Reading the room in a room of power contracts, I keep returning to a single number that refuses to resolve. Thirty-two gigawatts. Morgan Stanley says America's AI data centers will face a power gap of that size by 2028. The figure lands somewhere between prophecy and PR, and the difference matters enormously for anyone positioning capital on a flat market that has stopped paying for narrative alone.
I've spent the past year auditing on-chain data flows and institutional research notes side by side for a Tallinn consultancy, and the pattern I keep encountering is this: the most powerful numbers are the ones with the thinnest methodology. Thirty-two gigawatts is a directional truth wrapped in a quantitative bluff. Let me show you why, and why the real trade isn't in the headline at all.
The directional claim is sound. AI's constraint is migrating from silicon to electricity. That much is now consensus, corroborated across independent sources.
Lawrence Berkeley National Laboratory reported U.S. data center consumption rose from roughly 76 TWh in 2018 to about 176 TWh in 2023 — around 4.4% of national electricity — and projects 325 to 580 TWh by 2028, or 6.7% to 12%. The IEA's Electricity 2024 sees global data center demand climbing from roughly 460 TWh in 2022 toward potentially over 1,000 TWh by 2026. Different institutions, different methodologies, same trajectory. The demand explosion is real; the specific gap is not, or at least not yet provable.
Where does 32 GW come from? Morgan Stanley does not say publicly. Is it a demand-side shortfall? A grid interconnection gap? Peak-load capacity? Does it net out projects already under construction? Does it include non-AI data centers? These aren't pedantic questions. They determine whether 32 GW represents a five-hundred-billion-dollar capex hole or simply a reallocation problem the industry solves by 2027.
The physics underneath is less ambiguous. Interconnection queues at PJM, ERCOT, and MISO are backlogged for four to seven years in some regions. Transformer lead times stretched from months pre-pandemic to three to five years. Gas turbine order books at GE Vernova and Siemens Energy stretch into 2027 and 2028. Rack power density jumped from 10–15 kW to 40–130 kW for AI training clusters; an Nvidia GB200 NVL72 cabinet draws roughly 120 kW, and a single Blackwell B200 chip burns over 1,000 watts. The bottleneck is not generation capacity in the abstract — it's grid-connectable capacity, and the queue is the queue.
That distinction reshapes the entire investment map. I've been running Transformer Lead Time and Interconnection Queue Depth as two of my attention indicators for months. Both point the same way: the scarce asset isn't electricity, it's the legal right to connect to the grid at a known price and date. That right is tradable, underappreciated, and often buried in old bitcoin mining contracts.
In 2022, during the modular blockchain research binge that became my Substack, I learned to separate settlement layers from execution layers when everyone else was still arguing about blockchains as monolithic things. The lesson transfers. AI's execution layer — compute — is being constrained by its settlement layer — power contracts and interconnection rights. The value is migrating to the settlement layer, where the queues are.
Look at the bitcoin miners first. They've been dismissed for two years as failed ASIC farms with bad balance sheets. But Core Scientific, Applied Digital, IREN and their peers hold something the hyperscalers would pay billions to acquire: existing power purchase agreements and queued interconnection capacity. When Microsoft signs a nuclear PPA with Constellation at Three Mile Island and Amazon contracts with Talen, the market registers a headline. When a bitcoin miner converts a 500 MW site into an HPC hosting facility, the market yawns. That asymmetry is the trade Crypto Briefing is quietly pointing at without ever naming it. The crypto media angle here is not incidental; it's the buried lede.
The miners become power-contract arbitrageurs in an environment where every additional megawatt of queued capacity trades at a premium. This is not speculation. It's a repricing of assets that were mispriced because nobody wanted to model the energy leg of the business.
Now the contrarian angle, and it's the one I think most market participants are getting wrong. The consensus reading of the 32 GW warning is "buy utilities, buy nuclear, buy electrical equipment." Utilities Select Sector SPDR outperformed the broader market in 2024 for the first time in years. SMR concept stocks and nuclear names like Constellation and Vistra have re-rated hard. Eaton, Vertiv, GE Vernova have all run.
