You mentioned on your earnings call that hyperscaler CapEx had jumped from $400 billion to $700 billion in just a few months. What's driving that acceleration, and how much runway does it still have?
The jump largely reflects hyperscalers resetting their capital expenditure projections for 2026, and across the board those numbers went up significantly. But what we see in many facets of our business is that the world is simply short on compute. When I think about AI and where we're headed, I think about it becoming the primary means of production for labor across virtually every industry.
What's been remarkable to watch is how quickly the conversation shifted: in late 2024, every interview I did was about the AI bubble. Now the question is which industry AI is going to disrupt next. That's a stunning change in just a few months.
What's even more striking is that enterprise adoption, which historically has always lagged consumer adoption by years, is now happening almost simultaneously. ChatGPT reached a billion daily queries in under two and a half years; Google took eleven. And now enterprises are piling in as well, at a pace we've simply never seen. The hyperscalers see this, they're all short on compute, and they're racing to add capacity. The honest answer is it's very hard to see where this slows down, because we're still in the genuinely early stages of adoption.
When you think about the binding constraints on AI infrastructure, is it power generation, grid interconnection, or something else entirely? And is US policy moving fast enough to address it?
It's a combination of all of those factors. Interconnection queue times are extraordinarily high and the process remains difficult, so regulatory reform there would help. But even if you wanted to order traditional gas turbines today to generate power, you'd be looking at delivery in 2031 or 2032. That's simply what the supply chain looks like right now. We're fortunate to have locked up a significant amount of power for our campuses, and we've made forward-looking moves to begin building our own generation capacity coming online in 2028, once we've exhausted the available grid power at our sites.
The constraints are real and, importantly, they're getting harder, not easier. Permitting is more difficult, finding power is more difficult, the supply chain for key components is stretched across every dimension. I think even the hyperscalers will struggle to meet their own build-out budget goals given these limitations, while demand continues to accelerate.It's a genuinely unusual position to be in: constrained supply meeting what looks like unbounded demand.
You created Base Electron to own your power generation rather than simply signing long-term supply agreements. What was the rationale, and does it eventually sell power to third parties beyond your own campuses?
Base Electron was created because we identified power as a primary constraint and wanted more direct control over how generation capacity gets built. We partnered with Babcock and Wilcox, a company with over 160 years of experience building boilers, which has done many retrofits of coal infrastructure to natural gas. We're deploying steam turbines on that gas, building the steam turbine portion of a combined cycle plant first and getting delivery in 2028, with the traditional gas turbine to follow. That timeline fits our build-out perfectly, and it's something we simply couldn't have achieved through conventional procurement.
Crucially, we're not doing behind-the-meter generation, where you just build power right next to your own data center. We're siting these plants on the grid itself, working directly with utilities in North Dakota, so that the infrastructure benefits all ratepayers across the broader grid, not just our campuses. We believe that's a more durable and genuinely collaborative model. It strengthens the overall grid ecosystem in the regions where we operate, and it positions Base Electron as something much larger than a captive energy source for Applied Digital.
You were early, building on spec in 2023 before you had a single customer. What did your outsider perspective let you see that incumbents in the data center world had missed?
The insight came from being on the front lines at exactly the right moment. We had already built GPU infrastructure for AI labs, and when I went out looking for additional data center capacity, I sat with the CEOs of some of the largest private data center companies in the country. Every single one of them asked me some version of the same question: what the hell is going on? They were seeing demand for high-power-density capacity at a rate they had never witnessed, but they didn't have the context to understand where it was coming from. I did, because I was talking directly to the frontier model companies creating that demand.
The other piece was a conviction I developed from years as a semiconductor investor: the GPU shortage would get fixed. TSMC and the broader semiconductor supply chain had solved problems like this before and would again. But building data centers, getting permits, securing power, working through supply chains for electrical distribution and cooling, that takes years, not months. So while everyone else was focused on chips, I looked at the map and saw that data center capacity was going to be the real, durable bottleneck. We started building for hundreds of megawatts based on that thesis, 19 months before we signed our first customer. It was stressful, but the logic was clear.
Looking ten years out, is Applied Digital still primarily a landlord for hyperscalers, or does owning the land, power, and cooling infrastructure eventually pull you closer to the compute layer itself?
We've been very deliberate about separating these businesses precisely because they carry very different risk profiles. The data center business, Applied Digital at its core, is about long-term leases, 15-year contracts on buildings we consider 30 to 50-year assets. Investors in Applied Digital should be able to model predictable, stable cash flows from high-quality hyperscaler offtake. We've said we expect to exceed a billion dollars in net operating income over the next five years, and our executive incentive structure is built around reaching two billion. That's the clarity we want to offer.
The compute layer is a different proposition: shorter cycles, real technology obsolescence risk, and a business that would partly compete with the very customers we're signing data center leases with. That's why we're spinning it out as a separate public company, Chronoscale, in the coming weeks. And Base Electron, the power generation business, has its own capital structure and risk profile as well. The goal is for investors to have a genuine choice: Applied Digital for long-term infrastructure yield, Chronoscale for compute-layer upside, Base Electron for the power generation opportunity. Each piece is clean and understandable on its own terms, rather than obscured inside a complicated vertical stack.
