AI Infrastructure’s Scarce Asset Is Not Capital

IREN’s June 2026 $3.65 billion GPU financing, linked to contracted cash flows and a Microsoft contract, shows why some AI infrastructure operators can borrow at scale while others cannot. DigitalOcean, Lambda and Nscale reinforce the same lesson from different angles: lenders are not simply backing equipment purchases. They are testing whether equipment can be deployed, used, paid for and recovered if the plan goes wrong. The disclosed figures also demand care: $3.65 billion represents roughly 63% of IREN’s stated $5.81 billion GPU capital expenditure, while the reported 3.31% average cost cannot be calculated from the stated SOFR-plus margins without the relevant SOFR rate.

IREN did not ask lenders to make a philosophical bet on artificial intelligence.

On 1 June 2026, IREN Limited closed $3.65 billion in GPU financing. The package included a $2.1 billion private placement at SOFR plus 2.13% and a $1.55 billion delayed-draw term loan at SOFR plus 2.25%. It was secured by GPUs and contracted cash flows connected to a $5.81 billion GPU capital expenditure programme for Microsoft, the US technology company that buys cloud computing capacity at global scale.

That sounds, at first glance, like a story about cheap money chasing expensive machines. It is not.

IREN said the financing had an average cost of about 3.31%. Yet the disclosed components are floating-rate loans priced above SOFR, the benchmark interest rate used in many US dollar loans. The brief does not state the SOFR rate, or the dates used to calculate that reported average. So the 3.31% figure cannot be independently reconstructed from the two margins alone. It should be treated as IREN’s reported aggregate cost, not as arithmetic proved by the headline terms.

There is a second catch. $3.65 billion is roughly 63% of $5.81 billion, not 96%. The financing is enormous. It just does not fund nearly all of the stated capital expenditure.

Those corrections matter because they point to the real lesson. A lender is not being asked whether GPUs are exciting. It is being asked whether a chain of things will happen in the right order: machines will arrive, a site will be ready, power will work, a customer will pay, and something valuable will remain if the original plan breaks.

Imagine you are the lender. A warehouse full of GPUs may be valuable. GPUs serving a contracted customer, with cash flows attached, are another category of asset altogether.

The machine is only one link

Most people see a GPU as the obvious collateral in an AI loan. It is physical, costly and in high demand. That makes this look like ordinary equipment finance.

DigitalOcean complicates that neat idea. The Broomfield, Colorado-based AI-Native Cloud provider announced a $725 million equipment finance facility on 10 September 2026, maturing on 10 September 2030, with an option to increase borrowing by up to $300 million. The collateral included GPUs, CPUs and other equipment.

MUFG, the Japanese banking group; BMO, the Canadian bank; and Wells Fargo, the US bank, were among the lenders named in DigitalOcean’s announcement.

DigitalOcean’s September 2026 facility illustrates how a mid-market cloud provider can use equipment collateral alongside projected revenue and balance-sheet strength. Photo: Dimon4ezzz / Wikimedia Commons, CC BY-SA 4.0.

A GPU sitting in a crate does not repay a loan.

It has to reach a data centre. The data centre has to be ready. Power must be available. The computing capacity has to be sold. And the customer has to pay. If one of those links fails, the lender needs to know what it controls and what it can recover.

This is where most people stop looking. They see collateral and assume risk has been solved. Collateral merely changes the question: what is that equipment worth if a site is late, power is unavailable, demand disappoints or newer hardware weakens the resale market?

DigitalOcean said its facility was supported by projected revenue streams and balance-sheet health. That is a disclosed fact. The stronger claim, that lenders have solved every deployment or recovery risk, would be an inference too far. Its public announcement does not disclose equipment types, customer contracts, sites, power arrangements, lender enforcement rights or the terms for accessing the extra $300 million.

An accordion, the option to enlarge a facility, is not a blank cheque. It is an option whose use depends on conditions that are not publicly detailed here.

GI Network’s view: The useful question is not, “Can this company borrow against GPUs?” It is, “Can a lender follow the GPU from purchase order to powered site to paying customer, and still identify a recovery path if the customer disappears?”

Lambda made the customer part of the asset

San Francisco-based Lambda, Inc. describes itself as a “Superintelligence Cloud”. On 27 August 2026, it closed a $926 million senior secured Term Loan B, rated Baa2 and priced at SOFR plus 3.00%.

The GPU servers securing that facility supported a known investment-grade customer.

That detail is doing a lot of work. A GPU server may have theoretical resale value. A GPU server assigned to a known creditworthy customer has a clearer route to revenue while the loan is outstanding.

Lambda used an SPV, a separate legal vehicle that can hold defined assets and cash flows. The structure does not eliminate risk. It makes the question more precise: which assets sit inside the vehicle, which payments support them, and what happens if those payments stop?

The brief also records Lambda’s $1 billion fixed-rate delayed-draw facility of 1 October 2026, rated A (low)/Baa1 and priced at 6.78%, across multiple investment-grade customers. It supports the wider pattern. But the public details in this brief do not reveal every recovery right, operating condition or customer commitment within either facility. A rating and a headline rate are evidence of lender comfort, not a public map of every risk transfer.

For a parallel on the danger of treating one financial metric as the whole credit case, see DSCR Is Not the Number That Sets Emerging-Market Project Debt. A repayment model can show capacity. It cannot prove that a machine will be installed and earning on time.

