Over the past two weeks, a basket of North American miners that have rebranded themselves as AI hosts has given back roughly a fifth of its autumn gains, while the semiconductor complex that supposedly drives that demand has held its bid. Nothing in the earnings calendar explains the split. Nothing in the hashrate prints explains it either. What explains it is a single clause buried in a sell-side note that most of the market read as a headline and discarded as a mood: only about 35% of United States data center projects currently under development are considered credible.
We have spent three years pricing electricity as though it were software. The AI trade taught us that compute is fungible, that capacity scales with capital, that a gigawatt announcement is a balance sheet item. It is none of those things. A gigawatt is a queue position, a transformer order, a county zoning hearing, a turbine slot three years out. The most important number in the AI-crypto complex right now is not a hashrate or a token supply; it is a credibility ratio, and it is telling us that the binding constraint has migrated from the wallet to the wire.
The provenance of that number matters more than the number itself, and I want to be forensic about it, because I have been burned by clean-looking ratios before.
The chain runs like this: a research desk publishes a sector note; a crypto outlet summarizes it; the summary reaches most readers as three sentences and one percentage. By the time it reaches me, the sample frame is gone, the time window is gone, the definition of the operative adjective is gone, and โ critically โ any per-name rating changes are gone. What survives is a mood with a decimal point.
A statistic without a denominator is a rumor with better typography. The figure has no stated denominator: is it all US data center projects, or only those above a certain megawatt threshold, or only AI-specific builds, or only those announced since a particular date? It has no stated criterion: does credibility mean a signed power purchase agreement, an executed interconnection agreement, a named tenant with a deposit, or merely a plausible construction schedule? Those are four different populations, and the ratio could be 35% under one definition and 70% under another. That ambiguity is not a footnote. It is the difference between a sector-wide solvency problem and a routine dispersion of project quality.
There is a third problem, subtler and more consequential. The note reportedly frames the causality as delay creating pressure for a moratorium โ projects slip, residents and utilities grow restive, local governments freeze approvals. From everything I know about American land-use politics, the arrow usually points the other way. Counties and municipalities respond to substation congestion, water draw, and noise complaints by issuing moratoriums first; the delays are the visible symptom of a decision already made. If that inversion holds, the delay is not a market phenomenon that capital can shorten. It is a political one, with a policy half-life measured in election cycles.
So let me set the frame properly. This is not a blockchain story. There is no protocol here, no consensus mechanism, no smart contract, no cryptographic assumption to evaluate. Anyone who tries to slot this into the ZK-versus-optimistic-rollup debate, or the modular-versus-monolithic debate, is committing a category error so total that it becomes a tell about how they think. What we are auditing is an engineering delivery problem: interconnection queues at PJM, ERCOT, and MISO that in the tightest regions run to multiple years; lead times on large power transformers, heavy-frame gas turbines, and high-voltage switchgear that have stretched toward the end of the decade; zoning and environmental review; a shortage of electricians and mechanical trades; and the retrofit cycle of moving from air cooling to direct-to-chip or immersion, which is not a firmware update but a construction project with its own supply chain.
We are auditing the ghost in the machine's soul, and finding that the ghost needs a substation.
Start with the valuation mechanics, because that is where the damage actually lands.
For two years, the equity story for listed miners has rested on an option: we own energized megawatts, AI tenants need energized megawatts, therefore our megawatts should be valued at AI infrastructure multiples rather than bitcoin-mining multiples. The option was never priced at face value โ good analysts haircut it โ but the haircut was implicit, informal, and almost always applied to the timeline rather than to the probability of conversion. Announcements were capitalized at something close to full pipeline value, discounted only for time.
A credibility ratio of roughly one in three changes the shape of that option, not just its tenor. When the probability of conversion collapses, the value of the underlying optionality does not decline linearly โ it decays toward the intrinsic value of the assets you already control. For a miner with signed hosting contracts, delivered racks, and an energized substation, intrinsic value is high and the option was a bonus. For a miner with a press release, a land option, and a queue position, intrinsic value is close to zero and the option was the entire thesis. Those two companies have been trading within a multiple of each other. They should not be.
