The chips get the headlines. But the hyperscalers themselves keep saying the real bottleneck isn’t compute — it’s electricity. There’s a whole layer of the AI value chain sitting one step past where the crowd is looking.
Ask most people how to play artificial intelligence and you’ll get some version of the same answer: chips. Nvidia, the semiconductor complex, the hardware that does the computing. That’s Layer 1 — and it’s the most crowded trade on earth.
But listen to what the people actually building this stuff are saying. The executives running the largest AI buildouts have been remarkably consistent: the constraint is no longer getting the chips. It’s finding the power to run them. As one of them put it, the world has moved from a place where the bottleneck was GPUs to one where the bottleneck is electricity. When the buyers of compute tell you their binding constraint has shifted, that’s not noise — that’s a map.
Follow that map down the value chain and the crowd thins fast.
The Three Layers — And Where The Attention Runs Out
Layer 1 is compute — the chips. Enormous, essential, and priced accordingly, with the whole world already crowded in.
Layer 2 is the equipment — the transformers, gas turbines, switchgear, and grid hardware needed to physically deliver power to a data center. This was the smart “first derivative” trade, and it’s now well-discovered; the equipment names have had their moment.
Layer 3 is the power itself — the utilities and merchant generators that own and sell the electricity, through long-term contracts, regulated rate-base growth, and capacity markets. This is the least-crowded layer, and it’s structurally different from the two above it. You’re not betting on who wins a manufacturing race or a chip cycle. You’re looking at entities that, increasingly, have contracted to sell power to hyperscalers for 15 to 20 years.
That contracting piece is what makes this layer analytically interesting, so it’s worth understanding how the money actually gets made.
How This Layer Monetizes The Boom
There are three distinct mechanisms, and they matter because they carry very different risk.
Capacity markets. In PJM — the largest US grid — suppliers get paid simply to guarantee they’ll have power available years ahead. That payment has gone vertical: the capacity clearing price rose roughly ninefold in a single year, from about $29/MW-day to $270, and the most recent auction cleared at the $329 FERC-imposed cap. Data-center load is the primary driver, with US data-center power demand running around 76 GW in 2026, up from roughly 50 GW in 2024.
Long-term PPAs. Hyperscalers are increasingly signing 15-to-20-year power-purchase agreements — some for dedicated nuclear output — locking in demand for decades. This turns a speculative “AI might need power” story into contracted, visible revenue. Some of these deals even have tech companies paying for new generation whether or not they use every megawatt.
Rate-base growth. For regulated utilities, the buildout means enormous approved capital plans — new transmission, generation, and interconnection — on which they earn a regulated return. Several utilities have sharply raised multi-year capital plans and load-growth forecasts specifically on data-center demand, in some cases revising a single quarter’s demand outlook up by double digits.
Two Archetypes, Two Very Different Risk Profiles
The important analytical distinction inside this layer is between the two kinds of company, because they are not the same trade.
Regulated utilities earn a set return on their capital base. The upside is steadier and more visible (approved capital plans, load growth), but the risk is regulatory and political: rate-case approvals, allowed-return decisions, and a growing affordability backlash as households see bills rise and regulators demand proof that data centers — not ratepayers — foot the bill. Some states are already creating separate rate classes for hyperscalers.
Merchant generators sell power into wholesale markets (ERCOT, PJM) at market prices, sometimes anchored by big PPAs. The upside is far greater if power prices stay elevated — they capture the spikes directly — but so is the volatility. Earnings swing with power prices, and analyst price targets on these names can span an enormous range, reflecting genuine uncertainty. This is the higher-beta, more “AI-pure” expression, and it behaves like it.
Same theme. Completely different behavior. Confusing the two is one of the most common mistakes in this space.
The Catch — And Why It Ties Straight Into The Rates Story
Here’s the part that most “AI power” write-ups conveniently skip, and it’s the reason this isn’t a one-way trade.
Utilities are the textbook bond-proxy sector. They fund massive capital programs with debt and equity, and their dividend yields compete directly with bond yields for income investors. Which means the single biggest macro event of the moment — global long-term yields spiking to their highest level since 2008, which I covered in the last piece — is a direct headwind to this entire layer. Higher yields make the debt and equity these companies need to fund their buildout more expensive, and they make a utility’s dividend less attractive against a suddenly-competitive risk-free rate.
This is the crux. The demand isn’t the question — the AI power buildout is real, and increasingly it’s contracted for 15 to 20 years, which is about as un-speculative as a growth story gets. The question is timing. A rate-sensitive sector running its most capital-intensive decade in history, straight into the most expensive funding environment since 2008, can see its re-rating delayed for as long as long yields stay elevated — even with the demand entirely intact.
And Two Risks Worth Respecting
Beyond rates, two things keep this honest. First, demand-projection risk: not every announced data center gets built, construction has shown signs of moderating, and PJM’s own market monitor has openly questioned the credibility of the long-range demand forecasts feeding its models. Second, the affordability backlash: as data-center growth pushes household bills up, the political and regulatory response — new rate classes, cost-allocation fights, approval delays — is a real, sector-wide overhang, particularly for the regulated names.
What This Actually Means
The reframe is simple but it changes where you look: the AI trade is not only a chip trade. The bottleneck has moved to power, and there’s a distinct, less-crowded layer of the value chain — the utilities and generators that own the electrons — that monetizes the boom through capacity markets, long-dated PPAs, and rate-base growth. It’s a genuinely two-sided setup: contracted, multi-decade demand on one side; a bond-proxy sector colliding with the worst rate backdrop since 2008 on the other.
The move that matters isn’t chasing the theme — it’s knowing how to tell the archetypes apart, which mechanism each name is actually levered to, and how to weigh contracted demand against the rates headwind so you can see whichnames the setup genuinely favors and when.
In the full breakdown for subscribers, I go deeper on the sector:
- The mechanism map — which specific names are levered to capacity markets vs. PPAs vs. pure rate-base, and why that determines how each behaves as yields move
- The regulated-vs-merchant framework — how to size the regulatory/affordability risk against the power-price upside, name by name
- The rate-sensitivity screen — which balance sheets face the largest equity-issuance needs into this yield environment, and which are the most insulated
- Where contracted demand is strongest and most visible versus where the “data-center” label is doing more work than the contracts
That’s the analysis that turns a theme into a framework. It’s for paid subscribers.
👉 The mechanism map, the regulated-vs-merchant framework, and the rate-sensitivity screen are available to subscribers. Subscribe to read it.
Disclaimer
This article represents personal analysis and reflects solely the personal views of the author. It is provided for informational and educational purposes only and does not constitute investment advice, a research recommendation, or an offer or solicitation to buy, sell, or hold any security. It does not assign a rating to, or recommend, any company or security named or described. Any companies referenced are discussed solely as illustrations of sector dynamics. The author is not currently licensed or certified to provide investment advice in any jurisdiction. Information is based on publicly available sources believed reliable at the time of writing but is not guaranteed as to accuracy or completeness. Past performance does not indicate future results, and all investments carry risk, including loss of principal. Readers should conduct their own due diligence and consult a licensed professional before making any decision.
Leave a Reply