New York, May 26, 2026, 07:01 (EDT)
IREN Ltd. co-founder and co-CEO Daniel Roberts is sounding a note of caution about the AI infrastructure race, saying that it could take until 2030 for a 1-gigawatt “AI factory” — the big data centers that power AI models — started today to go live. IREN shares were quoted at $56.83 before the U.S. open, putting the Nasdaq-listed firm’s market cap around $19 billion. Benzinga
That’s important now since IREN isn’t just seen as an ex-bitcoin miner with extra energy. Nvidia and IREN said this month they’re teaming up on as much as 5 gigawatts of Nvidia-linked AI infrastructure. Nvidia picked up a five-year option to buy up to 30 million IREN shares at $70 each—worth up to $2.1 billion if they use it.
Nvidia posted record quarterly revenue of $81.6 billion last week, driven by $75.2 billion from its data center unit. At the same time, the International Energy Agency predicts data centers will need over twice as much power by 2030, rising to 945 terawatt-hours. The importance of chips is clear. Power to run them looks just as important.
IREN has a separate five-year deal with Nvidia worth $3.4 billion, where it’s set to run managed GPU cloud services for Nvidia’s own AI and research work. The setup will use Nvidia’s Blackwell AI chips at about 60 megawatts of IREN’s existing data center space in Childress, Texas. Roberts said the agreement proves IREN can offer “fully managed cloud solutions, not just bare metal,” so it’s more than just leasing out servers. IREN
Nvidia CEO Jensen Huang said, “AI factories are becoming foundational infrastructure for the global economy.” Roberts said the deal puts Nvidia’s AI systems together with IREN’s power, land, data-center, and infrastructure business. NVIDIA Newsroom
Roberts gave a more serious warning in comments to 24/7 Wall St.: “If you wanted to start today and build a gigawatt AI factory, you are looking 2030 before you get the first compute online.” He wasn’t talking about GPUs getting easier to buy. The holdup is land, substations, permitting for transmission, and actually getting things built. 24/7 Wall St.
Roberts said much the same in a Friday post, CoinDesk reported. “AI demand grows exponentially. Infrastructure doesn’t,” he wrote. He broke down IREN’s model into three parts: physical infrastructure, compute infrastructure, and software tools sold to enterprise customers. He said today’s value is mostly in the first two layers. CoinDesk
IREN’s shift is proving costly. Third-quarter revenue dropped to $144.8 million, down from $184.7 million last quarter, and IREN posted a net loss of $247.8 million. The company blamed the revenue drop on weaker bitcoin mining and taking mining rigs offline before installing new GPU equipment and billing, though AI cloud sales partly made up for the hit.
IREN has brought in new money. The company finished a $3.0 billion convertible notes sale this month and took in around $2.96 billion after costs. Convertible notes are debt that could turn into stock. Reuters said companies tied to AI have fueled a wave of these deals, and Michael Youngworth at Bank of America Securities told Reuters that a lot of the funds are being raised “to build out capital expenditure, particularly AI.”
The contest for AI cloud business is opening up. Reuters called IREN a “neocloud,” saying the company is selling Nvidia-based compute, in the same group as CoreWeave and Nebius. Microsoft agreed to a $9.7 billion cloud deal with IREN last year, and also signed a $17.4 billion infrastructure deal with Nebius. The moves show big tech customers are splitting their demand among new AI cloud players. Reuters
Execution is the main risk. IREN said the expected annualized revenue target is not guaranteed. The company said GPUs have to be delivered and commissioned, and it pointed to possible problems with power supply, grid connections, permitting, and customer performance all diverging from plan. IREN also flagged Texas ERCOT large-load interconnection rules as another issue that could cut or delay available power.
IREN is under pressure to show it can convert megawatts to billable GPU time, not just sign more AI deals. The company said its 2026 goal for 480 megawatts is still in place, and 1.2 gigawatts of AI cloud capacity are being built for 2027.