The arithmetic is crude but clarifying. Take a developer with 1.2 gigawatts announced. At consensus AI-infrastructure multiples of, say, $3 million per megawatt of energized capacity, the implied enterprise value of that pipeline is $3.6 billion. Apply a 35% credibility factor and the deliverable capacity falls to 420 megawatts, or roughly $1.26 billion โ a 65% haircut on the capacity story before you touch a single discount rate. The precise multiple matters far less than the direction: the market has been capitalizing announcements, and announcements are the one input in this model that has just been repriced.
I learned this lesson in the ugliest possible way. When FTX came apart, I spent three weeks rebuilding Alameda's cross-collateralization structure from on-chain traces and exchange flows, trying to reconcile a balance sheet that had been presented as solvent. The discrepancy I eventually isolated was roughly $1.2 billion in unallocated stablecoin reserves โ not stolen, not lost, simply unaccounted for, a hole in the ledger that everyone had agreed not to look at because the headline numbers were large and the counterparties were reputable. I came out of that exercise with a permanent reflex: I do not trust a ratio until I have seen its denominator, and I do not trust a pipeline until I have seen which specific links have been signed. This figure triggers both reflexes at once. It is directionally informative and operationally useless, which is a strange and dangerous combination โ informative enough to move a narrative, useless enough that no one can be held to it.
So let me do what the summary did not: decompose the pipeline into the links that actually gate delivery.
The first gate is interconnection. A large load cannot simply arrive; it must pass through a transmission operator's study process, and those queues have lengthened dramatically as load growth from data centers, electrification, and reshored manufacturing collide with a grid built for flat demand. A queue position is not a contract. Projects withdraw, restudy, and requeue; historical withdrawal rates in some queues have been high enough to make the headline pipeline a poor proxy for deliverable capacity. If the credibility factor was derived from a sample dominated by announced-but-unsigned gigawatt projects, then what it is really measuring is the gap between press releases and paperwork โ and puncturing that gap is a service, not a scandal.
The second gate is equipment. Transformers and turbines are not commodities you order when the concrete is poured; they are long-lead items with allocations and backlogs. A project can be fully financed, fully permitted, and still sit idle waiting for a transformer. This is the part of the bottleneck that most annoys me, because it exposes how badly this industry confuses capital with capability. Money does not compress a factory's cycle time. Money does not train a journeyman electrician in a quarter.
The third gate is local permission, and this is where the moratorium risk lives. It is also where the risk is most geographically fragmented. The same developer can be frozen in one county and welcomed in the next; a state-level transmission policy can help one site and strand another. Sector-level statistics cannot substitute for site-level diligence, and the single most common analytical error in this trade is applying a uniform haircut across names that sit in completely different regulatory jurisdictions.
The fourth gate is thermal. Density is rising faster than air can carry heat away, which pushes operators toward liquid cooling. That transition is happening, and it is real, but it is also a construction and supply-chain event with its own queue. Here I will confess a bias born of watching a different kind of operator bleed. I have spent the better part of a year arguing that ZK proving costs remain absurd relative to the fees they can capture, and that unless gas returns to bull-market levels the operators running those systems are quietly subsidizing their own users. The pattern rhymes. In both cases the visible product looks healthy โ proofs verify, racks fill โ while the invisible cost line eats the operator alive. In AI hosting the invisible cost is the retrofit plus the idle capital during the wait. In rollups it is the prover bill. Neither shows up in the metric everyone quotes.
Now the part that almost nobody has priced, and the part I think is the actual trade.
If new capacity is credibility-constrained, then existing energized capacity is a scarcity asset. Grid capacity is zero-sum in a way that token supply never is. A substation with headroom, an interconnection agreement already executed, a power purchase agreement already signed, a county that has already granted its approvals โ these are not reproducible by competitors at any price within a useful time horizon. Every quarter a new build slips, the relative value of the already-energized site rises. This is the reverse inference of the whole story, and it is the one I would build a position around.
It also reframes the retrofit path. If new construction is gated by interconnection and equipment, the cheapest route to AI capacity is not greenfield โ it is converting an existing mine that already has power. The friction is lower by an order of magnitude. That advantages operators with large, already-energized footprints and disadvantages the pure-announcement developers whose entire value proposition was that they would build.
There is an upstream consequence too. When the binding constraint is equipment and interconnection, pricing power migrates to whoever supplies those things: transformer manufacturers, turbine makers, EPC contractors, transmission and distribution utilities. The AI trade, properly understood, is partly a utilities trade wearing a technology costume. I find that less exciting than the consensus story and considerably more durable.
