The 32 GW Ghost: What the AI Power Panic Actually Prices

Last cycle, PJM Interconnection's capacity auction cleared at $269.92 per MW-day. The prior year, roughly $28. A 9x repricing inside a single auction window. Markets do not move like that quietly. They move like that when something structural has broken underneath the price.

Somewhere between those two numbers sits a claim that traveled from a Morgan Stanley research note into crypto media within days: the United States faces a 32 GW power gap for AI data centers by 2028. The figure spread fast. It touched utilities. It touched the nuclear SMR complex. It touched Bitcoin miners. Almost nobody asked the boring question first. What is the denominator?

I have spent enough of my career reconstructing ledgers to recognize a single-source number the moment it enters a narrative. It stops being a hypothesis. It becomes a fact. And facts, in this market, get priced before they get verified. The 32 GW figure is now doing valuation work it was never qualified to do.

Separating the fact layer from the forecast layer

Let me pull these apart, because the two are being quoted as if they were the same thing.

The directional claim is solid. U.S. data center electricity consumption rose from roughly 76 TWh in 2018 to about 176 TWh in 2023 — around 4.4% of national load — per Lawrence Berkeley National Laboratory. LBNL projects 325 to 580 TWh by 2028. The IEA separately models global data center demand moving from roughly 460 TWh in 2022 toward 1,000 TWh by 2026. Two independent sources. Same direction. That is the fact layer. I trust it.

The physical constraint layer is also real. Interconnection queues in PJM, ERCOT, and MISO are backlogged. New large loads wait four to seven years in some regions. Transformer lead times stretched from months to three-to-five years. Gas turbine order books at GE Vernova and Siemens Energy are sold into 2027 and 2028. These are disclosed, checkable, and mutually consistent.

Now the number everyone quotes. 32 GW. No methodology. No base year. No breakdown between demand-side gap, interconnection gap, or peak-load gap. No stated treatment of projects already under construction. In forensic terms, this is an unlabeled exhibit entered into evidence without a chain of custody.

That distinction matters more than the number itself. A directional truth and a precise falsehood can coexist in a single sentence, and that sentence can still move billions.

The asset nobody is pricing

Here is where the crypto market should be paying attention, and mostly is not.

If power — not silicon — becomes the binding constraint on AI compute, then the scarce asset is not a GPU. It is an energized interconnection right. A signed power purchase agreement. A substation with headroom. These take years to create and cannot be conjured by capital alone.

Bitcoin miners own exactly that.

This is the trade the coverage buried. Let me trace it the way I trace any ledger.

A Bitcoin mining operation is, structurally, a power contract with a hashing attachment. The hardware is depreciating and replaceable. The electricity contract — often fixed-price, often with curtailment rights, frequently sited in ERCOT or the mid-continent wind belt — is the durable asset. When hashrate economics compress, the contract does not vanish. It waits.

So when AI and HPC demand explode and new interconnections take half a decade, the fastest path to energized compute is to acquire the miner. Not the chips. The contract.

The operators who understood this early stopped describing themselves as miners. They reframed as digital infrastructure. HPC hosting. The revenue mix shifted. The multiple followed. A mining multiple runs roughly 3–6x EBITDA and is hostage to Bitcoin's price. A data center REIT multiple runs 15–25x and is hostage to contracted cash flow. Same physical asset. Different story. Different price. This is not sentiment. It is a reclassification of the cash flow's durability.

I ran this comparison across the disclosed fleet. The pattern held. Operators with secured, long-duration power and existing interconnection — particularly in ERCOT, where the queue moves faster — traded at a persistent premium to pure hashrate peers. The market was pricing the contract, not the coins.

This is the same forensic signature I found reconstructing the 2017 crowdsales. The narrative said decentralized community. The ledger said 68% of early holders were interconnected entities. Story and structure diverged. Structure won. It always does.

Let me be precise about what is being repriced. Three things, and only three.

First, the power purchase agreement. A fixed-price PPA signed in 2021 at $25–30/MWh is deeply in the money when spot power in constrained regions clears multiples higher. That contract carries a mark-to-market value that never appears on a mining income statement.

Second, the interconnection right. In a queue that takes five years, an already-approved 300 MW interconnection is not an operating expense. It is a moat. It can be monetized, transferred, or used to host someone else's GPUs at a hosting margin. The queue itself is becoming a secondary market in energy rights.

Third, the curtailment option. AI inference loads want to run constantly. Miners are trained to curtail. A hybrid site — hashing when power is cheap, hosting AI when power is scarce — is an optionality structure, not a single business.

