Summary. We believe Data Center fundamentals remain extremely strong, yet spreads on Data-Center-backed credit, particularly ABS and CMBS, have widened and demand for issuance has struggled to keep pace with rapid growth. We view ABS and CMBS backed by Data Centers as cheap relative to underlying fundamentals driven predominantly by growing issuance. Inference demand is arriving at scale, driving a step-function increase in Hyperscaler and AI-lab leasing that should translate into elevated capex well into 2030. We see strong relative value in Data Center ABS and CMBS, particularly Colocation and Enterprise shelves, which we expect to tighten as the basis to recent Investment Grade (“IG”) secured corporate deals compresses. In our experience, we also have found durable alpha via structural analysis and security selection in the sector.
An update to our September 2024 primer. Eighteen months ago, we walked clients through the Data Center ABS market from the ground up: what a Data Center is, how Hyperscale, Colocation, and Edge differ, how deals are structured, and where we saw the best relative value across ABS, CMBS, and corporate credit. AI was a clear tailwind, but demand rested primarily on Enterprise Cloud migration, hybrid compute, and traditional digitization. Since then, the market has grown by 2.5x: outstanding US Data Center securitized credit now sits near $60B, with $9.5B issued in Q1 2026, the heaviest quarter on record**. Far larger deals are being executed in the corporate finance market as well*.
The clearest signpost for what has changed is not in any research report; it is in your morning coffee order. A Starbucks customer can now open ChatGPT, describe a mood or craving, and have the model recommend and customize a drink before routing to pickup. Walmart, Target, Home Depot, and Uber have live integrations as well. A small convenience for consumers; a meaningful signal for everyone else.
Agentic AI has quietly moved from lab demos into everyday transactions, and each interaction runs on inference compute – the processing power, memory, and energy required to run a trained AI model – that did not exist at meaningful scale when we first published. We believe the arrival of inference at consumer scale turns a strong sector into an accelerating one, reshaping the Digital Infrastructure credit backdrop in real time. This commentary updates the thesis, with inference now as the accelerant in our view.
1. Fundamentals Are Strong: Spreads Are Wide for the Risk
Structural undersupply persists. Industry occupancy sits near 97%, and roughly 77% of Data Center capacity under construction is already pre-leased, per JLL and CBRE**. That combination provides multi-year utilization visibility before a single new kilowatt comes online. JLL's downside scenarios still project sub-5% vacancy through 2030, underpinned by long development cycles and high pre-commitment levels**. Generally, even if demand disappoints, the supply side cannot respond quickly enough to meaningfully dent occupancy.
Demand is diversified and durable. We believe AI is accelerating growth in Data Center needs, but it is only one of multiple contributing factors. Cloud revenue has compounded at roughly 20% annually over the past decade, global internet traffic at around 21%, and data creation and storage at around 23%**. Most facilities inside ABS and CMBS structures have limited or no direct AI exposure: these are multi-tenant Colocation and Enterprise-oriented assets serving a long tail of durable compute demand. That breadth is often overlooked when investors paint the sector with an "AI bubble" brush.
Development constraints make existing digital infrastructure more valuable. Greenfield project timelines (i.e., those that are developed from scratch without the constraints of existing buildings or infrastructure) average roughly two years, and power availability has emerged as the binding constraint. Grid interconnection queues and regulatory hurdles have lengthened, and fewer than 10% of facilities are self-powered**. Per TD Cowen's April 14 channel checks, lead times on reciprocating engines (widely used in industrial power generation) have gapped to 3–4 years from roughly a year in early 2025; J.P. Morgan (“JPM”) notes large combined-cycle gas turbines (highly efficient power generation systems that combine gas and steam to create) now carry 3–5 year lead times; and electricians have emerged as a bottleneck given the 4–5 year apprenticeship cycle*. We believe these constraints make existing Digital Infrastructure more valuable even as construction continues at pace.
QTS case study: tenants are chasing data center operators, not the reverse. When Blackstone acquired QTS Realty Trust, a Data Center REIT in 2021, the company operated roughly 360 Megawatts (“MW”) at an enterprise value near $10 billion. Today QTS operates approximately 5 Gigawatts (“GW”), with a booked-but-not-built pipeline of 7.3 GW over the next three years, on typical 15-year leases with two 10-year renewal options. Enterprise demand at scale is now arriving alongside unrelenting Hyperscaler demand: one Hyperscaler CEO recently escalated to the top of QTS to ensure delivery, and the ultimate ask came in at double the initial request. This is not a market where operators are chasing tenants; it is the opposite.
Data Center ABS/CMBS are infrastructure assets with low technological risk. Generally, the bondholder owns infrastructure: power systems, cooling, network connectivity, and backup power. Computing equipment sits with the tenant, meaning technological obsolescence risk is downstream of the landlord's collateral. In these cases, cash flows derive from long-term leases with high expense pass-through, low churn, and strong debt service coverage. Morgan Stanley notes weighted-average Debt Service Coverage Ratios (“DSCRs”) in the ABS universe remain near 1.9x, with LTVs actually lower than historical levels even as trusts have grown.
And yet spreads have widened. Despite all of the above, Morgan Stanley observes that Data Center ABS spreads widened materially through Q1 2026, outpacing both the IG corporate index and other esoteric ABS segments. Per Morgan Stanley’s pricing analysis, A-rated 5-year A2 tranches priced in the +170–245 range and BBB- tranches in the +275–330 range, against heavy supply ($7.5B of US ABS and $2B of US CMBS in Q1 2026) rather than any fundamental deterioration. Spreads are wide for the risk before we even get to the inference demand story***.
2. Inference Has Arrived, Driving Data Center and Fiber Demand
The leasing data is unambiguous. TD Cowen's April 14 channel checks show a record ~9.4 GW of US Hyperscaler and AI-lab Data Center leasing in Q1 2026, eclipsing the prior record of ~7.4 GW in Q3 2025 and roughly 5x the Q1 2025 pace. Amazon led with over 2.5 GW, followed by Google, Meta, OpenAI (~1.2 GW in Texas), Microsoft (500 MW with Equinix xScale in Chicago), Oracle, Fluidstack, Anthropic, Nvidia, and CoreWeave. The active pipeline sits near 11 GW, many deals exceeding 500 MW*. Given the 18–24 month build cycle, TD Cowen expects 2027 Hyperscale capex to materially exceed 2026 levels, with deals signed in the coming months representing 2028 commitments.
Training is giving way to inference. Per JPM, 2026 is likely the year inference compute consumption surpasses training. Anthropic's disclosed revenue run-rate reached approximately $30B this month, up from roughly $9B at year-end 2025, a trajectory that would have seemed implausible eighteen months ago*. Inference workloads are latency- and geography-sensitive in a way training is not: they locate near end users, fiber trunklines, and population centers. That dynamic lifts both Data Centers and the fiber connecting them, making network access a first-order site-selection variable alongside power and water. JPM estimates roughly 122 GW of global Data Center projects require financing over the next five years, and inference is why the curve is steepening.
ABS/CMBS collateral is not the ai bet, which is a feature, not a bug. ABS and CMBS collateral consist largely of enterprise, colocation, and multi-tenant wholesale assets with limited direct AI training exposure. ABS investors thus capture the broad digital infrastructure tailwind (cloud, enterprise migration, inference-adjacent workloads, fiber growth) without concentrated exposure to any single frontier-model thesis. Generally, if AI capex disappoints, the collateral is largely unaffected; if AI accelerates, the rising tide lifts the full complex. That asymmetry should command a premium, not a discount.
3. ABS and CMBS Offer Strong Relative Value and Should Tighten
The external case for Spread compression. The compression case starts with QTS. In April, QTS priced a $4.6B 10-year IG secured corporate bond at +137.5, backed by a 300+ MW campus fully leased to Microsoft under a 20-year triple-net lease. A month earlier, the QTSII ABS shelf priced A- tranches at +195 (5-year) and +225 (7-year), backed by AA-rated tenants under 15-year triple-net leases. Same sponsor, comparable tenant quality, and higher DSCR on the ABS (~1.6x vs. 1.3x), with stronger structural protections (amortizing post-Anticipated Repayment Dated (“ARD”), cash-sweep triggers), yet the ABS prices nearly 90 bps wider.
