a guaranteed minimum value clause within its lease agreement. More recently, RD Michigan Property, a financing vehicle for the Related-Blackstone joint venture that is building another very large data center to be leased to Oracle, placed approximately $14 billion of bonds into the investment-grade market. Finally, we are also aware of some ABS issuances that are backed by data center leases.

On the corporate-level side, issuance by both the neoclouds and the data center owners has been heavily dominated by convertible bonds, rather than “straight” bonds. In fact, to our knowledge, CoreWeave is the only below-investment-grade AI infrastructure company to have issued nonconvertible non-asset-level bonds to date, and these bonds have been fairly volatile since issuance. We believe that the broader high-yield market is still getting comfortable with the long-term business risk of neoclouds and powered shell data centers, making equity-linked paper an easier path.

Our analysis of CoreWeave shows that its equity valuation and level of corporate debt coverage are highly sensitive to a number of assumptions, including contract pricing, run-rate margins and GPU residuals.

As detailed above, we estimate that just the publicly traded neoclouds and non-investment-grade independent powered shell owners we track have raised approximately $107 billion of funded and committed debt to date.

Roughly two-thirds of this total was raised by neoclouds, with CoreWeave, in turn, comprising the lion’s share of that subtotal. The issuer mix within the data center space is more balanced, with the “big four” publicly traded issuers (TeraWulf, Core Scientific, Applied Digital and Cipher) each accounting for $5 billion to $6 billion.

In credit support terms, approximately $73 billion of the $107 billion raised was asset-based in nature, with GPU-backed delayed-draw term loans and secured data center bonds each approaching $30 billion in total. As noted earlier, convertible bonds dominate corporate issuance in the sector, with CoreWeave accounting for all of the “straight” corporate-level high-yield issuance.

Much of the $107 billion in debt was raised in the last 12 months. Looking at only the three bond categories, AI infrastructure issuers were responsible for over $56 billion of issuance since last May, with $30 billion of that total coming in the first five months of this year, as shown below:

Each of the issuers we cover are targeting significant capacity growth to meet what they see as the relentless and continuing demand for compute capacity. The four-largest publicly traded powered shell companies are aiming to increase their AI-focused capacity to over 10 gigawatts over the next few years from approximately 445 MW today. On the basis of our analysis, we believe only a quarter of this increase has been funded so far. Using capital intensity assumption of $12.50 per watt, this translates into a funding need of approximately $91 billion, before any cash flow from already-funded deals is taken into account We believe that much of this is likely to be funded in the bond market, both at the asset and the corporate level.

Neocloud growth targets are similarly lofty. The three publicly traded neoclouds (CoreWeave, Nebius and IREN) have communicated long-term goals that aggregate to at least 16.5 GW of capacity, up from just above 1 GW today. A significant portion of this capacity is already contracted with customers, though not yet activated. We estimate that approximately 14 GW of the projected increase is presently unfunded. At $25 per watt, this would require approximately $344 billion of financing, before accounting for any capital returns from the current portfolio.

As the sector is relatively novel and “untested,” most of the debt issuance has come to market fairly wide to its rating. For example, the 11 data center SPV notes we track priced at an average option-adjusted spread, or OAS, that was roughly double the spread on the ICE BofA BB High Yield Index at the time of issuance, as shown below. On an issue-by-issue basis, the spread at issue tightened somewhat earlier in the year before the outbreak of the Iran war but has subsequently increased again.

In the secondary market, these bonds tightened by 84 bps relative to the high-yield index, on average. However, as a group, they continue to trade approximately 105 bps wide to their rating.

CoreWeave, the only neocloud company to issue high-yield bonds to date, has also seen its debt trade wide for its rating. Its four bond issues trade at an average OAS spread that is 186-bps higher than the ICE BofA B High Yield Index, as shown below:

The minor amount of average spread tightening since issuance masks the significant volatility that these bonds’ prices have exhibited. As shown below, CoreWeave’s inaugural high-yield bond, the 9.25% due 2030, has traded as tight as 430 and as wide as 845 over comparable Treasurys. In price terms, the same bond traded as low as 90 and as high as 104.5. As we have previously noted, CoreWeave’s bonds have also been among the most widely traded in the high-yield market.