By Emily Jones, Token Fund Investment Manager
Hook: The Ghost at the Grid
Here's a narrative shift that should make every investor pause mid-scroll: the AI trade has left the chip, left the software, left the code entirely. It is now sitting inside a cooling tower in Pennsylvania. Three Mile Island—the site of America's most infamous nuclear accident—is coming back online to feed an AI data center. That isn't a metaphor. That's a 920-megawatt power purchase agreement, signed in ink, backed by an 18.5-year term. We don't just track trends; we hunt their origins. And the origin story of the next market cycle is not being written in Python. It's being written in megawatt-hours.
Context: The Electric Kool-Aid Test
Let me give you some background that frames this shift. The AI industry has hit a wall that no amount of algorithmic optimization can solve: power. A single modern AI training cluster—think 100,000 NVIDIA H100 GPUs—can draw peak loads in the hundreds of megawatts. That's the electricity consumption of a small city, running at 90% utilization, 24/7. Traditional data centers were designed for 5-10kW per rack. AI needs 10 to 100 times that, with zero tolerance for downtime. This isn't just a scaling problem; it's a structural mismatch between a digital economy and a physical grid that was never built for it.
Based on my audit experience across protocol infrastructure, I've learned that when a system's core input becomes constrained, the narrative doesn't die—it migrates. We saw it with gas fees on Ethereum post-CryptoKitties, and we're seeing it now on the grid. The bottleneck of AI was always going to be compute, but the bottleneck of compute has become power. And the market is responding accordingly. Four companies have emerged as the vanguard of this power play: Constellation Energy (CEG), Talen Energy (TLN), Vistra Corp (VST), and GE Vernova (GEV).
Core: The Narrative Mechanics of a Utility Revival
Let's dig into the numbers. This is where the story gets its teeth.
Constellation Energy (CEG) is the nuclear heavyweight. They operate the largest nuclear fleet in America, and they're in the process of restarting Three Mile Island—a facility that hasn't produced power since 1979. The new contract, a 920MW PPA with a weighted average term of 18.5 years, is the kind of revenue visibility that public utilities dream about. They've raised their adjusted EPS guidance to $11.50-12.50. But here's the kicker: the stock is down 34% from its 52-week high of $412.70, trading around $273. At that level, you're looking at a forward P/E of roughly 22-24x. That's rich for a utility, but it's a bargain if you believe the AI demand curve is real.
Talen Energy (TLN) is perhaps the most fascinating structure. They're not just selling power; they're co-locating data centers directly at their nuclear plant in Pennsylvania. The AWS deal—a 1,920MW long-term contract—is a game-changer. They have a 4GW data center option pipeline. The stock is down 32% from its high, trading around $305. With adjusted EBITDA guidance at $20.25-22.25 billion, the EV/EBITDA multiple is in the 15-18x range. That's double the traditional utility multiple, but you're not paying for electricity—you're paying for a technology platform with a power source.
Vistra (VST) is the diversified play. They're not just selling power; they're building a joint venture called Helix with NVIDIA, KKR, and the Kuwait Investment Authority. That's a big signal—it says they're building "power plus compute" as a single product. They raised their EBITDA guidance by 30%+ in the quarter, yet the stock is down 39% from its high of $219.82. Their EV/EBITDA is around 10-12x, which is the most reasonable relative valuation in this group.
GE Vernova (GEV) sits at the upstream end. They manufacture the gas turbines and grid infrastructure that make all of this possible. With a $176 billion backlog and AI data center orders doubling, they have 116GW of gas turbine backlog. The stock is down 21% from its all-time high, trading around $942, implying a P/S ratio of 4-5x. That's rich for a manufacturer, but the order book provides a degree of certainty you rarely see in capital goods.
The Core Insight: A New Kind of Security
Now, here's the technical nuance that most people are missing. The AI power demand isn't just a quantity problem—it's a quality problem. It's a base-load problem. It requires power that is not intermittent, power that can run at a 90%+ capacity factor without flinching. This is why nuclear and natural gas are winning over pure renewables. Solar only works when the sun shines. Wind only works when the wind blows. AI clusters need the power equivalent of a heartbeat—stable, continuous, high-quality. This is exactly why the "narrative velocity" of this sector is accelerating. We're seeing power generation shift from being a commodity to being an infrastructure.
This is also why the social layer of this narrative is so sticky. When you have tech giants signing 20-year contracts to secure power, you're not just buying electricity; you're buying a partnership with a primary resource. Security is the canvas; liquidity is the paint. The liquidity is the capital flowing into these projects. The security is the physical infrastructure providing the power.
The Contrarian Angle: The Bedrock of the Risk
But hold on. Let me play devil's advocate. I've been burned before. The Terra/Luna collapse taught me that narrative is fragile when it's not anchored to economic reality. And there are three things that this AI power narrative is not showing you.
First, the grid bottleneck. Even if all these power plants are built, the electricity has to travel across the U.S. grid. The average transmission line takes 7-10 years to get approved and built. We're talking about a massive backlog of interconnection requests. If the electricity can't get to the data center, the PPA is just a piece of paper.
Second, the rate sensitivity. These are capital-intensive businesses. They're running on high debt. If the Fed holds interest rates high, financing costs will eat into margins. The 20-40% drawdown might be the market starting to price in this exact risk.
Third, and most importantly, the AI capex cycle. The entire thesis rests on the assumption that Microsoft, Google, Amazon, and Meta will continue to spend hundreds of billions of dollars on AI infrastructure. If they decide the returns are not materializing—or if they pivot to a more efficient model—they could renegotiate or break these contracts. The exit is easy; the narrative is the hard part.
Takeaway: The Next Story
So what's the bottom line? The AI power narrative is structurally real, but the price is not exactly cheap. These stocks have corrected significantly, but they're still trading at premiums to the broader utility sector. The opportunity is in the execution. I'm watching Three Mile Island's restart timeline, the AWS contract's specific pricing terms, and the next major AI capex guidance from the tech giants.
The market is learning that finding the human heartbeat inside the cold code means understanding that the code lives in a physical world. The next frontier is not just about "digital gold" or "DeFi yield." It's about who owns the physical assets to power the machine. The question isn't whether AI will consume more power; it's whether the grid can deliver it—and at what cost. That is the next big trade. We don't just track trends; we hunt their origins. The origin is now a data center in Pennsylvania, powered by a resurrected nuclear ghost. The future is not in the code. It's in the current.