IREN provides GPU cloud infrastructure for AI training and inference, supported by data centers it develops and operates. Its proposition is the connection between physical capacity and computing service: land, power, construction, GPU clusters and operations are coordinated within one business. A buyer still needs to establish the exact cluster, location and delivery commitment available for its workload.
- 01The offer Dedicated GPU infrastructure supports AI training and inference.
- 02The reader Model developers and enterprises procuring substantial dedicated AI compute.
- 03The boundary A future campus or contracted allocation is not the same as capacity available to a new buyer.
01 / ProductAI compute connected to the infrastructure beneath it
The AI Cloud page lists NVIDIA GPU configurations and directs prospective customers to sales. It describes bare-metal access, customer frameworks and containers, private networking and storage options. This is infrastructure for running a customer’s models and software, rather than a ready-made model API with one public token tariff.
IREN’s vertical-integration explanation describes internal responsibility across site development, facility construction and technical operations. That can provide a more direct route for resolving problems that span the building and computing stack. The company’s claimed efficiency benefits remain vendor claims; they are not measured cost or reliability results from this article.
The FY2026 annual filing records the August 2026 acquisition of Mirantis, adding cloud infrastructure software, orchestration and enterprise support capabilities. That ownership change should not be mistaken for proof that every acquired feature is already part of every IREN customer agreement. Confirm the software and support scope actually being supplied.
02 / AudienceFor teams procuring a meaningful block of AI capacity
IREN is relevant to model developers and enterprises that need dedicated GPU infrastructure for sustained training or inference. A useful candidate has an identifiable workload, technical operators and a reason to control the software environment. Procurement can then focus on the cluster topology, storage, service support and delivery schedule that make the workload usable.
It is a less natural starting point for a small application needing a few sporadic completions. A managed model API can remove much of the operating work. A GPU cluster requires the customer or a contracted partner to handle model software, workload scheduling, observability and recovery unless those services are explicitly included.
The CoreWeave blueprint offers a comparison in specialized GPU cloud infrastructure. The Nebius blueprint is another useful read for AI compute and platform services. Compare the contracted operating layer and exact configuration rather than assuming that the same accelerator family creates an equivalent service.
03 / WorkflowA proposed training run with an explicit delivery plan
Imagine a company adapting a multimodal model for inspection images from its manufacturing process. This is a proposed workflow, not a performed IREN benchmark. Begin with the training objective, licensed dataset and a reproducible software environment. Establish how much work must finish by the business deadline before deciding how much capacity to reserve.
Prepare a representative job using the actual image sizes, storage layout and training framework. Identify the expected communication between nodes and the rate at which data must reach the GPUs. The purpose is to give the provider a concrete workload specification, so the proposal can address the surrounding network and storage rather than only accelerator quantity.
Ask IREN to identify the offered cluster’s location, GPU configuration, network topology and storage arrangement. Distinguish local scratch storage from durable data that must survive a job or node failure. Confirm the supported container and driver environment, access method and how responsibility is divided when a problem crosses software and hardware boundaries.
Stage the dataset and verify its integrity before starting the expensive run. Keep model checkpoints in a durable location with a documented recovery procedure. A training process that can resume from a recent checkpoint may tolerate a different interruption pattern from an interactive inference service. Match the support agreement to the actual workload instead of using one generic availability requirement.
The GPU buyer’s guide landing page highlights networking, service levels and data-transfer economics. Use those topics to shape a scoped proposal: specify the data path, required support response and transfer charges before comparing cluster offers. The AI Cloud product page uses contact-sales pricing, so obtain the tariff for the exact configuration and commitment being considered.
Plan acceptance before the capacity arrives. Agree how the customer will verify the contracted hardware, storage access, network behavior and basic job recovery, within the permitted service terms. Keep a record of the accepted configuration and distinguish provisioning acceptance from model-quality acceptance. A healthy cluster can still train a model that fails the business task.
Finally, define the exit path. Record how datasets, checkpoints and logs will be exported and which resources must be removed at the end of the agreement. A temporary training project needs a clean end state, while a continuing inference service needs a renewal or migration plan. Both benefit from deciding those responsibilities before the first large transfer.
04 / PricingA cluster agreement replaces a universal public hourly price
| Item | Published route | Specify in the agreement |
|---|---|---|
| H100 / H200 clusters | Contact sales | GPU count, location, term and allocation |
| B200 / B300 / GB300 NVL72 | Contact sales | Exact system, availability and delivery milestones |
| Storage and networking | Configuration discussion | Durability, performance, transfer and support boundaries |
| Managed versus bare-metal scope | Contracted service arrangement | Customer software duties, support and acceptance criteria |
Commercial scope checked 23 September 2026 on IREN AI Cloud and the GPU buyer’s guide landing page. Listed GPU configurations use contact-sales pricing; no universal public hourly tariff is established here.
