Investment Gems

FinTwit can short the second derivative. Apparently the customers didn’t get the memo.

There’s something hilarious happening in AI infrastructure right now. The market has decided we’re approaching the end of the capex supercycle, neoclouds are glorified leveraged GPU warehouses and everyone is about to discover what depreciation means. Fair enough, those are legit concerns.

But then $IREN reports earnings and casually tells us that customers are paying more for compute, old H100s are getting recontracted at DOUBLE their previous price, and one customer just agreed to prepay 90% of the GPU capex on a 5 year contract. Meanwhile management is deliberately slowing expansion because financial markets have become less receptive to funding it. Read that again. The customers want the compute. The financiers are getting cold feet. These are two VERY different markets, and right now they’re telling us two completely different stories.

The numbers, before we start hallucinating about a bubble

First, the actual quarter. IREN is transitioning from Bitcoin mining into an AI cloud business, and we’re finally getting something resembling an operating income statement rather than a collection of construction-site drone videos.

Metric	Q4 FY26	Q1 FY27	Change
AI Cloud revenue	$71M	$200M	+182%
Bitcoin mining revenue	$67M	$28M	−58%
Total revenue	$137M	$228M	+66%
AI Cloud gross margin*	86%	86%	Flat
Adjusted EBITDA	$19M	$66M	+247%
Adjusted EBITDA margin	14%	29%	+15 pp
Operating ARR	$0.5B	$1.0B	+100%
Revenue backlog	~$16.6B	$20.3B	+22%
GAAP net loss	−$684M	−$263M	Improving

AI Cloud segment gross margin excludes depreciation and amortization and certain site-level costs. Operating ARR is an annualized metric, not GAAP revenue.

The $200M of AI Cloud revenue includes $135M in services and $65M in lease revenue. So this isn’t just GPUs running some experimental workload. Microsoft has accepted Horizon 1, and operating infrastructure is actually generating recognizable revenue. Total company revenue is still slightly below the $240M reported a year earlier, because the Bitcoin business is disappearing. That’s an important distinction. This isn’t yet a clean year-over-year growth story for the whole company. It’s a rather violent transformation of what the company actually does.

IREN has a $20.3B backlog. About $9.6B belongs to Microsoft, $3.4B to NVIDIA and $7.2B to AI developers, frontier labs and channel partners. Roughly two-thirds of that backlog is with investment-grade counterparties. That’s important because a $1B contract with Microsoft and a $1B contract with a venture-funded AI startup are not interchangeable when you’re trying to convince a lender to finance $800M of hardware. Credit quality changes the economics of the contract. Also, yes, NVIDIA is both a supplier and a customer within this ecosystem. The circularity crowd isn’t completely insane to pay attention to that. But the relevant test isn’t whether two companies do business with each other. It’s whether contractual cash flows arrive and the underlying assets generate acceptable returns.

And the recent pricing is pretty difficult to square with the idea that compute demand has suddenly fallen off a cliff. After quarter-end, IREN signed a five-year contract at approximately $22M of annual revenue per IT megawatt, a shorter-duration deal at approximately $40M per IT megawatt annualized, and expanded an existing agreement with a frontier AI lab. It also recontracted older H100 GPUs at twice their previous contracted price, reportedly for a multiyear term.

Don’t be an amateur and extrapolate a short-term $40M/MW deal into five years of revenue. Nor does doubling one H100 contract prove the entire GPU market is repricing higher. But it does present a rather awkward problem for the crowd insisting older GPUs automatically become worthless the minute Jensen puts on a new leather jacket.

The CEO basically admitted our entire macro thesis

The most interesting part of the earnings call wasn’t revenue. It was an analyst noticing that IREN’s capex expectations had come down.

Management’s response was unusually revealing. Yes, they’re deliberately reducing the pace of spending. No, they don’t attribute that decision to a shortage of customers or development opportunities. They attribute it to the availability and cost of capital in the current financial environment.

Well, there it is.

We’ve been discussing the possibility that the market is confusing a liquidity contraction with the collapse of an investment cycle. Here we have an operator telling us, almost explicitly, that its investment decisions are being constrained by financing conditions while its customers are still bidding aggressively for compute.

