On August 24, 2026, Constellation Energy (CEG) announced the restart of Three Mile Island—the site of America’s worst nuclear accident. The news wasn’t buried in a regulatory filing; it was front-page financial press. Why? Because Microsoft had signed a 20-year power purchase agreement to buy the output. This is not a nuclear revival. It’s a signal of the most significant structural shift in energy markets since the shale revolution—and the crypto-native analysts who ignore it are missing the next narrative cycle.
I’ve been watching this pattern since 2017, when I first modeled Chainlink’s oracle incentives. Back then, the narrative was ‘trustless data.’ Today, it’s ‘trustless power.’ The underlying mechanism is the same: a bottleneck emerges, and the market rushes to price in the solution. But as with DeFi’s liquidity mining craze, the real winners are not the flashy frontends—they are the infrastructure providers that most retail investors overlook.
Context: The Historical Narrative Cycle
Every major crypto narrative has an infrastructure counterpart. During the ICO boom, the infrastructure was Ethereum’s smart contract platform. During DeFi Summer, it was automated market makers and oracles. During the NFT mania, it was decentralized storage. Now, in the AI era, the infrastructure is electricity. The narrative is no longer about which AI model will win; it’s about which grid will power the training.
AI compute demand is not linear—it’s exponential. A single 100,000-H100 cluster consumes 300-500 MW, equivalent to a small city. And these clusters run 24/7 at 90%+ utilization. The power density per rack has jumped from 5-10 kW to 50-100 kW. Traditional data centers were designed for bursty workloads; AI is a base-load monster. The grid was not built for this.
Core: The Mechanism of AI Power Demand
Let’s deconstruct the narrative. The market says: ‘AI needs power, so buy utilities.’ But that’s surface-level. The real mechanism is a structural mismatch between the demand curve (spiky, high-density, low-carbon) and the supply curve (intermittent renewables, slow-to-build nuclear, flexible gas). The winners are those who can match the two.
Constellation Energy (CEG) is the largest nuclear operator in the US. Nuclear provides 24/7 carbon-free base load—perfect for AI. The Three Mile Island restart is not just a symbolic win; it’s a 920 MW PPA with an 18.5-year average term. That’s $1.5B+ in locked-in revenue. CEG raised its adjusted EPS guidance to $11.50-12.50, implying a forward P/E of 22-24x. For a utility, that’s expensive. But the market is pricing in a ‘narrative premium’—the idea that nuclear will be the backbone of AI power.
Talen Energy (TLN) takes a different approach. It owns the Susquehanna nuclear plant and is co-locating data centers on-site. Its deal with AWS for up to 1,920 MW is the largest known PPA in history. TLN’s EV/EBITDA is 15-18x, far above traditional utilities. The justification: TLN is not just selling power; it’s selling power adjacency. Data centers on-site avoid transmission losses and interconnection delays. That’s a competitive moat.
Vistra (VST) is the diversified play. It owns nuclear, gas, and renewables. Its Helix joint venture with NVIDIA and KKR is a novel structure: VST provides the power, NVIDIA provides the compute, and KKR provides the capital. This is a vertical integration of narrative—the energy provider becomes a co-investor in AI infrastructure. VST’s EBITDA is growing 30%+, but its EV/EBITDA is only 10-12x. Why the discount? Because the market is uncertain about the partnership’s execution. I’ve seen this before: in 2020, I analyzed Compound’s governance token distribution and found that 40% of liquidity was speculative. The same skepticism applies here.
GE Vernova (GEV) is the picks-and-shovels play. Its $176B backlog includes 116 GW of gas turbines, and AI data center orders doubled year-over-year. Gas turbines are the flexible backup for renewable-heavy grids—they can ramp up in minutes. GEV’s P/S ratio of 4-5x is high for a manufacturer, but the backlog provides 3-5 years of visibility. The catch: gas turbine orders are cyclical. If AI capex slows, the backlog could shrink.
Contrarian: The Blind Spots Everyone Ignores
Now, the counter-narrative. The market is pricing in a perfect execution scenario. But I’ve been auditing narrative decay for years, and I see three cracks.
First, grid interconnection is the real bottleneck. The US has 1,200 GW of solar and wind projects waiting in interconnection queues, with average wait times of 5-7 years. Even if generation is built, the transmission lines to carry that power to AI data centers don’t exist. The 2026 Inflation Reduction Act allocated $10B for grid upgrades—a fraction of what’s needed. Without transmission, the PPA is just a piece of paper.
Second, tech giants are building their own energy assets. Microsoft is investing in small modular reactors (SMRs). Google is backing geothermal startups. If SMRs achieve commercial deployment by 2028, the independent power producers (IPPs) could lose their pricing power. The narrative of ‘AI needs IPPs’ might be a multi-year story, not a decade-long one.
Third, AI capex is not guaranteed. The 2025-2026 AI spending boom is driven by a handful of companies (Microsoft, Google, Amazon, Meta). If the returns on AI models diminish, or if a new regulation caps data center energy use, the capex cycle could turn. The 30-40% drawdowns in CEG, TLN, VST, and GEV from their highs suggest the market is already pricing in some risk. But the valuation multiples still imply perfection.
Takeaway: The Next Narrative Shift
So where do we go from here? The next narrative is not ‘buy power stocks’—it’s power sovereignty. Tech giants will increasingly own their own generation assets, either through PPAs, joint ventures, or direct construction. The value will shift from the generators to the grid operators and transmission builders that enable the connection. Think of companies like Quanta Services or NextEra Energy Partners—not yet in the spotlight, but poised to benefit from the infrastructure buildout.
I’ll leave you with a question: If the bottleneck is not power but the grid, then what is the narrative that captures that? As a narrative hunter, I’m already tracking the next prey.
