MY TOP 3 AI UTILITY NAMES AI today hits two hard limits power & space. There are plenty of $NVDA GPUs but not enough energized capacity to run them. The winners are the ones who turn cheap power into reliable compute that stays full & running. Heres my top 3 favorite names 🧵

1. $NBIS | Nebius Cloud Utility for the AI Age Nebius is the closest example of a true integrated AI utility. It sells capacity before construction begins which means revenue is locked in long before the concrete is poured. The $17B $MSFT deal confirmed this model works since those multi-year contracts secure cash flow and reduce financing risk since the utilization is guaranteed. Nebius’ data centers are built ~20% more power-efficient than typical cloud sites which means every GPU-hour they sell costs less to produce. In an industry where power costs can exceed 40% of total operating expense, this delta compounds quickly and allows Nebius to price below $AMZN AWS or Azure while still maintaining stronger margins. What makes Nebius distinct from other neocloud providers is its software layer. It offers integrated developer tooling (ClickHouse for data handling, MLOps pipelines & AI-native infra management) bundled into the same environment as compute. This makes its platform “stickier. Once developers deploy workloads on Nebius, switching costs rise because both storage and training pipelines are optimized for Nebius’ internal systems. The result is a full-stack flywheel: cheap energy, efficient hardware utilization & embedded software tools all reinforce each other. It earns like an energy company but grows like a software platform.

2. $IREN | IREN Limited Energy-to-Compute Engine of the AI Era IREN is structurally the most vertically integrated company in the entire AI infrastructure segment since controls the full value chain from power generation to compute delivery. Its renewable energy costs average ~3.5 cents per kWh (among the lowest in North America) and it owns the physical land, substations & cooling systems needed to bring additional capacity online quickly. What differentiates IREN is its ability to convert that cheap power into rentable GPU capacity faster than most peers. The company already operates its own AI cloud with several thousand active $NVDA GPUs and expansion plans call for tens of thousands more across Canada & Texas. These facilities can be commissioned within quarters because permitting, power interconnection & infrastructure groundwork are already in place. The internal economics is what attractive me to it since IREN owns both the energy and the shells, which causes no pass-through charges & gross margins on GPU compute can exceed 70%. Management has hinted at a roadmap to 60K GPUs by 2026 & ~$2.5B in cloud service provider rev once utilization ramps. When you own the power, the land & the hardware like IREN does then you capture the full margin instead of splitting it across vendors.

3. $CIFR | Cipher Mining Landlord of the AI Buildout Cipher’s plays a more traditional but still relevant role in the AI Utility theme. It acts primarily as an infrastructure landlord by leasing powered capacity to hyperscale tenants. The Barber Lake campus (designed for extreme rack density & multi-gigawatt scalability) is already validated by a multi-billion-dollar, decade-long lease with $GOOGL. That deal locks in steady revenue for Cipher and takes most of the near-term risk off the table since utilization stays high and counterparty risk is low. The trade-off is a bit structural since Cipher gets fixed rent instead of the upside from rising GPU-hour prices. As long as GPUs are scarce tenants will pay up for capacity but once supply catches up or hyperscalers build more in-house those margins tighten fast. Unlike Nebius or IREN, Cipher doesn’t own the compute layer since it owns the land, power & shells. Growth depends on raising capital through notes or share offerings to build more sites. The model looks more like an industrial REIT with steady cash flow but limited upside. In a tight capacity market that works well but once supply opens up it turns defensive instead of growth driven.

All three are solving the same problem that there isn't enough powered & GPU-ready space. • $NBIS builds across the software & developer layer • $IREN builds down into power & compute • $CIFR rents out the physical layer through long leases AI infrastructure is moving from buying more GPUs to energizing the ones we already have with power, interconnects & cooling as the new choke points.

Less than an hour until $IREN reports earnings. Here are the estimates: