The number arrived clean. Thirty-five percent. Only 35% of US data center projects are "credible." A sell-side note said so. A crypto outlet repeated it. And not one version of the claim โ not the note, not the summary, not the thread that landed in my feed โ told you what the other 65% were, or what the word "credible" was supposed to mean.
I went looking for the denominator. It was not there.
No sample size. No time window. No definition of the word doing all the work. No list of projects. Just a percentage, floating in a vacuum, dressed up as a finding. The code spoke, but the metadata lied โ except here there was no code at all, only a headline that had already been laundered through three layers of translation before it reached my screen.
That is the part worth writing about. Not the delay. The fact that an industry preparing to spend hundreds of billions on physical infrastructure cannot produce a number with a denominator attached to it โ and expects to be believed anyway.
For two years, the dominant trade in crypto-adjacent equity has been the miner-to-HPC pivot. The logic was seductive and simple. Bitcoin miners own land, substations, and โ critically โ grid interconnection capacity. AI needs exactly those things. Therefore, the market reasoned, every megawatt a miner controls is a call option on the AI buildout and should be valued at AI-infrastructure multiples rather than bitcoin-mining multiples.
That repricing happened. It is still happening. Companies that spent a decade being valued on hash rate and electricity spreads suddenly found themselves discussed in the same breath as hyperscalers. EV per megawatt became the metric of the moment. Announcements became pipelines. Pipelines became forecasts. Forecasts became, somehow, valuations.
Into this, Bernstein published research suggesting that only 35% of US data center projects are "credible," that significant delays are underway, and that those delays could threaten AI infrastructure growth โ possibly even prompting local moratoriums.
The crypto press picked it up. The framing wrote itself: delays are coming, the AI buildout is stalling, the miners' option value is impaired.
Maybe. But before anyone trades that conclusion, they should notice what is missing from the version of the report that actually circulated. Which is almost everything.
Start with the word doing all the work. What does "credible" mean? It could mean a signed power purchase agreement. It could mean an executed interconnection agreement. It could mean a firm tenant commitment. It could mean equipment on order. It could mean nothing more than a construction start date on a slide. Each of those definitions produces a wildly different 35%.
If credible means signed PPA plus approved interconnection, then 35% is a catastrophe โ it means two-thirds of the announced book cannot be financed into existence. If credible means "has a named tenant who expressed interest," then 35% is a rounding error on normal development attrition, and the note is telling you nothing you did not already know about how real estate gets built. Without the definition, the number cannot be modeled. Garbage in, permanence out โ and the permanence here is the valuation multiple someone just derived from it.
Then the universe. Thirty-five percent of what? Global data center projects? US only? AI-only? Only builds above 100 megawatts? The circulating summary says none of this. A percentage without a denominator is not a statistic. It is a mood with a decimal point.
Here is what I do know, from watching infrastructure from the inside rather than from a deck. The delays are real. They are also not financial. They are physical and administrative, and that changes everything about how long they last.
Five chokepoints, in the order they bite. First, the interconnection queue: large loads must clear transmission-operator processes โ PJM, ERCOT, MISO โ and in several regions that queue is measured in years and has worsened as load growth exploded. Second, equipment lead times: heavy-duty gas turbines, large power transformers, and high-voltage switchgear are backordered years out; some categories are effectively allocated. Third, permitting and zoning: rezonings, environmental review, water-use permits. Fourth, skilled labor: electricians and HVAC crews are scarce and booked. Fifth, the cooling retrofit: moving from air to direct-to-chip or immersion liquid cooling adds its own supply-chain and construction cycle.
Capital is abundant. Time is not purchasable. You cannot wire money to a transformer factory and get a transformer tomorrow. The 35% is not a capital problem. It is a physics-and-permits problem, and that is precisely why it will not clear on the next rate cut or the next funding round.
