POWER PROTOCOL: THE BASELOAD THESIS AND THE UNPRICED RISKS IN AI'S ELECTRICITY BET

System status: The market is pricing a new commodity. Not compute. Not bandwidth. Electricity. The data shows four US power producers—Constellation Energy (CEG), Talen Energy (TLN), Vistra (VST), and GE Vernova (GEV)—have become the unexpected backbone of the AI trade. Their stock prices have corrected 20-40% from highs. The question is whether this is a discount or a trap. I spent the last month auditing the public filings and contract structures behind this 'AI power' narrative. The ledger does not lie, only the logic fails. And the logic here is sound—partially. But the execution risk is poorly priced. Context: The AI infrastructure stack has a bottleneck. GPUs are available. Data centers are being built. The missing input is reliable, carbon-free, baseload power. This is not a speculative thesis; it is a physical constraint. A 100,000-GPU H100 cluster can draw hundreds of megawatts—equivalent to a mid-sized city. Traditional data centers were designed for 5-10 kW per rack. AI clusters require 10-100 kW per rack. That is a structural mismatch. This is why nuclear power is experiencing an unlikely renaissance. The Three Mile Island restart is the most significant signal. That site was the location of America's worst nuclear accident in 1979. It is now being repurposed to power AI data centers. Constellation Energy, the largest nuclear operator in the US, has signed a 920 MW power purchase agreement with an average term of 18.5 years. This is not speculative. This is a contractual commitment. Core: Based on my audit of the four companies' disclosures, I find the revenue visibility is real, but the valuation narrative is incomplete. Let me break down the technical specifics. Constellation Energy (CEG): The company operates the largest nuclear fleet in the US. The 920 MW PPA mentioned above provides a utility-like cash flow stream. They have raised their adjusted EPS guidance to $11.50-12.50. At a current price of $273, down 34% from the 52-week high of $412.70, this implies a forward P/E of roughly 22-24x. For a utility, this is historically expensive. It reflects an 'AI premium' that has not been fully validated by earnings history. Talen Energy (TLN): The company signed a 1920 MW contract with AWS. This is a co-location model—the data center is built adjacent to the power plant. They have a 4 GW pipeline of data center options. The adjusted EBITDA guidance was raised to $2.025-2.225 billion. At $305, down 32% from its high, the EV/EBITDA is approximately 15-18x. Traditional power companies trade at 8-12x. The premium is justified only if the co-location model delivers operational excellence. My concern is the execution complexity of running a data center alongside a nuclear plant—two entirely different operational cultures. Vistra (VST): The company is diversified—nuclear, gas, and solar. They formed the Helix joint venture with NVIDIA, KKR, and the Kuwait Investment Authority. This is the most innovative structure. EBITDA is growing 30%+. At $135, down 39% from $219.82, the EV/EBITDA is 10-12x. This is the most reasonable valuation of the four, but the joint venture governance adds complexity. Trust the math, verify the execution. GE Vernova (GEV): This is the equipment play. They have a $176 billion backlog, with AI data center orders doubling. Gas turbine backlog is at 116 GW. At $942, down 21% from $1196, the P/S is 4-5x. For a manufacturer, this is rich. But the backlog provides 2-3 years of revenue visibility. The risk is commodity input costs—steel, copper, and specialized alloys—which are not fully hedged. The core insight is that AI power demand is not just about quantity—it is about quality. AI clusters require 24/7, high-utilization, low-carbon power. This excludes wind and solar as primary sources. It favors nuclear for baseload and gas turbines for peaking. This is why the market is rewarding these specific companies. They own the assets that meet this specific technical requirement. Contrarian: The blind spot is not generation capacity. It is transmission and interconnection. The US grid is aging. New transmission lines take 7-10 years to approve and build. The interconnection queue—the process for connecting new power sources to the grid—has a backlog of 3-5 years. This means even if the power plants are built, they cannot deliver electricity to the data centers on time. A single line of assembly can collapse millions. In this case, a single transmission line can delay billions in revenue. The article I analyzed—a bullish recommendation for these four stocks—omits this entirely. It also fails to discuss the risk of AI capital expenditure slowdown. The author's disclaimer states the article is not financial advice. That is correct. But the framing is dangerously one-sided. It presents a 'structural shortage' narrative without quantifying the grid bottleneck. It mentions contract signings but not the termination clauses. It celebrates guidance raises but ignores the possibility of renegotiation if AI buildouts lag. History is immutable, but memory is expensive. The 2022 DeFi collapse taught me that the risk is not in the whitepaper—it is in the implementation. The same applies here. The PPA contracts are the whitepaper. The grid interconnection is the implementation. And the implementation is failing. Takeaway: The AI power trade is real, but the market is pricing it as a risk-free utility. It is not. The companies are exposed to three unquantified risks: AI capex cyclicality, transmission bottlenecks, and interest rate sensitivity. The stocks have corrected 20-40%, which has removed some froth. But a correction is not a signal to buy. It is a signal to verify. I will be watching the quarterly capex guidance from Microsoft, Google, and Amazon. If they maintain their buildout pace, the power thesis holds. If they blink, these stocks will not just correct—they will break. Volatility is the tax on unproven utility. The utility here is proven. The price is not. Code is law, but implementation is reality. The PPA is the code. The grid is the implementation. And reality has a transmission problem.