The collateral trade: how Bitcoin miners became the AI boom's most interesting credit story in 2026

There is a specific kind of financial engineering happening in the AI infrastructure sector right now, and it does not get covered as financial engineering. It gets covered as a growth story. A company that spent the last seven years mining Bitcoin signs a contract worth more than most countries' annual budgets, a bank rates the resulting debt as investment grade, and the headline is about artificial intelligence. The more interesting headline is about what actually got securitised, and what has not been earned yet.

This is not a story about whether AI infrastructure demand is real. Microsoft, Nvidia and Anthropic are not signing multi-billion dollar contracts as charity. It is a story about the gap between a contract and a cash flow, and about what happens to a highly leveraged, recently-diversified business if that gap closes more slowly than the credit markets currently expect.

From gold mine to brain factory, on paper

IREN's origin has nothing to do with artificial intelligence. The company built its business identifying regions with abundant, cheap, underused electricity, hydroelectric country in British Columbia, wind country in the Texas Panhandle, and using that power to run Bitcoin mining hardware, specialised chips called ASICs that do exactly one thing: solve the cryptographic puzzles that earn new Bitcoin. It was, by most measures, a well-run mining operation, with all-in production costs reported in the low forty-thousand-dollar range against Bitcoin trading well above that for most of the past year.

What that mining business actually built, almost as a byproduct, was a portfolio of grid connections and secured power, now reported at more than 5 gigawatts across sites in the US, Canada and Australia. That is the asset the entire AI pivot depends on, because in 2026 the binding constraint on new AI data centre capacity is not chip availability, it is the multi-year wait to connect a large new electricity load to the grid. A company that already holds the power does not have to wait in that queue.

The deals that changed the valuation

On November 3, 2025, IREN signed a five-year, approximately $9.7 billion contract to provide Microsoft with access to Nvidia GB300 GPU capacity, with a 20 percent prepayment and deployment phased through 2026 at IREN's 750-megawatt Childress, Texas campus. IREN separately committed to buying the GPUs and ancillary equipment from Dell for about $5.8 billion.

Then, on May 7, 2026, IREN signed a second five-year contract, this one worth $3.4 billion, to provide Nvidia itself with managed GPU cloud services for Nvidia's internal AI and research workloads. Alongside that contract, Nvidia received a five-year warrant to purchase up to 30 million IREN shares at $70 each, a potential $2.1 billion equity stake subject to regulatory approval. The company that manufactures the chips is now also, potentially, a shareholder in the company renting them back.

The financing behind the Microsoft contract is arguably the more remarkable development. In June 2026, IREN closed a $3.65 billion GPU financing facility, structured as a $2.1 billion US private placement plus a $1.55 billion delayed-draw term loan, and rated investment grade, A by Fitch and A(low) by DBRS. Combined with Microsoft's prepayment, that financing covered roughly 96 percent of the $5.8 billion GPU capital expenditure tied to the contract, at a blended cost of debt reported near 6 percent. For a business that seven years ago was financing warehouses full of Bitcoin mining rigs, an investment-grade credit rating on GPU-backed debt is a genuinely significant milestone, and it is the clearest evidence available that institutional lenders view Microsoft's contractual commitment as safe, bankable collateral.

The number that matters more than the contract value

Here is where the story gets more complicated than the headlines suggest. A signed, multi-year contract is not the same thing as revenue. It becomes revenue only once the data centre is built, the GPUs are delivered, installed and commissioned, and the customer actually accepts and uses the capacity. Until then, it is what the industry calls annualised recurring revenue "under contract," a forward-looking figure, not a trailing one.

At the time of IREN's Q3 FY2026 results in May 2026, the company reported roughly $3.1 billion in contracted ARR. In the same quarter, its actual AI Cloud services revenue was $33.6 million, having nearly doubled sequentially from $17.3 million the prior quarter. Annualised, that trailing revenue figure is close to $134 million, against a $3.1 billion contracted number. That is a real, verifiable, and enormous gap between the promise and the cash. Total company revenue for the quarter, including the shrinking Bitcoin mining segment, was $144.8 million against analyst expectations of roughly $220 million, a miss of about a third, alongside a net loss of $247.8 million, driven partly by a $140 million write-off from decommissioning mining hardware ahead of GPU installation.

None of this means the contract is not real. Microsoft is about as creditworthy a counterparty as exists, and the prepayment structure de-risks a meaningful share of the build. But it does mean that a market capitalisation built substantially on contracted ARR is pricing in a construction and commissioning timeline executing close to perfectly, on a scale IREN has never operated at before.

Concentration, then diversification

For most of the past year, the more specific risk in IREN's story was customer concentration: by the company's own disclosures, Microsoft represented somewhere in the range of half to three-fifths of contracted ARR. A single customer accounting for the majority of a company's forward revenue is a structural fragility, regardless of how creditworthy that customer is, because any delay on Microsoft's side becomes a delay in IREN's entire growth story.