That creates a pretty interesting feedback loop. Neocloud equities collapse, equity issuance gets more expensive, lenders demand higher returns, operators defer GPU purchases and construction, fewer new clusters get commissioned, and the capacity that IS commissioned potentially becomes more valuable. Then some genius points at the slowdown in investment and announces that AI demand has peaked. Congratulations, you’ve discovered that supply is sensitive to the cost of capital.

This doesn’t make IREN immune. The company itself has to finance a mountain of additional investment. But it does suggest that the second derivative of capex may tell us more about financing conditions than about the underlying demand for intelligence.

And that’s a much more interesting question than whether hyperscalers spend $800B or $900B next year.

The 90% prepayment is the real financial engineering here

Here’s where the story gets properly nerdy.

One of IREN’s recent five-year customers agreed to prepay an amount covering approximately 90% of the associated GPU capex. Not 90% of total data center investment, mind you. Just the GPU component. Still, that’s huge.

Management’s presentation provides this illustrative contract:

Per IT megawatt	No prepayment	90% GPU prepayment
Annual revenue	$20M	$20M
Data center capex	$20M	$20M
GPU capex	$40M	$40M
Total physical capex	$60M	$60M
Customer prepayment	$0	$36M
GPU cash payback	2.4 years	0.4 years
GPU + data center cash payback	3.5 years	2.4 years

These are illustrative assumptions, not the forecast for any particular contract. But the mechanism is extremely important.

The customer is effectively advancing $36M against future service payments, leaving IREN with $24M of initial net cash funding to arrange against a $60M deployment. The same infrastructure is built, the same GPUs are purchased and the contracted lifetime revenue hasn’t magically increased. What changes is the timing and source of capital.

This is where I want people to be careful. That 0.4-year GPU payback number looks like something produced by a hedge fund analyst after his third espresso. It is NOT saying a $40M GPU investment generated $40M of economic profit in five months. A large part of the future customer’s bill has simply been collected upfront. Economically, you’re monetizing future receivables before providing the service.

But that is still extremely valuable when external financing costs are rising.

Customer prepayments can reduce the need to borrow at 9%, issue equity at depressed valuations or keep expensive GPUs sitting around waiting for the revenue to arrive. They shift financing risk away from IREN’s shareholders and lenders toward a customer that actually needs the compute.

This is the kind of capital recycling that could seperate the surviving neoclouds from the ones that end up as very expensive lessons in spreadsheet optimism.

And the beauty is that the physical data center remains IREN’s asset. The GPUs may need refreshing in several years, but the land, grid connection, electrical infrastructure and cooling systems are designed to last much longer. Which brings us to the part of the business I think the market may still be underappreciating.

The GPU is the depreciating asset. The powered site is the option.

IREN has approximately $6B of data center assets that are currently unencumbered. Management targets $10B by June 2027. These include land, buildings and construction in progress, with the long-lived physical infrastructure generally depreciated over roughly 20 years.

Compare that with GPUs, where obsolescence and replacement cycles are much shorter.

It’s basically two businesses with two very different clocks. One is buying extremely expensive computers that can lose economic relevance relatively quickly. The other is developing locations where those computers can actually be operated, cooled, networked and connected to electricity. The second business is much harder to replicate on short notice.

This is why I found management’s discussion of its next-generation development strategy more interesting than the usual gigawatt chest-beating. They discussed using NVIDIA’s power-management software to install more GPUs inside a fixed grid power envelope, dynamically managing consumption instead of assuming every GPU runs at maximum rated power all day. They also discussed operating at higher temperatures to reduce cooling power consumption and increase the share of electricity available for computation.

None of that has yet produced a quantified, independently demonstrated increase in returns. So let’s not prematurely mark it as free money. But the direction is fascinating. If the scarce asset is a grid-connected megawatt, increasing billable compute per megawatt might create more value than simply owning another rack of GPUs.

And Mirantis fits into the same picture. IREN isn’t just trying to own servers. It’s building the software and operational capability to make entire AI clusters usable and supportable. The point is to sell reliable, commissioned compute rather than hand someone a warehouse full of blinking lights and wish them good luck with Kubernetes.

