The 24 hours ending 28 August 2026 at 09:01 in Tehran exposed the conversion problem inside the AI infrastructure boom. IREN said it now has $4 billion of contracted annualised run-rate revenue attached to 2026 capacity, but only $1 billion was operating as of 26 August. It also disclosed new GPU financing for deployments serving non-investment-grade customers at a fixed rate as high as 9%.
Marvell supplied the more mature side of the chain: its quarterly data-centre revenue reached $2.172 billion, up 46% year over year. In the UK, the government proposed an innovation objective for the Bank of England's oversight of payment systems, while keeping financial stability explicitly superior. The common lesson is not that demand has weakened. It is that a contract, financed equipment, commissioned capacity, recognised revenue and resilient operation are five different milestones.
The day in four lines
- IREN's 27 August results put $4 billion of contracted 2026 AI-cloud ARR beside $1 billion operating as of 26 August.
- Marvell's 27 August earnings release reported $2.739 billion of total revenue, including $2.172 billion from its data-centre end market.
- HM Treasury's 27 August payments announcement proposed a secondary innovation objective for the Bank of England when it regulates payment systems and digital settlement assets.
- The Bank of England separately deferred its November RTGS standards release after Swift delayed payment-message changes because global readiness was uneven.
IREN's $4 billion is a run rate, not annual revenue already earned
IREN defines ARR as contracted GPU-hour pricing for commissioned GPUs multiplied by 8,760 hours, with annualised storage and ancillary revenue included. Its fresh release says $1 billion of that rate was operating on 26 August. The company targets the full $4 billion to be operational by 31 December, but its own note says that target depends on internal assumptions about utilisation and pricing. Revenue is expected to ramp only after data-centre delivery, commissioning, testing and customer acceptance.
That wording matters. “Contracted ARR” is a useful commercial and capacity measure; it is not the same as GAAP revenue recognised during the fiscal year, cash collected or a guarantee that every planned GPU will enter service on schedule. Multiplying a contracted hourly price by every hour in a year describes the revenue rate of available capacity. It does not remove downtime, delayed acceptance, deployment sequencing, customer credit or renewal risk.
The $4 billion headline is not merely a repeat of an old target. IREN's 20 July filing said more than $4 billion of year-end ARR was targeted and approximately 85% was under contract. The 27 August update now describes $4 billion as contracted, adds a new multi-year agreement with an unnamed frontier AI laboratory and says 2026 capacity is largely sold out. The name, value, commencement schedule and credit quality of the new customer were not disclosed.
Operationally, the most informative ratio is therefore one-to-four: $1 billion operating against $4 billion contracted. It is not a margin or probability estimate. It is a commissioning gap that should shrink only when sites, power, cooling, GPUs, software and customers are accepted together.
The filed year shows the transition before the run-rate ramp
IREN's 2026 Form 10-K filed on 27 August covers a different clock: the year ended 30 June. It reported $707.0 million of total revenue, of which AI Cloud Services contributed $128.8 million and Bitcoin mining contributed $578.2 million. AI-cloud revenue was almost eight times the prior year's $16.4 million, but it remained 18% of annual revenue.
The quarter shows the direction more clearly. AI-cloud revenue rose to $70.5 million from $33.6 million in the March quarter, while mining revenue fell to $66.7 million from $111.2 million. That does not reconcile directly to the $1 billion operating ARR announced in August: the quarter ended almost two months earlier, ARR uses a point-in-time annualised contract rate, and subsequent capacity can enter operation without appearing in June revenue.
The transition also carried a large accounting cost. IREN recorded a $702.6 million net loss for the year after an $86.9 million profit in fiscal 2025. Management attributed most of the change to $638.8 million of non-cash impairment, primarily for Bitcoin-mining hardware being decommissioned as sites convert to AI cloud. Adjusted EBITDA was $245.7 million, down from $269.7 million. Neither number should be read alone: the impairment records the loss of value in the old fleet, while adjusted EBITDA excludes costs that still affect capital providers and shareholders.
The balance sheet confirms the physical scale of the pivot. At 30 June, IREN reported $5.896 billion of cash, $1.670 billion of current restricted cash and $6.753 billion of property, plant and equipment. Total liabilities were $11.604 billion. High liquidity and high commitments can coexist when cash is raised or restricted for equipment that has not yet become an operating service.
Financing reduces the upfront gap but adds a repayment clock
IREN separated its GPU financing by customer quality. A $3.6 billion investment-grade financing for its Microsoft contract carries a 6% rate; together with customer prepayments, IREN says it funds 96% of the associated GPU capital expenditure. For non-investment-grade customer deployments, the company announced $2.8 billion of new financing. That includes $2.4 billion led by Blue Owl and Pacific Investment Management Company, acting as adviser to certain investors, at a 9% fixed rate for the Mackenzie expansion, funding 90% of the associated GPU capex.
Recent customer prepayments represent 45% to 55% of GPU capex, according to IREN. Those structures can reduce the equity cash required before deployment and align some funding with customer contracts. They do not make the build free. Interest, principal, construction, power, networking, buildings, operations and equipment not covered by the GPU facility still need to be paid.
