Anthropic Adds IREN to a Cloud Portfolio Now Worth $53 Billion
Anthropic has signed a multi-year cloud contract with IREN, an Australian-listed supplier of computing capacity, putting one more name on a list of infrastructure commitments that analysts now value at roughly $530 billion.
The Information reported the agreement on October 11. Neither company disclosed the size of the contract or how long it runs. Shares of IREN rose more than 5% in premarket trading on the news.
Taken alone, the deal reads as a footnote. Set beside Anthropic’s agreements with AWS, Google, Lambda and TeraWulf, it completes a year of commitments to data center leases that analysts estimate at about $530 billion.
The capacity numbers move in the same direction. Anthropic’s footprint stood at 1.4 gigawatts last year. It has grown to 5 gigawatts this year, and the plan points to 9.5 gigawatts next year.
Reuters and Korean media outlets reported the same day that Anthropic also intends to lease facilities near Baton Rouge, Louisiana, running Google’s AI server chips from the second quarter of 2027.
What the company is buying is not equipment so much as time. A frontier model is trained on tens of thousands of chips running for months, and the labs that win a given release cycle are usually the ones that secured the buildings and the electricity before their rivals did. A lease signed in 2026 decides which model can be trained in 2028.
The mix matters as much as the total. Anthropic has spread its commitments across hyperscale providers and smaller suppliers of rented compute, a strategy that keeps it from leaning on a single vendor and leaves room to shift workloads as chips change.
Renting also keeps capital spending off its own balance sheet while the build-out is under way. The trade is that a lease is a fixed obligation, payable whether or not the model it serves earns its keep.
That distinction has started to draw attention. Commitments of this size are increasingly funded by debt and by contracts that stretch years into the future, and the companies signing them are betting that demand for AI keeps compounding at something close to the pace of the past three years.
The pattern runs across the industry. Labs rent, cloud providers build, and the two sides sign agreements that tie them together for a decade, a structure that spreads risk in theory and concentrates it in practice, because a slowdown in one corner shows up in the other.
For Anthropic itself, the run of deals reflects a gap between ambition and cash. The company sells access to its models to businesses and developers, but training the next generation costs far more than the current one, and the difference is covered by investors and by contracts rather than by revenue.
The IREN contract does not answer the question investors keep asking: how much capacity Anthropic actually needs, and when it will be used. Companies rarely publish utilization rates, and there is no independent check on whether the 9.5 gigawatts planned for next year matches work already booked or work the company hopes to win.
There is a version of this arrangement that is prudent and a version that is speculative, and the two look identical from outside. A lab that signs capacity just ahead of the demand it can forecast is managing a supply chain. A lab that signs capacity against demand it hopes to create is financing a bet with other people’s buildings. The difference shows up only when the models ship and someone measures what they earn.
The suppliers have their own reasons to welcome the contracts. A signed lease underwriting a new data center helps a developer raise construction debt, which is why compute providers have been willing to lock in long terms at prices large customers find attractive. Each deal makes the next one easier to finance, and each one adds to the total that analysts tally.
What is clear is the shape of the strategy. Anthropic is assembling power, buildings and chips years ahead of demand, and accepting obligations that will outlast any single model it ships. The logic is that capacity, once secured, is an advantage no competitor can rent away at short notice.
The $530 billion figure deserves care. It was assembled by analysts from disclosed agreements and reported negotiations, not filed in one place by the company, and estimates of this kind have a way of growing as more leases are signed and more are described in public.
The bet underneath the number rests on one assumption: that appetite for frontier AI keeps rising fast enough to fill the capacity now under contract. If it does, the leases will look cheap in hindsight and the companies that signed them first will look prescient. If it does not, the obligations stay on the books while the revenue that was supposed to cover them does not arrive.
There is a second, subtler risk. A lab that has committed to gigawatts of capacity has also committed to using it, and using it means training and running models at a scale that demands continual improvement to justify the bill. Contracts do not create demand. They only make the cost of being wrong larger and more visible.
Neither IREN nor Anthropic would say what the contract is worth or how long it lasts. IREN’s shares moved on the news anyway, which is the market’s answer to the part of the deal it can see.
The company has now signed enough of these agreements that the individual announcements blur. The aggregate is the story, and the aggregate says Anthropic expects to be a far larger consumer of electricity, buildings and silicon in 2027 than it is today. Everything else in the strategy follows from that expectation being right.
