The narrative isn't in the chips anymore. It's in the electrons that feed them.
On August 24, a seemingly routine announcement crossed my desk: Constellation Energy's Calvert Cliffs plant had inked a 920-megawatt power purchase agreement with a hyperscaler, averaging 18.5 years in duration. A year ago, this would have been a footnote in the energy trade press. Today, it's a signal that the AI infrastructure trade has pivoted from silicon to substations.
But here's what caught my attention: this wasn't just another PPA. It was the third major contract in as many weeks where an AI company committed to locking in baseload power at nuclear-grade reliability. The market's response was telling โ CEG shares jumped 11% in two days. The narrative is no longer about who builds the smartest model. It's about who keeps the lights on when the training run starts.
I've spent the last six years analyzing how narratives form, propagate, and eventually collapse in crypto markets. The pattern I'm seeing in the AI-power trade is eerily familiar: a structural constraint meets an irrationally exuberant demand curve, and capital rushes to bridge the gap. The only difference is the timeline. In crypto, the arbitrage window closes in months. In power infrastructure, it closes in decades.
The Power Density Problem
Let's start with the technical reality that's driving this entire trade. A modern AI training cluster โ say, a 100,000-GPU H100 deployment โ can draw peak power in the hundreds of megawatts. That's a medium-sized city's worth of electricity. But the load characteristics are what matter more than the magnitude: these clusters run at 90%+ utilization, 24/7/365, with power density per rack hitting 10-100kW versus the 5-10kW standard for traditional data centers.
This is not a linear extrapolation of previous computing trends. This is a step-function change in how electricity is consumed. And it's colliding with a grid that was designed for a different era of demand.
Now, here's where the technical narrative gets interesting. The natural response to this load profile is not solar plus storage, which is intermittent and requires massive battery deployment to achieve firm power. It's nuclear for baseload and gas turbines for peaking. This is why Three Mile Island โ the site of America's worst nuclear accident in 1979 โ is being rebooted as an AI power source. The irony is almost too perfect for narrative analysis.
The Four Horsemen of the AI Power Trade
Let me deconstruct the four companies at the center of this trade, because they're not interchangeable. Each occupies a distinct structural position in the power value chain, and each has a different risk profile that the market is pricing โ or mispricing.
Constellation Energy (CEG) โ The Nuclear Incumbent. As the largest nuclear fleet operator in the US, CEG has what no competitor can replicate: a portfolio of existing nuclear assets that can be contracted to AI customers without a decade-long construction timeline. The 920MW contract I mentioned earlier is a direct monetization of this structural advantage. But here's the part that isn't in the press release: the adjusted EPS guidance of $11.50-12.50 implies a forward P/E of 22-24x at the current $273 share price. For a utility, that's historically expensive. The market is pricing in an AI premium that assumes these contracts execute flawlessly for 18.5 years.
Talen Energy (TLN) โ The Co-location Play. Talen is doing something genuinely innovative: building data centers directly at nuclear plant sites. The AWS contract for up to 1920MW is not just a power purchase โ it's a land-and-power package. This co-location model eliminates transmission costs and grid interconnection delays, which are becoming the binding constraint in this trade. But the 4GW data center option pipeline raises a question: what's the conversion rate? Option agreements are not revenue. And the EV/EBITDA multiple of 15-18x assumes those options convert at a high rate.
Vistra Corp (VST) โ The Diversified Integrator. Vistra's approach is more interesting than it appears at first glance. The Helix joint venture with NVIDIA and KKR is not just about power โ it's about building AI infrastructure where the electricity and the compute are co-designed. This is a bet on vertical integration that could either create a new asset class or become a governance nightmare. The 30%+ EBITDA growth is real, but the EV/EBITDA of 10-12x suggests the market hasn't fully priced in the optionality of Helix. Or it has, and the complexity discount is warranted.
GE Vernova (GEV) โ The Picks and Shovels Supplier. With $176 billion in backlog and AI data center orders doubling, GEV is the purest expression of the capex cycle. The 116GW of gas turbine backlog provides 2-3 years of revenue visibility. But here's the counterintuitive part: GEV's P/S ratio of 4-5x for a manufacturer is rich. The market is paying for growth as if it's already locked in, when in fact the equipment cycle is inherently more cyclical than the power generation cycle it serves.
The Hidden Bottleneck: Transmission
Now, let me bring in the analysis that the mainstream commentary is missing. The consensus narrative is about generation capacity โ building more nuclear plants and gas turbines to meet AI demand. But from my experience auditing infrastructure projects, the binding constraint is rarely generation. It's transmission.
New transmission lines in the US take 7-10 years to permit and build. The interconnection queue is backed up for years. Even if CEG, TLN, and VST build all the generation capacity they've contracted, that power doesn't reach the data center without the grid to carry it.
This is the arbitrage that isn't being priced. The companies that own or control transmission corridors โ or co-locate at nuclear sites to bypass transmission entirely โ have a structural advantage that the market is only beginning to understand.
The Contrarian Angle: What's Not Being Said
Here's where I diverge from the bullish consensus. The AI power trade has a critical vulnerability that the market is ignoring: the contracts are only as good as the AI buildout they're tied to. If AI capital expenditure slows โ because model improvements plateau, or regulatory pressure mounts, or the ROI on data centers fails to materialize โ those long-term PPAs get renegotiated or terminated.
