A few people asked over the weekend what the calls to slow the pace of frontier AI mean for the buildout. Some thoughts: Even if models never improved from here, just rolling out what they can already do would take more compute than the world can build for years. The debate about how fast AI should be allowed to improve is a fair one to have. It's about future generations of models. Existing demand is the part I think people are misreading. Anthropic CEO Dario Amodei said in May they'd planned for 10x growth and were running at an 80x pace in the first quarter, and that's why they've had trouble supplying compute. OpenAI president Greg Brockman said in July they'll be in a compute shortage no matter what, and are choosing which products to scale. Google says it's processing 7x the tokens it did a year ago. Hundreds of millions of people use these models today, and most of them use a small part of what the models can already do. And every time more compute comes online, usage steps up again: limits come off for people already using it, customers who were turned away get on, and new use cases show up that weren't in anyone's plan. Now supply. The constraint is HBM, the memory that sits inside every major AI chip. Three companies make it and all three are sold out this year. A new memory plant takes years to build. TrendForce has HBM shipments growing 50-60% next year. NVIDIA, the biggest buyer of it, expects its revenue to grow about 70% next year and calls that outlook 'supply-constrained', noting its customers' forecasts point closer to 100%. On our own bottom-up work, the memory constraint lands in about the same place as NVIDIA's growth number. Then the chips need a building with power connected, which takes longer again. Goldmans reckons only about half the US capacity scheduled over the next two years will actually be built on time. The risk to demand continues to seem heavily weighted to the upside. The risk to supply continues to seem weighted toward less capacity getting built, not more.

A few people have Thursday's numbers tangled, so let me untangle them. The majority of the $684m loss is the cost of retiring Bitcoin miners as we convert those sites to AI Cloud. Non-cash. The cloud business underneath ran ~87% gross margins (ex D&A). Every megawatt that comes off mining goes back on at multiples of the revenue. Recent 3-year AI Cloud contracts are at >$20m per MW (IT), more than double late last year. And the $25-30bn of forecasted FY27 capex isn't an equity number. Customer prepayments can cover about half the GPU capex. Lenders can fund most of the rest. We raised ~$19bn over the last twelve months and only ~$3bn of it was equity. We haven't borrowed a dollar against the data centers yet either. $4bn of ARR is contracted, with three sites commissioning between now and year end to bring it online. And that's the 2026 story. The real ramp is 2027, when Sweetwater and the next wave of capacity come into play. It's delivery time.

Every new AI data center creates more demand for AI, not less. That's what people referring to past booms keep getting wrong about this one. Better AI needs more computing power to run. Cheaper AI gets used more, not less. And every step change in capability creates use cases that didn't exist a year ago. Demand grows at the speed of software, but supply cannot. Supply grows at the speed of the real world: power connections, transmission lines, concrete, steel, and the thousands of people it takes to build. In past booms, technology helped supply catch up. This time, technology is on the other side of the equation. It's fueling exponential demand, and the real world can't build fast enough to keep up.

Nine months ago we signed a contract. Today Microsoft has accepted delivery of Horizon 1: 50MW of direct-to-chip liquid cooled AI infrastructure, built, commissioned and handed over. A greenfield build of this scale typically takes 2-3 years. Owning the land, the power, the substations and every workstream in between is how you compress that. Proud of the 3,000+ people on site who make it happen. Three more Horizons to go this year.

Texas is asking the right questions about data center growth: how to ensure Texans reap the benefits while protecting their natural resources. At IREN, that has been our approach from day one. Read more from @danroberts0101:

Data centers. Compute. Software. Under one roof. Three layers. One compounding advantage. Yesterday, we closed the acquisition of Mirantis, strengthening layer three with a team that has spent over a decade building and operating cloud infrastructure for 1,500+ enterprises globally.

8 years ago Will and I started accumulating powered land because we thought compute would eat the world. Some weeks the market agrees with us more than others. What we know today: demand for our capacity exceeds everything we can build, 85% of our $4bn+ 2026 target is signed, and there are thousands of people on our sites right now pouring concrete and racking GPUs. We've been through way worse than this. Back to it.

One of the most common questions we get about AI data centers is how much water do they use? Fair question. Older designs cooled servers by evaporating water and people naturally assume ours do too. They don't. $IREN's sites are different. Direct-to-chip liquid cooling in a closed loop. Same idea as a car radiator or a fridge. Fill it once and the water just goes round and round. Over the life of a 200MW site, the initial fill plus every top-up averages out to about as much water each year as 3 households. Not a town. Not a suburb. Not a street... 3 houses. Water matters in the regions we build in. Which is precisely why our cooling barely uses any.

12 months ago we had ~3MW of self-built AI Cloud capacity. Today: 480MW being delivered this year, $2.8bn in new contracts signed, and our 2026 ARR target raised to $4bn+ with ~85% already under contract. Demand continues to exceed everything we can build. Recent contracts include customer prepayments covering ~45% of the associated GPU capex, with weighted average contract terms of ~4 years across the portfolio. Data centers. Compute. Software. The three-layer thesis, executing as written: We're in a good spot. Proud of the team.