Management says Horizon 1 went from civil construction to an NVIDIA-validated GB300 cluster in 14 months. It compares that with an 18-month industry benchmark for data center construction alone. That’s not a perfectly matched comparison, but handing over an operational cluster is nevertheless a much more meaningful achievement than announcing another site acquisition.

The increasingly scarce product may be time-to-compute, not the chip itself.

Unfortunately, the balance sheet did not get the memo

Now for the part the bullposters tend to scroll past.

IREN still needs a truly absurd amount of money.

Funding position, September 30	Amount
Remaining planned FY27 capex	$15.0B
Cash and restricted cash	$5.8B
Committed GPU funding and prepayments	$3.2B
Additional financing still targeted	$6.0B
Total debt principal outstanding	~$10.0B
Unrestricted cash	~$3.5B

Management identifies $12–15B of potential financing capacity from data center collateral, GPU financing, prepayments and other corporate sources. Sounds reassuring. Except potential funding capacity isn’t committed funding. Much of it depends on target collateral values, future assets, actual lender terms and market conditions. You don’t get to put a financing presentation in the bank and withdraw the money.

And the financing ain’t cheap everywhere. The Microsoft-related GPU financing carries roughly 6% weighted-average stated cost, supported by a high-quality customer. Other GPU financings carry rates around 9–9.5%. Importantly, some of these secured equipment financings are guaranteed by IREN itself, not cleanly ring-fenced away from the parent company. That’s a meaningful distinction when people start describing everything as harmless asset-level leverage.

There is another inconvenient detail in the 10-Q. IREN reports approximately $18.9B of purchase and development commitments, including hardware and infrastructure obligations. A separate approximately $6.7B Dell hardware order signed in September provides for deliveries beginning in the second half of 2027, with IREN guaranteeing the purchasing subsidiaries’ obligations. These amounts overlap, so don’t add them together. Some purchase commitments can be cancelled or adjusted, subject to contractual costs. But the idea that IREN can simply turn off all capex tomorrow and sit on its land like a patient Berkshire Hathaway is abit too cute.

The company has flexibility. It does not have unlimited flexibility.

And the $4B operating ARR target has slipped from December to January, potentially February. Microsoft’s requested network changes and supplier timing explain the delay, and the 10-Q indicates contractual protection from delay credits attributable to Microsoft-requested changes. Still, this is a reminder that the last mile matters. A data center with energised racks is not the same as customer-accepted, revenue-producing compute.

The funded approximately 300MW program is one thing. Getting from there toward the 800MW year-end 2027 capacity target is another. That’s where the remaining funding gap becomes relevant.

About that 86% gross margin everyone is celebrating

Here is the bit where we have to ruin the party with accounting.

IREN’s AI Cloud segment gross margin is 86%, but that is calculated before depreciation and amortization, and excludes certain site-level expenses. Useful for understanding direct service economics? Absolutely. A measure of the return available to common shareholders? Not remotely.

Look at the quarter.

Q1 FY27 item	Amount
AI Cloud revenue	$200M
AI Cloud segment gross profit, before D&A	$172M
Total adjusted EBITDA	$66M
Depreciation and amortization	$148M
Stock-based compensation	$106M
Finance expense	$73M
GAAP net loss	−$263M
Operating cash flow	+$529M
Increase in deferred revenue	+$645M
Cash spent on PP&E and computer hardware	$4.82B

Yes, adjusted EBITDA is improving. Yes, a substantial portion of depreciation relates to assets being deployed ahead of the full revenue ramp. Both points matter.

But the cash-flow statement is telling us something else too. Operating cash flow was positive $529M largely because customers advanced cash. Deferred revenue increased by $645M. Remove just that increase and operating cash flow would have been approximately negative $116M, before adjusting anything else. Those advances are real financing advantages, but they’re not recurring profits from the installed base.

And then there’s the shareholder count.

IREN had approximately 380M ordinary shares outstanding at the end of June, 407M at the end of September and 424M by October 30. That’s nearly 12% more shares in four months. The at-the-market issuance program has been a genuine source of funding, but existing shareholders are paying for part of the expansion through dilution.