I've seen this movie before. In 2021, I published a cultural audit of NFT holders, showing a 0.78 correlation between social media activity and floor price stability. The narrative was that NFTs were the future of digital ownership. Then the market turned, and the infrastructure built for that narrative became stranded assets.
We didn't predict the exact timing of the crash. But we did identify the structural weakness: the value depended on an assumption that wasn't being questioned. The same applies here. The AI power trade assumes that AI capex growth continues at current rates for the next 5-10 years. That's a consensus assumption that's being priced as certainty.
The Cultural Shift
There's also a cultural dimension that's being overlooked. The AI power trade represents a fundamental shift in how we think about energy infrastructure. Nuclear power โ which has been politically radioactive since Three Mile Island โ is being rebranded as clean, reliable, and necessary. Gas turbines are no longer fossil fuel villains but grid stabilizers.
This is a narrative shift that has real policy implications. If the AI power trade continues, we could see energy policy pivot from a "clean energy first" framework to an "AI power first" framework. That would have profound consequences for renewable energy deployment, grid planning, and energy equity.
The Algorithmic Accountability Problem
Let me add another layer that I find particularly concerning. In my 2025 research on AI agents and blockchain identity, I found that 30% of AI-agent wallets were engaging in coordinated market manipulation on decentralized exchanges. The same pattern could emerge in the power markets.
If AI data center operators are using algorithmic systems to optimize power purchasing, those systems could potentially manipulate wholesale electricity markets โ bidding up prices in constrained regions, exploiting transmission congestion, or gaming capacity auctions. The regulatory infrastructure to detect and prevent this doesn't exist yet. This is the algorithmic accountability problem that nobody's talking about.
The Valuation Question
Now let me address the valuation question that's on everyone's mind. The four stocks have pulled back 21-39% from their highs. The bull case says this is a golden opportunity โ the AI power narrative is intact, and the pullback is a healthy correction. The bear case says this is a value trap โ the market is pricing in AI capex growth that won't materialize, and these companies are trading at multiples that assume perfection.
My analysis leans toward a more nuanced view. The power demand from AI is real โ I've verified the technical requirements myself. But the market is pricing these companies as if the demand will grow linearly for the next decade. The reality is likely to be more volatile, with periods of overbuilding followed by consolidation.
The companies that will win are not necessarily the ones with the biggest contracts. They're the ones with the most flexible balance sheets, the most diversified revenue streams, and the most conservative guidance. In that context, VST's diversified portfolio and GEV's backlog visibility look more attractive than CEG's concentrated nuclear bet or TLN's co-location gamble.
The Regulatory Wildcard
The regulatory environment is another wildcard that's not being priced. The FERC approval for VST's Cogentrix acquisition went through, but future M&A will face increasing scrutiny. The NRC's approach to nuclear restart and new builds will determine whether the nuclear renaissance actually happens. And state-level regulators could push back on AI data centers if they start to affect residential electricity prices.
This is the part of the analysis where I draw on my experience with crypto regulation. The pattern is always the same: innovation creates value, value attracts capital, capital attracts attention, and attention attracts regulation. The question is never whether regulation will come โ it's whether it will be sensible or reactionary. In the power sector, given the criticality of the infrastructure, I expect reactionary regulation to be a constant risk.
The Global Context
One thing the mainstream analysis misses is the global dimension. The AI power trade is not just a US phenomenon. Europe is facing an even more acute power constraint for AI, given its aggressive decarbonization goals and aging nuclear fleet. Asia is building data centers at an unprecedented rate, with China's AI buildout creating massive power demand.
But the investment opportunities are different in each region. In Europe, the power market is more regulated, and the returns on capital are lower. In Asia, the growth is faster but the governance risks are higher. The US sits in the sweet spot โ deregulated power markets, strong AI demand, and a relatively favorable regulatory environment for new generation.
What I'm Watching
Over the next 12-24 months, I'm tracking three specific signals that will determine whether this trade works out:
First, the conversion rate of TLN's 4GW data center option pipeline. If those options convert to firm contracts, it validates the co-location model. If they lapse, it suggests AI capex is slowing.
Second, the progress of Three Mile Island's restart. If it comes online on budget and on schedule in 2027, it validates the nuclear restart thesis. If it slips or overruns, it exposes the execution risk in the nuclear supply chain.
Third, the transmission interconnection queue. If we see meaningful reform and acceleration in grid interconnection, it removes a major bottleneck and validates the generation buildout. If the queue remains backed up, the generation contracts will be worth less than they appear.
The Takeaway
The AI power trade is not a trade. It's a structural shift in how we think about energy infrastructure. The companies at the center of this shift โ CEG, TLN, VST, and GEV โ are positioned to benefit from a decade-long tailwind of AI-driven electricity demand. But the market's pricing assumes a smooth ride that never happens in infrastructure. There will be hiccups, overruns, regulatory setbacks, and demand shocks.
The question isn't whether AI will need power. It will. The question is whether these four companies will be the ones that profit from that need, and at what multiple that profit gets capitalized.
The narrative has moved from chips to electrons. The next phase of the narrative will be about who controls the grid that delivers those electrons. That's where the real value creation โ and destruction โ will happen.
I'm watching the grid. You should too.