We’ve been thinking this year about a simple equation: Electrons = Intelligence. Every token generated, every model trained, every inference run ultimately comes down to electrons moving through silicon. No power, no intelligence. The open question isn’t whether intelligence requires electrons, but where those electrons sit. Some inference will remain centralized in large data centers; some will move closer to the edge, onto devices. Either way, the constraint doesn’t disappear—it shifts bottlenecks and changes buildout strategy.
A year ago, we wrote about constraint. This year, our thinking has evolved. After spending time with founders, operators, hyperscalers, utilities, and customers, we’ve come to see energy not as a universal explanation for AI adoption today, but as a defining factor in what actually gets built over the next decade. The technology works. Models can reason, code, see, and increasingly act. Yet adoption and value creation remain uneven, shaped far more by organizational realities than by model capability alone.
Part I: The Constraint
Energy is a local game. It’s won locally—and it’s a hearts and minds game. AI infrastructure doesn’t scale abstractly; it scales through communities, utilities, permitting offices, labor markets, and timelines. Demand is real. Capital is abundant. What’s scarce are electrons that can be delivered on predictable schedules. The scale has shifted dramatically: projects that once seemed enormous are now baseline, and full gigawatts are increasingly table stakes. Execution, not ambition, determines outcomes.
The grid was built for a different era—centralized generation and predictable demand. AI workloads are neither. Training clusters and inference at scale create load profiles the system wasn’t designed to handle, all while electrification accelerates elsewhere. Developers can have land, capital, permits, and customers—and still wait years for power. This doesn’t stop AI progress, but it does shape where it happens and who can participate. Capital is abundant. Execution is not.
Part II: The Adoption
AI outcomes are gated by ownership, incentives, process redesign, integration, and trust—not by model capability. As we enter 2026, adoption is becoming mandatory. Many teams are being asked to deliver the same service levels on materially smaller budgets. That shifts AI from a discretionary pilot to an operational necessity, and this is likely the first year it shows up meaningfully in operating metrics at scale.
Enterprises often frame adoption as a security-versus-utility tradeoff. That’s understandable as a starting point, but the faster learners are allowing broader experimentation, even if it means mistakes along the way. Models improve every six months; organizations that don’t learn in parallel fall behind.
Most companies still spend AI budgets where politics point rather than where returns live. They buy vitamins when they need painkillers. Vitamins are nice-to-have improvements: discretionary spend, unclear ownership, easy to delay. Painkillers address measurable cost or risk, have a named P&L or operational owner, and create a forcing function that makes deployment unavoidable. At Nadia, we only originate companies that are painkillers.
Part III: What’s Coming Next
Nuclear will matter, but not on the timelines headlines suggest. The bridge—natural gas, microgrids, pragmatic generation—matters enormously in the interim. Just as important, we’re running out of people who know how to build this infrastructure. Skilled labor shortages in electrical and construction trades are becoming binding constraints. You can have capital and permits; without crews, nothing happens. We expect consolidation around execution capability.
On the software side, the next phase is about context and memory. AI systems need to know what to remember, when to act, how to be evaluated, and how to operate within clear controls and auditability. Systems of record aren’t disappearing; they’re evolving. The opportunity is in connecting systems, embedding intelligence where work actually happens, and enabling digital labor that does real work.
What We’re Watching Closely
We’re watching robotics and its timing, distressed renewable energy and battery storage, and the emergence of memory as a first-class layer in software—state management, evaluations, and governance. These are all pain points, not features. We don’t view these as predictions. They’re simply areas where we’re trying to stay close to the ground and honest about what’s changing.
Our mission hasn’t changed. We exist to unlock potential in others to make a difference in the world. We build companies from zero to one to scale by identifying constraints and compounding advantages where we can be genuinely helpful. Over time, we’ve found that durable businesses tend to grow out of very ordinary problems that happen to matter a great deal.
None of this work happens in isolation. Nadia is a partnership in the truest sense, and this thinking reflects countless conversations, debates, and shared work. I want to explicitly recognize John Halpin and Jack O’Connell, my partners at Nadia. Their judgment, discipline, and operating rigor shape every decision we make. Any clarity in this letter is a direct result of their work not mine.
We’re grateful for our partners, our CEOs, and everyone building alongside us.
We’re all in.
Aidan Kehoe
Founder and Managing Partner
Nadia Partners