Morgan Stanley attributes the gap to demand depth, not credit risk*. Their thesis: the corporate bid has been deeper than the securitized bid this cycle, and that imbalance should reverse as ABS investors are drawn in by the spread pickup or as more assets migrate to the secured corporate channel. They expect the tightest ABS shelves to settle near +180–200, project $22 billion of 2026 ABS issuance and $9 billion of CMBS, and prefer Colocation and Enterprise shelves over Hyperscaler.
We broadly agree with their direction and would add five points.
First, the structural cushion goes beyond what the headline spread gap implies. ABS and CMBS structures layer amortizing post-ARD mechanics, DSCR cash-trap and amortization triggers, and LTV-based sweep tests on top of comparable or better collateral, precisely to protect bondholders from tail events the corporate form does not address. We believe that on a risk-adjusted basis, the 90 bps gap not only looks wide; it understates the true cushion.
Second, the perceived real alpha is in security selection, not the headline basis trade. Pricing diverges from fundamentals around three recurring themes: call optionality near ARD, premiums on newer issuers without track records, and concessions during heavy supply periods. These are trust-by-trust dislocations, not sector-level moves. Investors positioned purely for spread compression capture less of the opportunity than those willing to do the underlying credit work.
Third, our preference for colocation and enterprise over pure hyperscaler is not a new call. We view Hyperscale exposure as one factor among many, not a defining determinant of credit quality. Colocation shelves carry diversified tenant bases, often dozens or hundreds of non-AI compute tenants near their end users, producing tenant diversification that single-tenant Hyperscaler paper lacks. Morgan Stanley’s Colocation preference aligns with ours, and we have held this view well before the current Hyperscaler supply wave.
Fourth, complexity is accelerating across the digital infrastructure capital structure. Over the past year the market has absorbed High Performance Computing (“HPC”) project bonds in high yield and IG secured format, JV IG secured deals, amortizing corporates, SASB CMBS, and traditional ABS shelves, each with different covenant packages, amortization profiles, and tenant exposures. That complexity is a headwind for passive participants and a tailwind for active ones. Triangulating across the full spectrum of Digital Infrastructure assets, from the QTS secured corporate to the QTSII ABS to a Cipher Mining project bond (an industrial-scale Data Center operator that is transitioning from pure Bitcoin mining to providing HPC), opportunities to capture spread for substantially similar risk continue to surface.
Lastly, Fiber ABS has outperformed Data Center offerings. Fiber ABS fundamentals are equally compelling, but the outperformance at the top of the capital structure, with Fiber ABS A-rated bonds trading tighter than Data Center AAA bonds, reflects technicals more than fundamentals. We do see value in Fiber ABS credit and have found alpha in name-specific stories related to make-whole premiums and market misunderstandings, particularly in the Fiber-to-the-Home segment.
Bottom line. In our view, Data Center fundamentals (occupancy, pre-leasing, supply constraints, lease durations, tenant quality) are the strongest in memory, and inference demand is only now arriving in earnest. Against that backdrop, ABS and CMBS spreads look wide relative to underlying risk and particularly wide relative to newly issued IG secured corporate comparables. Morgan Stanley has made the broad compression call; we agree but believe the more durable source of return is structural analysis and security selection across the full capital structure rather than pure-beta to the basis trade.
*Some information in this presentation has been taken from third-party sources that are believed to be reliable but which have not been verified for accuracy or completeness. Except where otherwise indicated, the information provided is based on matters as they exist as of the date of preparation and not as of any future date. This information will not be updated or otherwise revised to reflect information that subsequently becomes available, or changes in circumstances or events occurring after the date hereof.
** Market data, channel checks, and specific spread and issuance figures referenced in this commentary are drawn from Morgan Stanley Research (“Data Center ABS: A Turning Point?”, April 14, 2026), TD Cowen (“Data Center Channel Checks,” April 14, 2026), and J.P. Morgan (“Building for the AI Boom,” April 9, 2026). Fundamentals data sourced from JLL and CBRE are from Bank of America Research (“CMBS Weekly, JLL Call Recap: Capital Markets Strong, Office Hits Inflection Point, & No AI Bubbles”, December 12, 2025).
*** Source: Bloomberg.