That picture has shifted meaningfully in the most recent quarter. On July 20, 2026, IREN announced $2.8 billion in new multi-year contracts with a broadened roster of AI developers, including Nvidia, Perplexity, Figure AI, Together AI, Fluidstack, Fireworks AI, Fal AI and Hume AI, and raised its year-end 2026 ARR target from $3.7 billion to more than $4 billion, with approximately 85 percent of that new target now under contract. That is a genuine diversification event, the kind that should, if it holds, reduce single-counterparty risk even as the absolute dollar figures grow. It does not resolve the contract-to-cash gap, since most of this new capacity is not yet built either, but it changes the shape of the risk from "one customer" to "one construction timeline across many customers."

IREN is not alone, and neither is the pattern

The same structure is appearing across the sector, which suggests this is a category, not an idiosyncratic story. TeraWulf, another former Bitcoin miner, announced on July 6, 2026 a 20-year lease with Anthropic at its Justified Data campus in Hawesville, Kentucky, covering approximately 401 megawatts and expected to generate about $19 billion in contracted revenue over the lease term, with initial capacity not arriving until the second half of 2027. On the same day, TeraWulf sold its 50.1 percent stake in a separate Texas joint venture with Fluidstack for roughly $530 million, monetising an existing position to fund the wholly-owned Kentucky build.

Anthropic's compute sourcing itself illustrates how far this circularity now extends. Beyond the TeraWulf lease, the company has committed roughly $1.25 billion a month to lease capacity from SpaceX's Colossus data centres, a $50 billion partnership with Fluidstack for custom US facilities, and was separately reported to be discussing a compute arrangement with Meta, one of its own hyperscaler competitors. Every frontier lab is now sourcing capacity from an expanding web of neoclouds, hyperscalers and, increasingly, each other.

That web became visibly fragile on July 1, 2026, when Bloomberg reported that Meta was developing its own cloud business, tentatively called Meta Compute, to resell excess AI capacity to outside customers. CoreWeave shares fell nearly 14 percent that day and Nebius fell about 17 percent, because Meta is a major contracted customer of both, worth a reported $21 billion to CoreWeave and up to $27 billion to Nebius. IREN also declined, roughly 6.5 percent, a smaller but real reaction given its more diversified customer base at that point. The lesson generalised across the sector in a single trading session: in a market this circular, today's biggest customer can become tomorrow's biggest competitor, and no amount of operational execution insulates a neocloud from that risk.

The Australian tender: a real catalyst, not a guaranteed one

The most closely watched near-term event for IREN specifically is a reported Anthropic tender for at least 1.4 gigawatts of Australian data centre capacity, first reported by the Australian Financial Review, with the total project value cited anywhere from $12 billion to $22 billion depending on the source and scope assumed. IREN is one of five shortlisted bidders alongside CDC Data Centres, AirTrunk, NextDC and Stack Infrastructure, leveraging its 800-megawatt Bundey campus in South Australia. Multiple reports suggest Infratil-backed CDC Data Centres is tipped to receive the largest allocation if Anthropic splits the award across providers, which industry sources consider the more likely outcome. IREN's stock has moved sharply, in both directions, on each new leak and denial in this process, which is itself a signal of how much unearned optionality is already priced into the shares around a deal that has not been signed.

What this means for how you read every neocloud headline from here

The lesson of the IREN and TeraWulf trade is not that these companies are frauds or that their contracts are fake. Microsoft, Nvidia and Anthropic do not sign multi-billion dollar agreements with counterparties they consider unserious, and the investment-grade rating on IREN's GPU financing is real, hard-won evidence that sophisticated credit markets believe in the underlying cash flow. The lesson is that a contracted ARR figure and a trailing revenue figure are different instruments entirely, one is a forward claim on future construction executing on schedule, the other is money that has already changed hands, and the gap between them is exactly the risk an equity investor is being asked to underwrite at today's valuation.

The same logic extends directly to how DN's readers should evaluate DePIN and decentralised compute tokens making similar claims about contracted demand or partnership pipelines. A signed memorandum of understanding, a headline partnership, or a projected network revenue figure is not a receipt. The Contract-to-Cash Ratio below is built to be run against any operator, centralised or decentralised, making a forward revenue claim.

For readers looking to position around neocloud and AI-infrastructure exposure directly, spot and derivatives markets are available through most major exchanges, including Bybit, OKX and MEXC. As always, this is not financial advice. The gap between a contract and a cash flow is measurable. Whether any individual company or token closes that gap on schedule is a judgment call that deserves independent due diligence.

Frequently asked questions