# AI capex — X 热门讨论 (2026-09-25 01:26 UTC)
## @tszzl (roon) · 09-25 00:02 · ♥351 ↻15 💬54 Iran / oil prices is a short term driver but broadly this is due to demand pressure from extremely attractive ai capex opportunities that are borrowing at nation scale > 引用 @Barchart: JUST IN 🚨: U.S. 10-Year Treasury Yield jumps to highest level since the run-up to the Global Financial Crisis https://x.com/tszzl/status/2103273813928759398
## @StockAnalystPro (StockAnalystPro) · 09-24 23:44 · ♥41 ↻2 💬6 EXCELLENT INTERVIEW of $IREN CCO @kentpdraper, answers Semi Analysis, Q4 2026, 2027, $NVDA , 2028. Thank you @McnallieM You are amazing.
My take on IREN’s Q4 FY26 update, the 2027 NVIDIA ramp, Mirantis, and what the company still has to prove.
Kent Draper’s interview reinforced my view that the AI infrastructure opportunity is enormous. It also put the main bottleneck in focus: IREN must turn secured power and contracted demand into reliable, operating AI Cloud revenue.
His response to SemiAnalysis is an important part of that story. Kent said IREN had not provided a live, representative managed-services platform for the review because that offering was still being developed. He also pointed to fuller power and mechanical redundancy in future builds. That explains a limitation of the managed-services assessment, but it does not make the reliability concerns disappear. IREN now has to demonstrate the improvement in operating clusters.
Here is my breakdown.
Q4 FY26: Contracted ARR versus operating ARR
IREN reported $4B of contracted annualized run-rate revenue (ARR) tied to 2026 capacity, of which $1B was operating at the FY26 update.
That distinction matters. A signed contract does not become operating ARR until capacity is built, commissioned, handed over, and running for the customer. FY26 recognized AI Cloud Services revenue was $128.8M, showing how early the revenue conversion still was relative to the contracted run rate.
For me, the question is no longer whether IREN can announce large contracts. It is how quickly the remaining contracted capacity becomes operational—and how reliably it runs once delivered.
Where NVIDIA fits in 2027
IREN expects approximately $700M of ARR from its NVIDIA cloud contract to ramp in 2027. Management explicitly said this is outside the $4B contracted ARR figure for 2026 capacity. I would keep those figures separate in any model.
The relationship also brings technical validation. IREN reported NVIDIA Exemplar Cloud status for its GB300 NVL72 deployment, while Mirantis has strengthened the software side of IREN’s NVIDIA ecosystem. The next milestone is turning that relationship into deployed capacity and operating revenue.
Horizon: Can IREN repeat the first delivery?
IREN delivered Horizon 1, the first of four 50 MW IT liquid-cooled deployments. At the FY26 update, Horizon 2 was in commissioning and Horizons 3 and 4 were in late-stage construction, targeting delivery in Q4 calendar 2026.
The value of a repeatable design is speed and learning. But each phase still needs successful construction, GPU installation, testing, customer acceptance, and stable operation. That is why I watch completed handovers more closely than planned megawatts alone.
Liquid cooling helps IREN support dense, power-hungry AI systems. It is an advantage when delivered well, and another system that must perform consistently once a customer’s cluster is live.
Vertical integration: The full stack
Kent described IREN’s platform across three layers:
1. Data centers: Land, power, substations, buildings, cooling, and connectivity.
2. Compute: GPUs, servers, networking, and storage.
3. Software and services: Orchestration, deployment, monitoring, customer support, and managed cloud offerings.
IREN also acts as its own general contractor on key projects. That gives it more direct control over construction schedules, equipment coordination, and delivery. In a market where customers want compute quickly, time to operating capacity can be a competitive advantage.
Owning the full stack also raises the execution bar. IREN has to build the facility, finance and deploy the GPUs, and deliver a dependable customer experience.
Why Mirantis matters
Large hyperscalers and frontier labs often bring their own orchestration software and may prefer bare metal. Smaller AI developers and enterprises may need a provider to manage more of the stack.
Kent said Mirantis helps IREN serve that second group through orchestration, deployment, monitoring, and enterprise support. It also creates a path toward reserved managed clusters and potentially on-demand compute. That could broaden IREN’s customer base and allow it to earn more value from the same underlying infrastructure.
I see the strategic logic. I also want to see the operating proof: a mature service, consistent support across sites, and customers renewing because the platform works well for them.
Kent’s answer to the SemiAnalysis question
SemiAnalysis’s September ClusterMAX 3.0 report was sharply critical of IREN. It cited customer-reported power, networking, storage, and GPU reliability problems at its British Columbia sites and questioned the maturity of its managed-cluster offering. SemiAnalysis also said the newer Childress and Sweetwater builds appeared better designed for redundancy. Those are SemiAnalysis’s findings and reported customer experiences, rather than incidents I have independently verified.
Kent’s response had two parts.
