Blue Owl Capital and PIMCO have structured a $2.4 billion debt facility for IREN Ltd, the Nasdaq-listed operator that is converting bitcoin-mining sites into AI compute campuses. Reporting on the deal indicates the proceeds are earmarked for purchasing Nvidia accelerators — the specialised processors that run AI training and inference workloads. Separately, Core Scientific announced $600 million in new credit facilities.

The two financings land alongside IREN’s statement that its 2026 capacity is sold out and that it is now negotiating contracts for 2027 and 2028. Together they mark the maturing of a financing structure in which the chips themselves, and the contracted revenue they generate, carry the debt.

Executive Summary

The headline number is $2.4 billion, but the more consequential detail is the structure. Blue Owl and PIMCO are both large private-credit managers — firms that lend directly to companies rather than arranging syndicated bank loans — and they have built a facility specifically tailored to GPU procurement. That framing implies a financing secured against a hardware fleet and the contracts that fleet serves, rather than against a diversified corporate balance sheet.

This matters because it decouples AI infrastructure buildout from equity issuance. A neocloud — an operator that rents out GPU capacity without the broader service portfolio of a hyperscaler like AWS or Azure — has historically had two ways to buy chips: sell shares, or fund from cash flow. Neither scales to multi-billion-dollar fleets. Asset-backed debt is the third path, and it is now open at institutional size.

The trade-off is symmetrical. Pre-selling capacity years forward gives lenders visible cash flows to underwrite against; IREN’s claim that 2026 is fully contracted is precisely the kind of evidence that makes such a facility underwritable. But it also fixes revenue in advance while leaving the borrower exposed to the residual value of assets that depreciate on a schedule nobody has yet observed across a full technology cycle.

What It Means to Pledge a Chip

Collateralised lending is old; the question is always what the lender can recover if the borrower stops paying. Real estate works as collateral because buildings are immobile, long-lived, and trade in a deep secondary market. Aircraft and shipping containers work because they are standardised, tracked, and re-leasable. GPUs are a genuinely new asset class in this respect: they are standardised and in acute demand, which argues for strong recovery values, but they are also installed inside purpose-built facilities with specific power and cooling requirements, which complicates repossession in any literal sense.

In practice, facilities of this type tend to rely less on physically seizing hardware and more on capturing the contracted revenue that hardware produces — the customer agreements, and the entity that holds them. That is why the sequencing in IREN’s case is notable: the company’s statement that 2026 capacity is sold out precedes and supports the financing logic. Lenders are underwriting a contracted book, with the chips as backstop rather than as primary recovery.

None of the public material specifies the security package, the advance rate against hardware cost, the tenor, or the pricing. Those terms are where the actual risk allocation lives, and their absence is the single largest gap in what has been disclosed.

The Residual Value Problem Nobody Has Solved

Every asset-backed structure embeds an assumption about what the asset is worth at the end. For GPUs, that assumption is unusually hard to defend. Nvidia has been shipping new accelerator generations at a cadence far faster than the multi-year amortisation periods typically applied to data centre equipment, and each generation has delivered large performance-per-watt improvements. A chip that is two generations old is not worthless — inference workloads, smaller models, and price-sensitive customers all provide a floor — but its rental rate is not the rate it commanded at launch.

This creates a specific mismatch. If a facility amortises over, say, a longer horizon than the period during which a chip commands premium pricing, the borrower must either re-contract older hardware at lower rates or refinance into a fleet upgrade. Both are manageable in a market with excess demand. Neither is comfortable if demand normalises while the debt schedule does not. The honest position is that no one has yet observed a full GPU depreciation cycle under sustained competitive supply, so residual-value assumptions in these deals are estimates, not history.

It is worth being even-handed here. The counterargument — that compute demand has repeatedly outrun supply forecasts, and that older accelerators have found ready secondary uses — is not unreasonable. The point is not that these facilities are unsound; it is that their soundness rests on a forward-looking judgment that has not been stress-tested, and that lenders are being compensated for taking it.

Winners, Losers, and the Private-Credit Angle

The clearest beneficiaries are the neoclouds themselves. IREN and Core Scientific both originated as bitcoin miners, meaning they already controlled the scarcest input in AI infrastructure — energised sites with interconnection agreements and power contracts. What they lacked was the capital to fill those sites with accelerators. Debt of this kind converts a land-and-power position into a compute business without diluting shareholders at every step.

Nvidia benefits indirectly and substantially: financing capacity is now a gating factor on GPU sales, and structures that unlock institutional debt expand the buyer pool beyond hyperscalers with investment-grade balance sheets. Private credit managers benefit from a new, large, yield-generating asset class at a moment when they hold substantial dry powder. Traditional banks are, for now, less visible in these transactions — which is itself informative about where regulatory capital treatment and risk appetite currently sit.

For buyers of AI capacity, the second-order effect is availability. More financed hardware means more contractable capacity, and IREN’s stated pivot to 2027 and 2028 negotiations suggests operators are trying to lock in demand well ahead of delivery. Enterprises signing multi-year GPU contracts should nonetheless treat counterparty durability as a real diligence item: a highly levered provider whose debt is secured against the very fleet serving your workload is a different credit risk than a hyperscaler, and contract terms should reflect that.

Background

Both IREN and Core Scientific began as bitcoin miners, businesses defined by the pursuit of cheap electricity at scale. That pursuit left them holding something the AI buildout badly needs: sites with signed grid interconnection agreements and multi-year power contracts, in a market where new interconnection queues can run for years. When AI compute demand accelerated, converting those sites to GPU hosting became a more attractive use of the same infrastructure. Core Scientific emerged from Chapter 11 bankruptcy protection in 2024 and continued that pivot; a proposed all-stock acquisition by CoreWeave was rejected by its shareholders in 2025, leaving the company independent.