But here's what I observed during the NFT cycle in 2021, and it's the analytical habit that has served me best since: when a narrative gets quantified with a single proprietary number, that number doesn't get stress-tested — it gets amplified. The 32 GW figure is a gift to thematic capital allocators. It anchors pitch decks. It justifies valuations that no bottom-up model would support. When a number is produced by an investment bank whose clients are also trading the theme, the number's warning function and its marketing function are indistinguishable.
I've watched this exact dynamic before. In 2021, the JPEG-to-access-key narrative was validated by price, then price stopped validating anything, and the narrative kept running for six months after the fundamentals flipped. The equities side of AI-power is showing the same signature. Somewhere in the核电 basket, there are SMR names trading at multiples that assume the 32 GW gap is factual rather than forecast. When the actual interconnection data lands — and PJM publishes queue reforms, and GE Vernova reports order books with specific delivery dates — there will be a split between names with signed offtake and names with signed press releases. That split will be brutal and it will be fast.
I don't claim to know which specific tickers are mispriced. I've never been good at that, and the ENFP tendency to chase five theses at once is honestly a liability in a positioning market. But I know the analytical signature of a narrative that has gotten ahead of its evidence, and I can recognize it here with the same clarity I recognized the NFT identity-marker shift in 2021 before the correction.
The deeper structural point is that the market is pricing AI-power into a rigid constraint frame. It isn't rigid. Efficiency hedges exist. Model sparsity, quantization, MoE architectures, liquid cooling pushing PUE below 1.1, and DC 800V/HVDC architectures all compress the joules-per-token denominator. The 32 GW gap assumes static efficiency. Any credible forecast of an eight-gigawatt annual gap must specify an efficiency trajectory, and Morgan Stanley does not.
Then there's the variable the report completely omits, and it's the one I think about most because it's the one I can't hedge. China added hundreds of gigawatts of generation capacity annually across 2023 and 2024, mostly solar and wind, with a grid investment intensity the U.S. permit and environmental review regime cannot match. The 32 GW gap is a U.S.-specific number in a global AI competition. If electricity is the new constraint, and one competitor can build the power plants faster, then the 32 GW figure is not a warning about AI's future — it's a warning about which country gets to build it. That's a strategic framing absent from every sell-side note I've read, and I suspect it's not an oversight but a scope choice.
My audit experience keeps pulling me back to a simple operating principle. Since 2020, when I was verifying Zcash zk-proofs in Python late at night at Tartu and learning that the deepest narratives hide in verification, I've tried to read market structure through the same lens: follow the incentives, not the announcement. The 32 GW warning has a user. That user isn't retail investors.
The next twelve months of monitoring signals are specific. PJM and ERCOT capacity auction results, quarterly. GE Vernova and Siemens Energy turbine delivery schedules, quarterly. Local moratorium filings in Virginia, Georgia, Ohio — event-driven. Hyperscaler capex guidance and disclosed PPA volumes. Nuclear PPA announcements from Constellation, Talen, Vistra. And critically, power contract and interconnection queue values embedded inside public bitcoin miner filings — the most under-covered dataset in the entire AI-power complex.
On a sideways market, chop is for positioning. The genuine insight is not that AI needs power; everyone knows that. The genuine insight is that power contracts, not power plants, are the binding constraint, and the entities holding queue positions — including the miners everyone stopped watching — are sitting on the only assets the 32 GW gap makes genuinely scarce.
Whether the gap turns out to be 32, or 20, or 45 gigawatts, the assets don't change. The queues don't clear faster because a number is questioned. The only question is who owns the queue, and whether the market prices that ownership before the interconnection data forces it.
I don't know the answer yet. But I know which filings I'm reading this weekend, and which order books I'm tracking, and I trust that the closer I get to the primary evidence, the less the 32 GW ghost will haunt the positions I build.
The gap is not the signal. The queue is.
