Why IREN Thinks AI Compute Supply May Never Catch Up With Demand

IREN’s Daniel Roberts sees AI infrastructure as a structural supply problem rather than a normal boom-and-bust investment cycle. His central argument is that every increase in available computing capacity can enable more capable models and new AI applications, creating additional demand, while the physical inputs behind that capacity take far longer to build. 13

That view helps explain why IREN is proposing as much as $30 billion in fiscal 2027 capital expenditure as it shifts from Bitcoin mining toward AI cloud infrastructure. It also explains why investors should distinguish between the company’s demand thesis and its ability to execute an unusually capital-intensive plan.

The argument: digital demand can expand faster than physical supply

Roberts’ position is that AI demand is self-reinforcing. More available compute can support larger training runs, higher-volume inference and more commercial applications. By contrast, the supply side depends on physical systems: power generation and transmission, grid interconnection, land, data-center construction, cooling, GPUs and the teams that deploy them. IREN has described this as a widening gap between digital demand and physical supply. 13

That is what makes the current cycle different, in Roberts’ telling. In a conventional commodity cycle, elevated prices can attract investment that eventually produces excess capacity. AI infrastructure may be harder to oversupply because building energized, operating GPU capacity is constrained by long lead times and power availability—not merely by willingness to spend capital. 113

Goldman Sachs’ own outlook supports the premise that power is a central constraint, while not proving IREN’s conclusion that supply will never catch up. Goldman raised its estimate for U.S. data-center power demand in 2030 to about 108 GW, from 83 GW previously, and lifted its growth forecast to a 3.5% compound annual growth rate through 2030. 19 Separate reporting attributed to Goldman estimates that U.S. data-center capacity could more than triple to roughly 125 GW by 2030; because capacity and power-demand measures are not necessarily identical, these figures should not be treated as interchangeable. 20

How IREN says it can fund up to $30 billion of capex

The key to IREN’s financing argument is that the $25–$30 billion figure represents total project capital expenditure, not an equivalent requirement for new common equity. Roberts said customer prepayments could cover roughly half of GPU capital expenditure and lenders could finance much of the remainder through secured or GPU-backed structures. 14

IREN reported that it had secured $9.3 billion of funding over eight months across customer prepayments, convertible notes, GPU leasing and GPU financing. 32 Reporting on the company’s capital plan also said that $3 billion of $19 billion raised had come from equity. 2

The most concrete example is IREN’s Microsoft agreement. IREN announced an approximately $9.7 billion, five-year contract to provide GPU cloud infrastructure using Nvidia GB300 systems, deployed in four phases totaling 200 MW of critical IT load. The agreement includes a 20% customer prepayment, according to IREN and Data Center Dynamics. 2533 A separate $3.65 billion GPU-financing facility was reported to support that Microsoft contract. 27

This structure can reduce the amount of upfront equity required, but it does not eliminate financing risk. The model depends on creditworthy customers, timely prepayments, functioning debt and equipment-finance markets, GPU collateral values, delivery schedules and the ability to bring data-center power online as planned.

Why the $684 million loss needs context

IREN reported a $684 million quarterly net loss during its transition from mining to AI cloud. Management and subsequent reporting attributed much of that result to non-cash charges associated with retiring or selling mining hardware: a $450.4 million impairment and a $102.1 million fair-value decline on mining hardware held for sale. 67

Those charges do not by themselves represent the cash cost of the AI expansion. Still, they are economically meaningful: they reflect the cost of winding down older assets while the company undertakes a much larger capital buildout.

IREN pointed to about $4 billion in contracted annualized run-rate revenue tied to 2026 AI capacity, which it said was largely sold out. Its AI-cloud revenue was reported at about $70.5 million in the cited quarter. 68 Contracted revenue can improve the bankability of infrastructure projects, but it is not the same as cash flow already received or proof that future capacity will earn comparable economics.

Rising contract prices are central to the economics

Management said its three-year contract pricing had risen about 125% since November, while five-year pricing had increased about 70%. Recent three-year contracts were priced above $20 million per MW of IT load, and management said active discussions were around $25 million per MW. It characterized the earlier level as supporting an approximate two-year payback on compute investment. 1217

Those statements are important because they link the supply-shortage thesis to project returns. If AI capacity remains scarce, providers with delivered power, installed GPUs and committed customers may be able to sign high-value contracts before capacity becomes broadly available. If pricing falls, hardware economics weaken, or facilities are delayed, the same capital intensity becomes a major vulnerability.

Where neoclouds fit against hyperscalers

IREN, CoreWeave and Nebius are part of a specialist GPU-cloud, or “neocloud,” market. Their pitch is more focused AI infrastructure, direct capacity access and potentially lower rates than a general-purpose hyperscaler cloud.

Published price comparisons suggest meaningful differences, but they should be treated carefully. One comparison estimated that eight B200 nodes cost $329,472 for 30 days at Nebius versus $742,349 at a hyperscaler; the comparison noted differences in pricing structure and sourcing, so it is not a like-for-like benchmark for every workload. 49 Another published comparison listed on-demand H100 pricing at $3.85 per GPU-hour at Nebius and $6.16 at CoreWeave. 53

List price alone is not the purchasing decision. GPU generation, interconnect, storage, geographic location, reservation term, availability, support and service-level commitments all affect the effective cost. The broader competitive point is that specialist clouds may compete with hyperscalers on speed of access and AI-optimized capacity as well as on advertised hourly pricing.

IREN’s major contracts

IREN has announced two agreements that underpin its AI-cloud positioning:

• A five-year, $3.4 billion computing-capacity lease agreement with Nvidia, reported as signed in May. 20
• An approximately $9.7 billion, five-year Microsoft agreement for Nvidia GB300-powered GPU cloud infrastructure across four 50-MW phases, totaling 200 MW of critical IT load. 2533

IREN also said the first 50-MW Horizon deployment at its Childress, Texas site had been delivered to Microsoft. 12

The bottom line

Roberts’ thesis is compelling in a narrow sense: AI workloads can expand rapidly, while power and data-center infrastructure cannot be built instantly. Goldman’s higher power-demand outlook and the value of IREN’s disclosed contracts show why investors are focused on energized AI capacity. 1925

But “supply may never catch up” remains an executive’s strategic judgment, not an established fact. IREN’s outcome will depend on whether it can convert contracted demand into operating, financed infrastructure—without power delays, hardware shortfalls, financing stress or a downturn in AI-compute pricing. The company’s large prepayments, financing arrangements and contracted revenue provide support for the plan; they do not remove its execution risk.