The financing question followed directly. Site and power are capital-intensive but financeable through familiar channels; filling those sites with accelerators requires very large equipment purchases that neither company could fund from operating cash flow. Equity issuance dilutes shareholders. That gap is what facilities like the Blue Owl and PIMCO structure are designed to fill, and it explains why the terms of these deals — not just their headline sizes — are the thing worth watching.

Source: Blue Owl (OWL.US) partners with PIMCO to structure a $2.4 billion GPU financing facility tailored for IREN (IREN.US) — coverage of the Blue Owl and PIMCO debt facility for IREN, reported alongside Core Scientific’s $600 million credit facilities and IREN’s statement that its 2026 capacity is fully contracted.

Market research firm Dell’Oro Group has published analysis describing a growing “memory tax” on AI infrastructure — the rising share of system cost attributable to high-bandwidth memory (HBM) and DRAM in AI servers and accelerators. The note, surfaced April 27, 2026, frames memory as an increasingly material and often under-examined component of AI capital spending.

Executive Summary

Dell’Oro Group, an analyst firm that tracks data center and telecom infrastructure markets, is calling attention to memory — specifically HBM, the stacked memory packaged alongside AI accelerators, and conventional DRAM used in servers — as a fast-growing cost component in AI infrastructure. The “memory tax” framing suggests that as AI models and the clusters that train and serve them grow, memory is consuming a larger slice of every infrastructure dollar.

The framing matters because most public discussion of AI capital expenditure centers on GPUs and, increasingly, on power and data center construction. If memory costs are rising as a share of the bill of materials — the itemized cost of the components inside a server — then budget models built around accelerator pricing alone will understate the true cost of AI capacity. That has implications for cloud providers, enterprises buying AI servers, and the memory suppliers positioned to benefit.

Readers should note what is available here: a headline and thesis from a recognized analyst firm, without the underlying figures, forecast horizon, or methodology visible in the source material. The direction of the claim is consistent with the widely reported tightness in memory supply driven by AI demand, but the magnitude is not substantiated in what we can see.

Why Memory Became a Line Item Worth Naming

AI accelerators are unusual among chips in that their usefulness is bounded as much by memory as by raw compute. Training and serving large models requires moving enormous volumes of data to the processor quickly, which is why modern accelerators are packaged with HBM — DRAM dies stacked vertically and connected to the processor over a very wide, short interface. HBM is expensive to manufacture, supply is concentrated among a small number of suppliers (SK hynix, Samsung, and Micron are the established producers), and each new accelerator generation ships with more of it.

Conventional DRAM matters too: the host servers around the accelerators, plus the storage and networking tiers of an AI cluster, all consume memory. When one demand source — AI — pulls hard on a supply chain with long lead times and few producers, prices tend to rise across the board. Dell’Oro’s “memory tax” label captures the effect from the buyer’s side: a cost that arrives embedded in system prices whether or not the buyer itemizes it.

Who Pays, and Who Collects

If memory’s share of AI system cost is growing, the immediate beneficiaries are the memory manufacturers, for whom HBM commands substantially better margins than commodity DRAM historically has. Accelerator vendors sit in the middle: memory is a cost input to their products, but strong demand has so far allowed system prices to carry it. The buyers — hyperscale cloud providers, AI labs, and enterprises — absorb the tax directly in capital expenditure, and indirectly it flows into the price of cloud GPU capacity and AI services.

There is a second-order effect worth watching. Rising memory prices do not stay confined to AI hardware. General-purpose servers, storage systems, and consumer devices draw on the same DRAM supply base, so a sustained AI-driven squeeze can raise costs for infrastructure buyers who are not purchasing AI systems at all. For data center operators and IT planners, that argues for treating memory pricing as a market variable in refresh budgets, not a constant.

An Analyst Thesis, Not a Dataset — Yet

It is worth being precise about the evidentiary weight of what has surfaced. Dell’Oro is an established infrastructure research firm, and the thesis aligns with observable market conditions. But the material visible here is a headline-level framing: it does not disclose how large the memory share of AI system cost currently is, how fast it is growing, or over what forecast period. “Growing” is directionally plausible and quantitatively unverified in this source.

That distinction matters for anyone using the claim to make decisions. A memory share that rises from, say, a modest slice to a dominant one would reshape supplier negotiations and cloud pricing; a gradual drift would be a planning footnote. Until the underlying figures are public, the responsible reading is that memory costs deserve a named line in AI infrastructure budgets — and that the size of that line needs data the summary does not provide.

Background

The AI infrastructure buildout that accelerated from 2023 onward has been discussed mostly in terms of GPUs, power, and data center construction, but every AI accelerator ships with a large complement of high-bandwidth memory, and every cluster consumes conventional DRAM in its servers and supporting systems. Memory is a historically cyclical market dominated by a small number of manufacturers — SK hynix, Samsung, and Micron — and AI demand has become a defining force in its current cycle.

Dell’Oro Group, founded in the 1990s and based in Silicon Valley, publishes recurring research on data center capex, servers, and network infrastructure. Its analysts’ framing of trends — in this case, memory as a “tax” on AI infrastructure — often shapes how vendors and buyers talk about market economics before detailed figures circulate publicly.

Source: The Growing Memory Tax on AI Infrastructure — Dell’Oro Group, analyst commentary on rising HBM and DRAM costs in AI infrastructure economics, published April 27, 2026.