The reviewed AI Cloud page presents contact-sales routes for its named GPU configurations. The commercial table therefore identifies the scope to establish rather than inventing a market-wide rate. A quote should state the compute allocation, term, location, storage, network and support commitments, plus when charging begins.
Large announced contracts are evidence of commercial activity, but their total values are not public list prices for another customer. Their duration, equipment, delivery milestones and service scope can differ substantially. Dividing an announced contract value by an assumed GPU count would produce a misleading comparison, so this blueprint does not do that.
For the inspection-model project, compare the total cost of completing the accepted training objective. Include data preparation, idle time while diagnosing software problems, checkpoint storage and recovery. A cluster that appears cheaper per hour can cost more per completed run if the application cannot use its available resources efficiently.
Capacity timing is also commercial. Ask whether the proposal covers installed equipment, an allocation from an operating pool or a future delivery. Those arrangements carry different dependencies. The contract should connect payment, acceptance and remedies to the actual promised capacity, especially where a business deadline depends on the start date.
05 / DistinctionsFacility ownership makes the capacity story more concrete
IREN’s Childress facility page describes a Texas site with air and liquid cooling, diverse fiber paths and company-owned substations. Those are physical characteristics relevant to a large computing service. They do not establish how much unallocated GPU capacity is available there, or whether that site is the one proposed to a new customer.
The owner-operator model can make it easier to discuss dependencies across power, cooling and cluster operation with a single organization. The practical advantage depends on the agreement: identify the team responsible for each layer and the escalation path between them. Vertical integration does not remove the need for transparent incident communication or customer-side workload expertise.
The current AI Cloud page also supports bringing frameworks and containers to bare-metal infrastructure. That is useful for teams with an established training environment or specialized serving software. It shifts attention toward compatibility, reproducibility and operational support, rather than just choosing a model from a hosted catalog.
06 / QuestionsHow much of the advertised capacity is available for this project?
The annual filing provides a useful distinction between delivery and plans. It says Microsoft accepted Horizon 1 in August 2026, while later Horizon tranches had future targeted delivery periods. That example illustrates why an announced campus, contracted capacity and an accepted operating service should be treated as different states. It does not imply that Microsoft’s capacity is available to other customers.
The public location navigation likewise distinguishes sites under construction or in development from operating locations. A buyer should request a current allocation and delivery statement for its own cluster. Do not infer immediate availability from the size of the company’s power portfolio or from a photograph of a facility.
What is included in isolation and support? The product page describes private networking, tenant separation and in-house support, but a sensitive workload needs the relevant contract, access controls and data-handling commitments. Confirm where administrators, backups and telemetry operate. The published overview is not an independent certification of a particular customer configuration.
How will software capabilities evolve after acquisition? Mirantis can add relevant orchestration and enterprise expertise, but integration takes specific product and service work. Ask which capabilities are supported now and which remain planned. Keep future capabilities out of the minimum requirements for the initial workload unless they are explicitly committed in the agreement.
07 / DecisionChoose IREN for a specified cluster and operating responsibility
IREN is worth considering when a team needs substantial AI compute and values a provider that also controls the facility layer. Start with a workload specification and request a concrete configuration with a delivery and acceptance plan. Judge the offered service by those commitments rather than by the company’s total announced development pipeline.
For the inspection-model project, useful success is a completed, reproducible training run and a portable accepted artifact. The GPU infrastructure makes that possible, while dataset quality, training software and model evaluation remain essential customer responsibilities. A clear division of work gives both sides a better basis for delivery.
Procure sustained AI capacity
Define the workload, topology and required start date before requesting a cluster proposal.
Expand an existing model platform
Verify the offered allocation and migration plan independently of future campus capacity.
Start with hosted inference
Use a managed model API when dedicated infrastructure would add unnecessary operating work.
A business worth understanding.
Suggest your business or one you find interesting. Tell us what you want to understand about its product, positioning, design or workflows.
Suggestions are free. Selection and publication stay with the desk.
- IREN AI CloudConsulted
- vertical-integration explanationConsulted
- FY2026 annual filingConsulted
- GPU buyer’s guide landing pageConsulted
- Childress facility pageConsulted
