AI Eats Watts: The Power Play Behind the Compute Boom

Hook: The Divergence Nobody's Watching

Four stocks. All down 20-40% from their 52-week highs. All with rising earnings guidance, exploding order books, and contracts measured in gigawatts. CEG: -34%. TLN: -32%. VST: -39%. GEV: -21%. The market is treating these as momentum plays that ran too hot. The data suggests something else: the market is mispricing the physical layer of the AI buildout. History is just data waiting to be backtested. And right now, the tape is screaming that power, not chips, is the binding constraint.

Context: The Structural Mismatch

AI training clusters are not your grandfather's data centers. A 100k-GPU H100 cluster can draw hundreds of megawatts—the equivalent of a mid-sized city. Power density per rack has jumped from 5-10kW to 100kW+. Utilization rates sit above 90%, running 24/7. This isn't a load profile; it's a base-load industrial consumer with the appetite of a steel mill and the uptime requirements of a hospital.

The market has finally noticed. Constellation Energy (CEG) is restarting Three Mile Island—the site of America's worst nuclear accident—to power AI. Talen Energy (TLN) signed a 1,920MW contract with AWS. Vistra (VST) formed a joint venture with NVIDIA and KKR. GE Vernova (GEV) has $176B in backlogged orders, with AI data center orders doubling year-over-year.

This is the paradigm shift from "chip" to "watt." The question isn't whether AI needs power. It's whether the market understands the quality of power required.

Core: The Order Flow Analysis

Let's dissect the numbers. CEG's 920MW new nuclear PPA, averaging 18.5 years, locks in utility-grade cash flows. Adjusted EPS guidance raised to $11.50-12.50. At $273, that's a forward P/E of roughly 22-24x. For a utility. That's not a growth multiple; that's a scarcity premium.

TLN's AWS contract is the tell. 1,920MW is not a power purchase. It's a co-location play—Talen is building data centers next to its nuclear plant, selling power and compute together. The 4GW option pipeline suggests this is a template, not a one-off. EV/EBITDA at 15-18x looks rich against traditional utilities at 8-12x, but you're not buying a utility. You're buying a toll booth on the AI highway.

VST's Helix JV with NVIDIA and KKR is the most interesting structure. It's not just selling power; it's co-investing in AI infrastructure. EBITDA growth of 30%+ with an EV/EBITDA of 10-12x. That's the cheapest optionality in the group.

GEV is the pick-and-shovel play. $176B backlog. 116GW of gas turbine orders. AI data center orders doubled. The equipment cycle is front-loaded—these orders convert to revenue over 2-3 years, providing visibility that most tech companies can't match.

Here's what the market is missing: these aren't speculative bets on AI adoption. They're contracted, long-dated, and backed by physical assets. The revenue is already booked. The question is execution.

Contrarian: The Blind Spots

Everyone's focused on AI capex. They're watching Microsoft, Google, Amazon, Meta's quarterly guidance like it's the oracle. But the real risk isn't demand—it's delivery.

Grid interconnection queues in the US average 3-5 years. New transmission lines take 7-10 years to permit. The power is contracted, but can it physically reach the data centers? This is the hidden bottleneck that no PPA can solve.

Second, the interest rate sensitivity. These companies carry massive debt loads to fund nuclear and gas turbine builds. If the Fed holds rates higher for longer, financing costs eat into margins. The market's 20-40% drawdown may be pricing this in—or it may be the beginning of a repricing.

Third, the tech giants aren't passive buyers. Microsoft is investing in SMRs. Google is exploring next-gen geothermal. If these alternatives commercialize faster than expected, the long-term thesis for independent power producers weakens. The contracts are real, but the renewal cycle in 2030 will look different.

Takeaway: The Trade

I've audited enough protocols to know that when the narrative and the data diverge, the data wins. The narrative says AI power is overhyped. The data says these companies have contracted revenue, rising guidance, and physical assets that can't be forked or upgraded. The drawdown is a gift, but only for those who understand the risks.

Watch the grid interconnection queue. Watch the Fed. Watch the SMR pilots. If those three variables move in your favor, the current prices are a discount. If they move against, capital preservation beats capital appreciation. History is just data waiting to be backtested. The backtest on AI power is just beginning.