The rate difference is itself operational evidence. Investment-grade contracted cash flows supported 6%; other deployments were financed at 9%. Actual interest expense will depend on draw timing, amortisation and terms that the earnings summary does not fully disclose. A buyer or operator should model the schedule of cash receipts against debt service, not compare an annualised revenue headline with a financing principal as if both arrived on day one.
This is the same distinction visible in the earlier Nebius convertible-financing analysis: access to capital can accelerate capacity, but it cannot prove commissioning, utilisation or return on the completed system.
Marvell shows the supplier side is already recognising revenue
Marvell's filed 8-K and earnings exhibit provide a cleaner measure of delivered supplier demand. Revenue for the quarter ended 1 August was $2.739 billion, up 37% year over year. The data-centre end market produced $2.172 billion, up 46% year over year and 18% sequentially, and represented 79% of company revenue.
That $2.172 billion should not be relabelled “AI revenue”. Marvell's data-centre category includes AI systems and servers, but also Ethernet switching, network-attached storage, general-purpose servers, storage systems and data-centre interconnect. Management said AI-related bookings were robust and forecast acceleration in its custom business during the second half; the category table does not isolate the amount earned from AI workloads.
The filed economics are nevertheless substantial. GAAP gross margin was 53.1%, GAAP net income was $308.0 million and operating cash flow was $605.5 million. Marvell guided to $3.150 billion of current-quarter revenue, plus or minus 5%, and a 52.9% to 53.9% GAAP gross margin. Those are forecasts, but the reported quarter already shows components, connectivity and custom silicon converting infrastructure demand into revenue.
Read beside the previous NVIDIA earnings and AWS capacity brief, the chain is clearer. Semiconductor suppliers can recognise record sales while an operator is still financing and commissioning the capacity that will use them. Supplier revenue is evidence of spending; it is not evidence of the end customer's utilisation or the operator's debt coverage.
UK payments policy kept innovation subordinate to stability
HM Treasury first published its announcement at 10:04 British Summer Time on 27 August, inside this brief's window. The government intends to add a secondary innovation objective to the Bank of England's regulation of systemic payment systems, including systems using digital settlement assets such as stablecoins. The Bank would report annually to Parliament on how it advanced that objective.
This is a policy proposal, not a rule already in force. The government says financial stability will remain the Bank's primary objective and the innovation duty would not require support for an arrangement that undermines it. Implementation requires amendments to the Financial Services and Markets Bill. Parliament's official bill API lists report stage from 7 September and confirms that the bill is not yet an Act.
The same-day RTGS delay makes the hierarchy concrete. Swift said more than 98% of payment instructions now use ISO 20022, but readiness for structured postal addresses remained uneven. The Bank's statement then deferred its entire November release for RTGS and CHAPS rather than separate interdependent changes late in the cycle. That is not opposition to innovation; it is change control around globally connected infrastructure.
For payment teams, the practical discipline remains the one in AI infrastructure: label the stage. A proposed objective is not enacted law; a standard is not an implemented message flow; and an innovative payment rail is not resilient until participants, reconciliation, fraud controls, recovery and customer outcomes work together. The payments operations control guide covers those release boundaries in more detail.
What is confirmed and what remains open
| Development | Confirmed during the window | Important limit |
|---|---|---|
| IREN capacity | $4B contracted ARR; $1B operating on 26 August | $4B operation depends on delivery, testing, acceptance, utilisation and pricing assumptions |
| IREN financing | $3.6B at 6%; new $2.8B package including $2.4B at 9% | Draw schedule, full covenants and realised debt-service coverage were not provided in the release |
| Marvell quarter | $2.739B total revenue; $2.172B data-centre revenue | Data-centre revenue is broader than AI, and custom acceleration is forward-looking |
| UK payments objective | Government intends a subordinate Bank innovation duty | Parliamentary approval and final text remain outstanding |
| RTGS standards | November release deferred to preserve alignment and reduce implementation risk | Revised dates were not yet set |
What operators should measure next
- IREN's accepted megawatts and commissioned GPU count, not only contracted capacity.
- Monthly recognised AI-cloud revenue and cash collection as the $1 billion operating run rate expands.
- Interest expense, restricted cash, customer concentration and debt-service coverage after the new facilities are drawn.
- Marvell's data-centre mix, gross margin and custom-silicon revenue as its second-half forecast reaches reported accounts.
- The exact Bank of England amendment, parliamentary changes and annual reporting measures for payment innovation.
- Swift and the Bank's revised ISO 20022 and RTGS milestones, including participant readiness evidence before cutover.
The strongest signal from the day is that AI infrastructure finance is becoming more specific. Companies are disclosing which revenue is earned, which is annualised, which capacity is contracted, which hardware is financed and which risks remain outside the headline. That vocabulary is valuable because it makes execution visible. The bridge from signature to service is real, but it still has to be built, funded, accepted and operated.
