IREN Achieves NVIDIA Exemplar Cloud Status on NVIDIA GB300 NVL72, Confirming Scalable AI Training Across its Production Fleet

IREN has achieved NVIDIA Exemplar Cloud status on NVIDIA GB300 NVL72 connected with NVIDIA Quantum-X800 InfiniBand networking for training workloads. To achieve this status, a provider must perform within NVIDIA's reference performance targets across every prescribed recipe for the hardware generation being validated.

This milestone arrives on the heels of IREN’s recent Exemplar Cloud achievement on NVIDIA HGX B300. Reaching Exemplar status on a mature platform is meaningful. Doing it again, back-to-back, on the NVIDIA GB300 NVL72 generation demonstrates repeatable execution across successive hardware architectures at production scale.

IREN achieved Exemplar Cloud status through close collaboration with NVIDIA’s engineering team. The process identified areas to optimize efficiency and performance, allowing IREN to implement the latest best practices to ensure that GPU, CPU, networking, and other configurations are optimized for modern AI and HPC workloads. 

IREN’s Exemplar Cloud status was achieved following NVIDIA’s testing of its NVIDIA GB300 NVL72 deployment at Horizon 1, which was delivered to Microsoft in August 2026. Horizon 1 is the first of four 50MW (IT load) direct-to-chip liquid cooled AI Cloud deployments scheduled for delivery to Microsoft at IREN’s Childress, Texas campus in 2026 under a five-year, $9.7bn cloud services contract.

What NVIDIA Exemplar Cloud Verifies

The NVIDIA Exemplar Cloud program narrows the gap between AI performance claims and real world results.

The NVIDIA Exemplar Cloud initiative was created to bring transparency and rigor to a market long governed by spec sheets and theoretical peak numbers. Rather than testing components in isolation, it evaluates providers against real, end-to-end AI training workloads using open source NVIDIA Performance Benchmarking recipes which run across multi-node accelerated computing clusters under conditions that reflect genuine production pressure.

To achieve the status, a provider must perform within NVIDIA reference performance targets across every prescribed recipe for the hardware generation being validated. It requires clearing benchmark thresholds on throughput, networking efficiency, and cluster reliability under sustained load. A cluster that looks fast in a demo does not always hold up at scale. Networking bottlenecks, suboptimal software configuration, and inconsistent behavior under load are precisely the variables that erode real training runs and precisely what this validation is designed to surface.

For teams choosing where to train frontier and foundation models, NVIDIA Exemplar Cloud is an independently verified signal that the infrastructure will behave the way NVIDIA’s own engineers expect it to run.

Why NVIDIA GB300 NVL72 Raises the Stakes

NVIDIA GB300 NVL72 fuses 72 NVIDIA Blackwell Ultra GPUs into a single liquid-cooled, rack-scale domain connected by fifth-generation NVIDIA NVLink™ scale-up, turning the entire rack into a single accelerator. On IREN AI Cloud, each NVIDIA GB300 NVL72 instance exposes roughly 20TB of HBM memory, 37 TB of fast system memory, 2,592 vCPUs, and 276 TB of NVMe, backed by 3.2 TB/s of non-blocking NVIDIA Quantum-X800 InfiniBand networking for distributed training.

That density is exactly what trillion-parameter architectures need and what makes validation hard. Extracting the platform’s full capability demands infrastructure engineered around its power density, direct-liquid-cooling requirements, and high-bandwidth networking. Expert-parallel groups stay inside the NVLink scale-up domain, collective communication takes the optimal InfiniBand scale-out path, and the whole system sustains that behavior run after run. Passing every recipe on NVIDIA GB300 NVL72 means power, compute, cooling, and fabric continuously work together to NVIDIA standard.

The IREN Difference Offers Vertical Integration You Can Benchmark

IREN provides a vertically integrated platform that secures the power, acquires the land, designs and constructs the data centers, and also operates the cloud platform on top of it. That end-to-end ownership is the seamless engine behind both Exemplar achievements.

Because there is no operational seam between the facility and the workload, IREN runs a continuous feedback loop that many providers structurally cannot. When an issue arises, the teams responsible for the physical plant and cluster operations are both within the same group.  For neoclouds operating in leased colocation or third-party facilities, that loop typically requires coordination across organizational boundaries.; IREN closes it internally. The NVIDIA GB300 NVL72 result is what that integration looks like when it is measured.

Validating NVIDIA HGX™ B300 and then NVIDIA GB300 NVL72 in quick succession demonstrates IREN’s ability to  bring NVIDIA platforms up to certified performance repeatably. This reflects the operational muscle that separates access to GPUs from outcomes on the NVIDIA accelerated computing platform.

What Does IREN Achieving NVIDIA Exemplar Cloud for NVIDIA GB300 NVL72 Mean for AI Builders?

For teams training foundation models, fine-tuning domain-specific systems, building multimodal and agentic applications, running large-scale inference, or accelerating simulation and scientific discovery, the Exemplar Cloud status on NVIDIA GB300 NVL72 is a signal for three practical outcomes.

First, performance you can plan around. The validation is a real-world baseline, so training timelines and total cost of ownership can be modeled with confidence.

Second, infrastructure engineered for NVIDIA Blackwell Ultra from the ground up, with 3.2 TB/s non-blocking NVIDIA Quantum-X800 InfiniBand networking, scalable storage options, and bare-metal with full root access to support major AI frameworks.

Third, direct, in-house support from the team that manages the entire stack, backed by IREN’s status as an NVIDIA Preferred Partner and 24/7 operations.
 

The Trajectory Forward

Exemplar Cloud on GB300 NVL72 is another milestone in IREN’s growth trajectory. With the NVIDIA Vera Rubin platform and beyond, the differentiator will not be who can acquire the newest accelerated computing platform but who can stand it up, validate it, and keep it performing predictably at scale. IREN has now demonstrated that capability twice in short order, on two generations, from the same vertically integrated foundation.

For enterprises and AI natives evaluating where to run their most demanding workloads, the takeaway is simple: for IREN AI Cloud, the performance is independently verified, the stack is owned end-to-end, and the momentum is building toward what AI will demand next.