I want to flag one more layer, because it connects to work I have done and because it is where the crypto-native reader is most likely to overreach. In 2025 I built a liquidity model around BlackRock's BUIDL fund integrating with Ethereum Layer 2s, and I quantified how tokenized real-world assets compressed settlement times โ the figure I kept landing on was a 94% reduction versus traditional rails, with compliance intact. It was a genuinely interesting result and I believed it. But the honest footnote, the one I have written about repeatedly since, is that traditional institutions do not need a public chain to settle; they need a legal wrapper and a database. Tokenization is a distribution story, not a settlement revolution, and three years of RWA storytelling has not changed that.
The same discipline applies here. AI tenants do not need your token, your chain, or your narrative. They need a substation, a signed interconnection agreement, and a cooling loop that works. Any project whose pitch is compute-marketplace abstraction without physical delivery evidence should be treated as a wrapper around an asset it does not control.
And this is where my own priors get uncomfortable. Last year I analyzed a dataset of roughly ten million transactions between autonomous AI agents and found that about 60% of them executed with no human in the loop. I wrote then that a machine economy was forming beneath the human one, and I still think that is right. But the machine economy does not float above the physical world; it is denominated in it. Every one of those agent-to-agent micropayments ultimately resolves into a watt, a rack, a cooling ton. The more autonomous the software layer becomes, the more brutally physical the constraint layer gets. That asymmetry โ infinitely scalable agents, stubbornly finite electrons โ is the real structural story of this cycle, and this credibility number is its first honest measurement.
The consensus read is that a credibility gap is bearish for the AI-crypto complex. I think that read is lazy in one direction and dangerously incomplete in the other.
Lazy, because the demand side has not weakened. Inference demand is real, training demand is real, and the backlog of tenants willing to sign is not shrinking. What has weakened is the market's assumption about the speed of supply. A trade built on the premise that demand is infinite and therefore every megawatt converts was always going to be tested; the test is now here, and it will produce winners, not a funeral.
Incomplete, because the causality inversion changes the duration of the risk. If delays cause moratoriums, then better financing and better project management can shorten them. If moratoriums cause delays, no amount of capital helps, and the risk has a policy half-life. My read, from watching how American local government actually behaves when a substation queue backs up and residents start showing up at planning meetings, is that the second version dominates. Moratoriums are contagious, too โ once one jurisdiction demonstrates that a freeze is politically survivable, neighbors copy it. That is the tail risk nobody is modeling: not delay, but cancellation. Delay retains option value. Cancellation zeroes it.
And here is the blind spot I hold against my own side of the trade. Everyone in this market is modeling demand. Almost nobody is modeling permission. We have built elaborate frameworks for token supply schedules, unlock cliffs, and emissions curves, and we have almost no framework for the one input that is genuinely non-elastic: the willingness of a county board to approve a 300-megawatt load. The scarcest asset in this cycle is not a token, a chain, or a model. It is an approval.
Finally, the honest contrarian point about the research itself. The most valuable artifact in that note was almost certainly the sample list โ which specific projects were judged not credible, and why. That is the part that got stripped in transmission. Without names, there is no trade here, only a mood. What we received is a signal about a distribution, and distributions do not have tickers. The two-stage adjustment this implies is worth naming: valuation gets repriced first, logic gets repriced second. We are in the first stage. The second arrives when operators start guiding down contracted capacity rather than announcing more of it.
The market is sideways, and chop is for positioning, not for conviction. So position on evidence, not on narrative.
Three signals will tell you whether this ratio is a warning or a eulogy. Watch large-load interconnection dockets at the major transmission operators โ the queue is the pipeline, and withdrawals are the tell. Watch power purchase agreement filings and county-level ordinance calendars, because permission precedes capital in this cycle. And watch the spread between operators reporting delivered racks and contracted revenue, and operators reporting announced capacity โ that spread is the pair trade this note implies, long delivery against short announcement, and it will widen before it narrows.
I have argued elsewhere that by 2030 something close to 40% of global GDP will be governed by algorithmic monetary policy embedded in central bank infrastructure. I still believe that, and I would add a corollary I did not have three years ago: the sovereign algorithm will run on the same electrons as everything else. The ledger bleeds red when trust decays into code, and it bleeds brownouts when the code forgets it needs a grid. The question for the next eighteen months is not whether compute demand is real. It is whether this industry can survive the discovery that its most valuable asset is a piece of paper signed by a county clerk.