IREN Limited (NASDAQ: IREN) is leveling up its AI cloud ambitions with fresh firepower from NVIDIA. The data center operator announced it has secured NVIDIA Preferred Partner status—a badge that opens the door to deeper collaboration—and added another 2,400 next-gen GPUs to its fleet, split between 1,200 air-cooled B300s and 1,200 liquid-cooled GB300s.

The deal, worth roughly $168 million, pushes IREN’s total GPU count to 10,900 NVIDIA units, a scale that puts it in striking distance of more established AI cloud players.

Liquid Cooling Meets Grid Power

The GPUs will be housed at IREN’s Prince George campus in Canada, where construction is underway on a 10MW liquid-cooled installation designed to handle more than 4,500 GB300s. That mix of air- and liquid-cooled hardware highlights a growing trend in AI infrastructure: traditional air cooling is hitting limits as power-hungry chips like the GB300 push thermal boundaries.

By integrating both cooling methods under one roof, IREN is effectively future-proofing its data center against the escalating demands of AI workloads. The company also claims nearly 3GW of grid-connected power, a figure that dwarfs many rivals and hints at room for massive GPU scaling.

Financing the GPU Arms Race

Of the $168 million bill, $96 million is covered through a 24-month lease with a high single-digit interest rate, while the rest comes from IREN’s existing cash reserves. That financing model—leasing GPUs rather than outright purchasing—mirrors what other cloud upstarts are doing to avoid diluting equity while still competing in a market dominated by hyperscalers like AWS, Microsoft, and Google Cloud.

The math matters: NVIDIA’s latest GPUs are not just expensive; they’re scarce. With cloud demand for AI training surging, being able to secure thousands of GB300s is as much a supply-chain victory as a financial one.

What IREN’s GPU Stack Looks Like Now

Once the new hardware is online, IREN’s AI cloud will pack:

• 800 NVIDIA H100s
• 1,100 NVIDIA H200s
• 5,400 NVIDIA B200s
• 2,400 NVIDIA B300s
• 1,200 NVIDIA GB300s

That’s a mixed fleet but heavily tilted toward NVIDIA’s latest B-series and GB-series accelerators—the same chips that are expected to drive AI model training through the second half of the decade.

The Bigger Picture

IREN isn’t alone in stockpiling GPUs. From CoreWeave to Lambda, specialized cloud providers are racing to secure NVIDIA’s bleeding-edge chips, betting that enterprises will prefer nimble, GPU-dense providers over generalist hyperscalers for AI training workloads.

But IREN’s twist is scale and flexibility: few upstarts can claim both liquid-cooled and air-cooled infrastructure at industrial scale. If the company executes, it could emerge as one of the largest independent AI cloud operators outside of the hyperscale triopoly.

For now, the takeaway is clear: the GPU land-grab is intensifying, and IREN just carved itself a bigger slice.