The causality is the tell. The report frames it as: delays may prompt moratoriums. I would flip the arrow. In actual US local zoning practice, the more common chain runs the other way โ residents and grid stress build, a local government imposes a moratorium, approvals freeze, and projects that were moving suddenly appear "delayed." If that is the true direction, the delay is policy-rigid, not capital-elastic. It does not respond to a better balance sheet. That distinction is the difference between a risk that lasts a quarter and one that lasts three years.
Now the implication that actually matters, and the one the headline buries. The problem is not a bad quarter of slippage. It is a haircut. If the market begins pricing announced capacity at 35% instead of 100%, every EV/MW multiple built on the full pipeline gets marked down โ slowly, systematically, across the whole announced gigawatt book. This is the Terra lesson applied to equity. When I mapped the UST wallet clusters in 2022, the failure was not in the headline โ it was in the stake concentration nobody disclosed. Same pattern. The pipeline is the headline. The signed PPAs are the flows. Trace the flows.
But โ and this is where the lazy short thesis breaks โ a sector statistic is not a project verdict. The 35% strengthens divergence, not uniform decline. On one side: projects with executed long-term hosting contracts, delivered racks, energized capacity. On the other: projects with a press release and a rendering. The gap between those two will widen violently, and it will widen regardless of what AI demand does next year. Applying a sector-level number to a single name is a category error. Statistics are not due diligence.
The reverse inference is the one the market keeps missing. If new capacity cannot get power, existing energized capacity becomes scarce. This is the same structure I documented in the NFT metadata investigation: the token said "owned," but the asset lived on a server someone else could switch off. Owning a pipeline announcement is not owning a powered site. A miner with a substation, an interconnection agreement, and spare capacity holds something a new entrant literally cannot replicate โ because grid capacity is zero-sum. The scarcity premium sits in the physical, permitted, energized asset, not in the deck. When a moratorium lands, the game shifts from competitive acquisition to allocation of what already exists, and the pricing power moves decisively to whoever already holds the interconnect.
The value migrates upstream too. If the bottleneck is equipment lead time and grid access, pricing power flows to the supply side: transformer makers, turbine manufacturers, EPC contractors, transmission and distribution utilities. The picks-and-shovels of the picks-and-shovels. Meanwhile the miner "AI option" โ that EV/MW premium โ faces compression precisely where it was most speculative, in the names that announced the most and delivered the least.
And now the part that should irritate any careful reader. The version of the report that reached the public stripped out the only actionable content: the sample, the criteria, the project list, and โ most importantly โ any associated stock ratings or target changes. Sell-side warning notes almost always come with rating adjustments. Those were removed. What survived was an emotional headline. The single most price-relevant component was surgically deleted, leaving a mood with a number attached. I don't trade headlines. I audit them. And this one fails the audit at the source-chain level, before we even get to the data.
None of which means the bulls are wrong. It means they are early and imprecise.
The demand is real. AI compute consumption is not a narrative โ it is measured in signed contracts and power purchase agreements from hyperscalers who do not sign for fun. The delay does not kill the thesis. It re-sorts the winners. The bulls are right that the physical layer โ power, land, interconnect, cooling โ is the scarcest asset in the entire stack. Where they drift is in assuming that scarcity accrues evenly, that every megawatt is equal, and that the timeline is theirs to command. It is not. The demand curve did not bend. The supply curve did. That is a positioning problem, not a directional one. In a market like this, chop is for positioning โ and the 35% is a positioning signal dressed up as bad news. The correct response is not to short the sector. It is to separate the energized from the announced, and to stop paying AI multiples for a PDF.
The 35% is not a forecast. It is a mirror. It shows how much of the AI-infrastructure story is actually plugged in and how much is a rendering. The question every holder should be asking is not "will AI demand grow" โ it will โ but "how much of my valuation is a signed PPA, and how much is a press release?" Until someone publishes the denominator, the number is noise. And noise, in a market like this, is exactly where the losses hide. Volatility is the product; loss is the feature. The only defense is the one thing the report never gave us: the list.