Over the past 12 months, I have watched a narrative solidify that transcends the usual crypto-cycle chatter. It is not about block size, gas fees, or the latest L2. It is about the physical layer of the digital economy: the gigawatt. In August 2026, Constellation Energy (CEG) signed a 920MW power purchase agreement tied to the restart of Three Mile Island, and Talen Energy (TLN) inked a massive 1,92GW deal with AWS. These are not incremental contracts. They represent a structural transfer of capital from the chip to the conductor. The ledger of the AI revolution is no longer written in TFLOPS; it is written in megawatts. And the market, having bid up these energy names and then watched them correct 30-40% from highs, is now trying to figure out whether this is a dip or a distribution. We do not build on hype; we build on consensus. And the consensus in the data is that we face a physical shortage that code cannot patch.
I have spent the last decade reading on-chain reserves as liquidity signals, but the liquidity signal for the next decade is the interconnection queue at PJM and ERCOT. The thesis is simple: AI compute expansion is a linear function of power availability, not chip fabrication. In this piece, I will break down the four primary vehicles capturing this trend—CEG, TLN, VST, and GEV—using the specific contract structures and order books that are currently public. We will then look at the systemic blind spots, specifically the transmission bottleneck and the interest rate regime, that the narrative often ignores.
The Context: The Physical Bottleneck
To understand why power is the new constraint, you have to look at the physical architecture. The latest generation of AI training clusters, such as the 100,000-H100 GPU class, requires a peak load of several hundred megawatts. That is not a server farm; that is a city. The power density per rack jumps from 5-10 kW in the traditional data center to 10-100 kW plus. The load is not just high; it is high-quality, meaning it requires a 99.999% uptime, a base-load capacity factor, and a carbon-conscious profile to satisfy the ESG mandates of the hyperscalers.
This is where the technical reality diverges from the 'renewables-only' crowd. Intermittent sources do not work for this load profile without massive storage, which is not yet economically scalable at the gigawatt level. That is why the market is pivoting to the old guard: nuclear and gas. The restart of Three Mile Island (the site of the 1979 incident) is the single most potent symbol of this shift. It shows that necessity overrides historical trauma. The technical suitability of nuclear for AI is not just about volume; it is about quality of power—continuous, carbon-free, and predictable. Similarly, the gas turbine, via GE Vernova (GEV), is not just a bridge fuel; it is the peak-shaving mechanism required to balance the grid when the base load is met but the sun isn't shining and the wind isn't blowing.
The Core: Four Players, One Demand Curve
Let us look at the scoreboard through the lens of liquidity forecasting. I have always said, "Follow the liquidity, ignore the noise." In this sector, the liquidity is not in the mempool; it is in the backlog and the PPA.
Constellation Energy (CEG): The Nuclear Base. The company operates the largest nuclear fleet in the US. They have signed a 920 MW PPA with an average duration of 18.5 years. This is the definition of an annuity. They have raised their adjusted EPS guidance to $11.50-$12.50. At the current price of $273, down 34% from the high of $412, the forward P/E sits at roughly 22-24x. That is not cheap for a utility, but it is a premium for a stock with a monopolistic, 18-year revenue lock. The risk here is not demand; it is the NRC. The regulatory approval for the TMI restart is the critical path.
Talen Energy (TLN): The Co-location Specialist. The AWS contract for 1,920MW is the largest single data center contract ever signed. It is not just a PPA; it is a co-location model where the data center sits behind the meter of the nuclear plant. This eliminates the grid bottleneck entirely—a crucial detail. The EBITDA guidance was raised to $2.025-$2.225 billion, with a pipeline of options for 4GW. The stock, down 32% from its high, trades at an EV/EBITDA of 15-18x. The risk here is concentration. If the hyperscaler pauses construction, the entire thesis breaks. But the structure suggests they are building a fortress of power and compute that is hard to replicate.
