IREN and BE Networks plan AI factory with 50,000 Nvidia GPUs

Scaling to tens of thousands of state-of-the-art GPUs is precisely what separates hobbyist AI from frontier labs. IREN’s planned Blackwell Ultra deployment, and the decision to build a high-fidelity digital twin with Nvidia DSX Air before racking hardware, shows how quickly AI “factories” are professionalizing. Network simulation at this scale is about more than uptime—it’s about guaranteeing that multi‑billion‑parameter training runs don’t stall because of subtle bottlenecks.

For the race to AGI, the interesting piece is not that another data center is buying more GPUs, but that design and validation are becoming software-defined. Using DSX Air to pre‑model the entire environment lets operators push utilization, experiment with new topologies, and rehearse failures in silico. That shortens iteration cycles for both infrastructure and model developers, compressing the time between generations of frontier systems.

It also underscores the growing dependence of independent AI cloud players on Nvidia’s full-stack ecosystem, from chips to simulation. If DSX Air emerges as the de facto planning environment, Nvidia’s influence will extend well beyond silicon, giving it significant leverage in how and where large-scale compute for advanced models is built.