So when somebody tells you AI Cloud revenue is growing at 182%, the appropriate response isn’t necessarily to sell the stock. It’s to ask how much of that growth is eventually going to belong to each share after interest, depreciation, taxes, future hardware purchases and whatever new shares get issued along the way.

It’s called investing. Apparently it’s become controversial.

So who is actually winning the AI capital cycle?

Here’s my working interpretation.

NVIDIA sells the hardware and captures much of the hardware economics when systems ship. IREN finances, constructs and operates the compute infrastructure, absorbing execution risk and the burden of ongoing capital investment. Microsoft, NVIDIA and other large customers obtain access to scarce capacity, sometimes helping fund it through contractual commitments or prepayments. Lenders collect interest and hold claims against equipment or other assets. Common shareholders are entitled to whatever value remains once all those obligations have been met.

The current financing squeeze changes the bargaining power between those groups.

If credit is abundant, almost anybody with a large enough PowerPoint can announce a multibillion-dollar GPU cluster. If credit becomes scarce, the ability to deliver working compute starts to matter more. So does customer credit quality. So does the ability to fund GPUs against contracted revenue. So does owning long-lived physical infrastructure without an enormous fixed colocation bill.

That should favor certain operators over others. But the winners won’t necessarily be the companies with the biggest gigawatt forecasts. They’ll be the ones earning the most durable return per dollar of capital actually at risk, and ultimately per share.

IREN’s new contracts suggest it may be improving that equation. The 90% prepayment is a much more meaningful data point than another grand announcement about future capacity, because it tells us customers are willing to put real capital behind future compute.

On the other hand, $18.9B of commitments, $10B of debt, substantial depreciation and a rising share count mean IREN is still very much exposed to the financial cycle. Owning the infrastructure makes it more interesting. It doesn’t repeal the laws of corporate finance.

What would make me wrong?

Pretty straightforward, actually. If IREN starts missing commissioning milestones repeatedly, if contracted GPU prices roll over across generations, if the attractive prepayments disappear, or if lenders will only fund the next expansion on increasingly punitive terms, then the scarcity thesis begins losing its force.

And I’d watch the gap between operating ARR and recognized revenue like a hawk. Management expects roughly $250M of revenue in the December quarter from the operating ARR base at September’s end, while cash operating expenses are expected to increase another $50–60M sequentially. The real payoff from the additional clusters will only become visible after delivery, testing and acceptance. The quarter following commissioning matters considerably more than an attractive annualized headline.

There’s also a trap in the company’s 35GW development pipeline. Management itself acknowledges that much of this is speculative and doesn’t expect a significant portion to become secured power. It’s optionality, not commissioned capacity, and certainly not $35GW worth of free cash flow. Let’s not turn an interesting infrastructure business into another Excel-generated religion.

My conclusion

I came away from these earnings more convinced about the AI infrastructure scarcity thesis, and somewhat more constructive on IREN’s competitive position. But I’m still not prepared to confuse the value of the assets with the returns available to shareholders.

The evidence from actual contracting is difficult to reconcile with an imminent, across-the-board compute glut. The evidence from actual financing is much easier to reconcile with a market where productive capital formation is being constrained by expensive and selective money.

And that’s the part I think deserves more attention.

IREN is discovering that the most valuable part of the AI buildout may not be the GPU. It may be the ability to turn scarce megawatts into functioning compute, and then turn contracted compute into financeable cash flows.

The customer prepayment deals are especially interesting because they begin to solve the financing bottleneck from inside the commercial relationship itself. The customer wants the compute badly enough to help pay for it before receiving the service.

But there’s no free lunch. Somebody still has to finance the buildings, bear the execution risk, replace the hardware, service the debt and keep the platform running. And right now the common shareholder is still writing some pretty large checks.

So yes, I like what I’m seeing operationally. The business is beginning to look less like a speculative GPU renter and more like an integrated compute infrastructure platform.

What I need to see next is that the company can translate that progress into returns without continuously expanding the share count.

Until then, my position is simple: bullish on the scarcity of commissioned AI capacity, bullish on the value of financeable infrastructure, and very much awake to the possibility that the lenders and customers capture the best part of the economics before common equity gets its turn.

The AI revolution can be real and the stock can still be a bad investment. Both things can be true. That’s why we read the damn filings.

---