Nscale’s money does not arrive all at once

Nscale took a different route. Around 2 September 2026, it closed two senior secured delayed-draw loans totalling up to $3.05 billion for AI data-centre campuses in Texas and North Carolina.

IDC Nova, an industry news publisher covering data-centre transactions, reported that the facilities had investment-grade ratings while noting that detailed terms were not fully disclosed.

A delayed draw means the lender commits money but releases it later when conditions are met. Here, those conditions were tied to construction and equipment milestones.

That is a simple solution to an awkward reality. A finished building without equipment earns nothing. Delivered equipment without power earns nothing. Funding in stages means the lender does not have to accept all construction, delivery and deployment risk on day one.

The interesting part is that Nscale’s structure does not pretend the campuses are already producing cash. It turns progress into a condition for receiving more money.

Put IREN, Lambda, DigitalOcean and Nscale side by side and a pattern appears that none of the announcements says outright. Each deal is trying to remove a different blind spot.

IREN points to contracted cash flows, customer prepayments and Microsoft. Lambda points to an identified investment-grade customer. DigitalOcean combines equipment collateral with projected revenue and balance-sheet health. Nscale controls when funds are released as development progresses.

The shared object is the GPU. The real subject is execution.

Test it on a bad day

Here is the part glossy deal announcements cannot settle: what happens when deployment goes wrong?

The research brief does not provide a named failed GPU deployment. It would be dishonest to invent one. But the downside can be tested through an illustrative example based on the risks the disclosed structures are designed to manage.

Imagine a borrower receives GPUs on schedule, but the data-centre connection is delayed and power is unavailable. The machines cannot serve customers. Billing does not begin. Meanwhile, a planned customer postpones its workload. If the loan depends on projected revenue alone, the borrower is left with interest payments and equipment that may be losing value before it has generated meaningful cash.

Now compare that with the cases above. A delayed draw can stop undisbursed money from entering the problem too early. A known customer can make demand less speculative. Contracted cash flows can give a lender something beyond management forecasts. A security package may improve recovery prospects.

May is the important word. The public materials cited here do not disclose the exact enforcement mechanics, residual-value assumptions or recovery rights in these facilities. We can see the risk-management direction. We cannot see every clause that decides what happens after failure.

The old infrastructure lesson inside new hardware

Airports, toll roads and telecom towers have long borrowed against physical assets paired with visible cash flows. AI infrastructure belongs in that family, with one uncomfortable difference: the equipment cycle may move faster.

Meta’s Richland Parish data-centre joint venture is the larger reference point. In October 2025, Meta, the US technology company behind Facebook and Instagram, used $27.3 billion in senior secured project-finance bonds with a residual-value guarantee mechanism, according to Legal & General Investment Management.

A residual value is what an asset may be worth after it has been used. It sounds like a footnote. In a GPU loan, it can be central.

If hardware can be reassigned or sold, debt looks safer. If it cannot, lenders depend more heavily on the original customer and the original demand forecast remaining intact. Debt does not replace operational discipline. It makes operational discipline a condition of finance.

What operators should bring before they ask for debt

Build one joined-up evidence file, not a slide deck with a GPU count.

First, map every equipment purchase to delivery timing, ownership, installation location and intended workload. Then map the site and power path. Where will the hardware run? What must happen between arrival and revenue?

Next, separate contracted demand from projected demand. IREN’s Microsoft-linked contracted cash flows and customer prepayments are not the same thing as a demand forecast. Lambda’s identified investment-grade customer is not the same thing either.

Finally, write the bad-day plan before a lender asks for it. Can equipment be moved? Can it be sold? Which cash flows are actually secured? What rights remain uncertain because contracts or facility documents are not yet complete?

GI Network would turn those strands into an underwriting case: an equipment register, a site-and-power dependency map, a schedule separating contracted revenue from forecast revenue, and a downside recovery analysis. We would then test whether the facts support equipment finance, staged borrowing, equity or a blended structure before capital providers begin diligence.

What investors should ask before applauding the headline

Investors often like secured borrowing because it can reduce the need to issue new shares. Fair enough. But headline facility size is not a risk assessment.

Ask what has actually been de-risked. In IREN’s case, disclosed support included GPU security, contracted cash flows, customer prepayments, ratings and a Microsoft contract. It does not follow that each feature caused the reported cost, or that the same rate is available to another operator. In Lambda’s case, the known investment-grade customer made the August facility easier to understand. Nscale used milestones to control when exposure increased.

Then look at the silent gap between delivery and billing. A company can own valuable servers while earning no income if power, construction or deployment slips. What Red Flags Do Institutional Investors Look for in Due Diligence? matters here because the costliest weaknesses often sit between a persuasive story and the documents beneath it.

The Chain of Five test

Before calling an AI infrastructure project debt-financeable, use the Chain of Five:

1. 1.Equipment: Is the hardware clearly identified, owned and capable of being secured?
2. 2.Place: Is there a ready data-centre location for it?
3. 3.Power: Is power available when the equipment is due to operate?
4. 4.Payment: Is demand contracted, creditworthy or otherwise evidenced rather than merely forecast?
5. 5.Plan B: If the first customer or deployment fails, can the equipment and relevant revenue rights be recovered, reassigned or sold?

Break one link and the GPU is just expensive machinery.

Hold all five together and it begins to look like infrastructure.