Vistra (VST): The Diversified Play. VST is the gas and solar giant, but their move to partner with NVIDIA, KKR, and the Kuwait Investment Authority in the Helix JV is the game-changer. They are not just selling power; they are building AI infrastructure. The EBITDA growth of 30%+ puts their EV/EBITDA at 10-12x, which is the most reasonable valuation of the four. They have the most significant exposure to the Texas grid (ERCOT), which has the highest demand growth. The risk is governance and execution. But if Helix lands its first clients, the market will re-rate this from a utility to a tech infrastructure play.
GE Vernova (GEV): The Picks-and-Shovels Monopoly. With $176 billion in backlog, the machine is already running. AI data center orders have doubled, and they hold a 116GW order book for gas turbines. The CEO has guided to break through 125GW by year-end. They are the only company in this list that does not carry the risk of contract termination—they sell the machines regardless of who the operator is. The P/S ratio is 4-5x, which is rich, but the order visibility is 2-3 years out. The bull case is that the transmission grid upgrade will be the second wave of their revenue, as they also supply the high-voltage equipment. The bear case is that they are a cyclical manufacturer at peak earnings, and the market will discount that.
The Contrarian Angle: The Blind Spots in the Ledger
While the four players above represent the direct route, we must analyze the systemic risk that the market is pricing in. The data tells us that the current electricity demand for AI in the US is roughly 50-100 GW. Current nuclear output is about 100 GW. That means we need to double the nuclear fleet or build massive gas capacity just to meet the AI demand alone, not to mention the electrification of the grid and the EV fleet. This is not just a challenge; it is a capital allocation challenge.
The market is currently priced for a continuation of this capex cycle. But I see three blind spots.
First, the transmission bottleneck. The generator is the gas turbine, but the grid is the power line. The average lead time for a high-voltage transmission line in the US is 7-10 years. The interconnection queue is backed up for years. The power may be generated, but it cannot be delivered. This is the "physical" constraint that the market hasn't fully priced in. This could delay the revenue recognition for CEG and TLN if they cannot get their power to the load centers.
Second, the interest rate shadow. These are not software companies. They are capital-intensive beasts with high leverage. The cost of capital is the lifeblood. If the Fed holds rates high, the financing costs of the $176 billion backlog for GEV will be an issue, and the debt refinancing for VST and CEG will erode the FCF. The discount rate applied to these future cash flows is high, which caps the upside.
Third, the self-building threat. The Hyperscalers are not just signing contracts; they are investing in Small Modular Reactors (SMRs) themselves. Microsoft, Google, and Amazon are already funding SMR startups. If this trend accelerates in 2027-2028, the reliance on independent producers could be weakened. This is the "decoupling" thesis. The market is betting on the continued existence of the independent power producer, but the hyperscalers are taking control of the physical supply chain.
The contrarian angle is that the best trade is not the utility that is getting the PPA but the equipment manufacturer that is selling to everyone, regardless of who wins the contract. But the blind spot is the grid—we might have the power, but we don't have the pipes to move it.
The Takeaway: Positioning for the Chop
We are in a sideways market for these equities. The initial hype is gone. The market is waiting for the Q3 earnings to see if the capex guidance holds. The construction is still in the early innings. The data from the last 7 days suggests that the pressure on the grid is not going away. The real test is the execution of the Three Mile Island restart and the AWS-TLN co-location.
In the long run, the "AI Electricity" is the only game in town. But we are building the infrastructure on a 10-year timeline. The price may be correcting, but the physical reality is not. The macro trend is set. The question is whether you have the balance sheet to survive the quarterly volatility. As for the current players, the ones with the lowest cost of capital and the highest grid security will be the winners. The ones with the leverage and the long approval times will be the ones that get re-rating down.
I have been through the ICO era, the DeFi Summer, and the NFT winter. The physical economy is always the ultimate settlement layer. The ledger remembers what the market forgets. And the market is forgetting that a gigawatt is a physical asset, not a line of code. The chase is not for the next token; it is for the next 1,000 megawatts.