First, he said IREN did not submit a live managed-services environment representative of its intended offering for testing at that time. The platform was still under development. That matters when interpreting a rating of its managed-service capabilities.
Second, he said future builds would include full power and mechanical redundancy as a standard design feature. That speaks to the infrastructure concerns, especially as IREN expands beyond its earlier sites.
My view is that Kent provided useful context, but the debate will be settled by results. A developing platform explains what was available to test; it does not answer every report about live-site reliability. The proof will be sustained uptime, resilient power and networking, effective incident response, and a managed service that customers and independent reviewers can test at scale.
Funding the next phase
IREN said it secured approximately $19B in funding over the preceding 12 months. That included roughly $3B of equity, with the rest coming from sources including customer prepayments, GPU financing, and convertible notes. The $19B should be understood as funding secured, not as unrestricted cash on the balance sheet.
IREN reported $3.6B of investment-grade GPU financing at about 6%. For non-investment-grade customer deployments, it secured $2.8B of GPU financing; a $2.4B Mackenzie portion carries a 9% fixed rate. Recent customer prepayments covered 45%–55% of associated GPU capex on the deals IREN described.
This is a powerful financing model if customer demand, asset performance, and delivery continue to support it. Yet IREN guided to approximately $25B–$30B of FY27 capex, so the cost and structure of future funding remain central to the investment case.
Texas and the longer pipeline
Texas gives IREN room to scale at Childress and Sweetwater. Kent highlighted Sweetwater’s 2 GW power position, existing grid infrastructure work, and the importance of building on a schedule customers can use.
IREN is also developing opportunities in Oklahoma, Spain, and Australia, alongside additional deployments at existing Canadian sites.
I separate secured power, planned IT capacity, contracted ARR, and operating ARR. They represent different points on the path to revenue. A large power position is valuable; each subsequent step still takes capital and execution.
IREN’s main growth bottlenecks
1. Converting power into usable IT capacity. Grid access is the starting point. Data halls, cooling, substations, and networking have to be ready together.
2. Delivering projects on time. Horizon 1 is a proof point. Repeating the handover across later phases and new sites is the test.
3. Commissioning GPU clusters. Hardware delivery alone does not create operating ARR. Clusters must be installed, tested, accepted, and available to customers.
4. Reliability. SemiAnalysis put a spotlight on power, networking, storage, and customer experience. IREN needs measurable improvement and sustained performance.
5. Scaling managed services. Mirantis brings software and expertise. IREN must integrate them into a dependable offering that customers will pay a premium to use.
6. Financing growth efficiently. Customer prepayments and equipment financing help fund GPUs. The scale of future capex means debt cost and potential dilution still deserve close attention.
7. Contracting future capacity across more customers. The next stage depends on matching 2027 and 2028 deployments with demand on attractive terms, while avoiding excessive dependence on any one customer or contract type.
My thesis remains that power and execution are the scarce assets in AI infrastructure. IREN has assembled an unusual combination of sites, construction capability, customers, GPU financing, and Mirantis’s software layer.
Now the company has to show the full system working together. I will judge that through operating ARR, on-time customer handovers, cluster reliability, renewals, and funding cost. Kent’s response to SemiAnalysis is part of the explanation; performance over the coming quarters will be the answer.
My analysis only, not investment advice. Future ARR and delivery dates are forward-looking. Do your own research. @brianfry01 @jiahanjimliu @ilzmcfly @OInvests @TheTechInvest @BitcoinAIGuy @franklee6924T @IREN_Ltd > 引用 @cazenove_uk: Great to welcome @kentpdraper, Chief Commercial Officer of @IREN_Ltd on the Power Analysis podcast to discuss highlights from their 2026 FY and provide the outlook for 2027 and 2028, plus much more!!! ➡️ https://t.co/BMtTwGKMX1 https://x.com/StockAnalystPro/status/2103269310047367458
## @moninvestor (mon) · 09-24 20:50 · ♥30 ↻0 💬4 When it comes to the AI infrastructure sector, there are so many different ways you can play this.
You have the data center builders, the neoclouds selling compute, power companies, construction companies like CAT, cooling, electrical equipment, GPUs, CPUs, memory, networking, fiber and photonics. Then you have commodities like copper, uranium and rare earth materials. They are all benefiting in different ways from the same buildout.
This is why I think investors need to look at AI infrastructure as a much bigger sector than just data centers or semiconductors. If we are going to build significantly more compute over the next five to ten years, we are going to need significantly more of everything.
What I find crazy is that this has already been going on for a few years, yet it still feels like we're just getting started because the demand keeps going up. We are only now starting to see agentic AI scale, while robotics, autonomous systems, wearables and other areas could add another huge layer of compute demand over the next few years.
Then look at what the hyperscalers are doing. Microsoft, Google, Amazon and Meta are spending enormous amounts of money on AI infrastructure, and I think Meta's CapEx will continue going higher. The amount of capital going into this sector is something we have never really seen before in technology. https://x.com/moninvestor/status/2103225630183829601