A forensic note on detection. The transition is verifiable if you know where to look. Three disclosures matter. Contracted megawatts versus energized megawatts — the gap is the latent capacity. Hosting revenue as a percentage of total revenue — the crossover point is where the multiple reclassifies. And power cost per MWh locked versus spot — the spread is the embedded margin. I track all three across the disclosed fleet quarterly. The operators whose hosting share crossed 30% did not announce a strategy change. The disclosure preceded the narrative. It always does.

A miner sitting on 500 MW of contracted, energized, curtailment-flexible power is not a Bitcoin proxy anymore. It is a scarce energy asset with an optional hashing overlay. The market has begun to price it that way, unevenly, which is exactly where inefficiency lives.

s silence.

Where the narrative overreaches

Here is where I stop agreeing with the crowd.

The 32 GW number is being used as an anchor. A warning became a valuation input. Every nuclear SMR concept, every transformer manufacturer, every AI-power thematic fund now cites a figure that has never been publicly stress-tested. I have audited models built on softer assumptions. The number is not wrong. It is unverified. Those are different things, and this market conflates them constantly.

Correlation is not causation, and a projected gap is not a realized shortfall.

Consider the efficiency hedge the narrative ignores. Model sparsity, quantization, mixture-of-experts routing, and inference optimization all reduce joules per token. Liquid cooling pushes PUE below 1.1. Rack density has already climbed from 10–15 kW to 40–130 kW — the GB200 NVL72 draws around 120 kW per rack — which sounds like more demand, but it is also the precondition for cooling efficiency that legacy air-cooled halls never reached. A rigid gap assumes demand grows on a fixed efficiency curve. It does not. If tokens-per-joule improves threefold over four years, the gap compresses materially.

The GPU constraint is also loosening even as the power constraint tightens. Supply of accelerated compute has normalized from the 2023–2024 scramble. That asymmetry matters. When chips were scarce, power was not the binding constraint. Now the ordering has flipped, and the market is still pricing the old bottleneck.

Consider also that 32 GW may already be net of nothing. If it does not subtract capacity under construction — PJM additions, nuclear life extensions, new gas — it overstates the shortfall. Morgan Stanley is a sell-side shop. Its research covers the utilities and electrical equipment it writes about. That does not make the number false. It makes it interested.

There is a second layer of interest. The outlet that carried the warning is a crypto publication. It did not run this because it cares about transformers. It ran it because the implied trade is miners pivoting to AI hosting. The article is a signal dressed as a warning. Two interested parties stacked on one number.

s silence.

The dimension the coverage skipped

There is a strategic layer the report omitted entirely, and it is the one I would put at the top if I were writing it.

Power constraints are relative. China added generation capacity at hundreds of gigawatts per year across 2023 and 2024, dominated by wind and solar, with grid investment intensity the U.S. does not match. America's new-build pipeline is throttled by permitting, environmental review, and community resistance. The comparative speed of grid construction is quietly becoming an AI variable — one that can partially offset a chip disadvantage.

That reframes the 32 GW gap. It is not only an American infrastructure problem. It is a relative-positioning problem, and relative problems do not get solved by domestic headlines.

Then there is the community layer, which the report touched in a single word. Moratoriums. Data centers raise residential rates, consume millions of gallons of water for evaporative cooling, generate noise, and take tax incentives while producing limited local employment. County-level pause orders are already appearing in Virginia, Georgia, and Ohio. That is a regulatory risk that lives below the federal level, moves faster, and carries a veto. A power gap can be financed. A county board cannot.

What I actually watch

I do not trade the 32 GW figure. I trade the observable proxies. The interconnection queue. The capacity auction clearing price. The turbine order backlog. The miner's contracted megawatts versus its energized megawatts.

Here is my pre-mortem, the failure conditions stated in advance. The thesis breaks if any of three things happen. One: interconnection reform — FERC action, fast-track approvals — compresses the queue and destroys the scarcity premium on existing rights. Two: efficiency gains outrun the demand curve, so the gap never materializes at the forecast scale. Three: local moratoriums spread from Virginia and Georgia into the core ERCOT and mid-continent siting zones, converting a power problem into a political one.

I am watching the next PJM capacity auction. If it clears softer than the prior cycle, the scarcity premium is already in the price. Watch the spread between the clearing price and the cost of new entry. When that spread compresses, the premium is decaying. That is the leading indicator, and it prints quarterly.

The gap is real. The number is a rumor. The difference between the two is where money is made and lost.

Logic is the only audit that never expires.