The Number Without a Denominator
The number that should have stopped the conversation arrived without a footnote: two hundred and twenty-eight billion dollars, attached to a category of AI data center projects reported as blocked, delayed, or otherwise stalled by public opposition. It was carried by a vertical crypto outlet, aggregated in the way vertical outlets aggregate โ quickly, plausibly, and thin on provenance.
I read it four times, hunting for the denominator. Was this planned total investment? Committed capital? The notional value of projects that will now never be built? There is no time stamp on the disturbance, no geography, no company names, no distinction between a project in litigation and a project that merely received a hostile public comment period. A number that large with that little scaffolding is not data. It is a mood, expressed in dollars.
So I am going to do something unfashionable. I am going to treat it as a directional signal and flag every inference I make from it. The number is unverified. The shape it describes is not. And the shape โ capital meeting a constraint it cannot engineer, buy, or lobby away โ is the single most important thing happening to the compute narrative right now, in AI and in crypto, because they are no longer separate stories.
To hunt the truth, one must first bury the hype. This one is buried deep.
What The Number Actually Says
Strip the dollar figure away and read the sentence underneath it. A set of very large, very capital-intensive physical facilities, whose entire economic purpose is to convert electricity into computation, has encountered resistance that is not technical, not financial, and not regulatory in the ordinary sense. It is social. Neighbors. Municipal boards. Ratepayer advocates. People who live near the substation.
That is new. For most of the last three years, the constraint on AI infrastructure has been described in the language of supply chains โ GPUs, HBM stacks, advanced packaging, export controls. Every one of those constraints is a problem you solve with money, time, and engineering talent. You queue for the chips, you pay the premium, you build the fab. Expensive, slow, but legible.
Permission is a different category of problem. It does not have a price, because the people withholding it are not selling.
There is a second thing buried in that headline. The projects are described as a group, which implies multiplicity, which implies a competitive structure. This is not one company's bad luck with one county board. This is a race โ the article's own word โ in which several large balance sheets are simultaneously pursuing the same scarce physical inputs: grid capacity, water, land with the right fiber and the right zoning, and the tolerance of the people who already live there.
When several players compete for a scarce input, the input reprices. That is the whole game. Everything I write below is downstream of that single mechanism.
I want to be honest about my own position here. My training is not in electrical engineering and it is not in municipal governance. It is in narrative and in capital structure, which means I am qualified to tell you how this reprices and unqualified to tell you how many megawatts a specific campus needs. So I have kept the technical claims to what is broadly established across the industry and labeled the rest as inference. If you want the number verified, call the utility. If you want to know what the market will do with it, keep reading.
Why I Recognize This Shape
Every cycle produces a constraint that the industry refuses to name, because naming it would mean admitting that the previous cycle's capital was misallocated. I have watched this happen four times, and I have been wrong in the same direction each time, which is the only reason I now recognize it.
To hunt the truth, one must first bury the hype โ and I have personally buried quite a lot of it.
In 2017, at thirty-three, I was working through the ICO boom out of Barcelona with a spreadsheet and a growing suspicion. I read something like fifty whitepapers that year, and the pattern was always the same: a genuine technical capability, wrapped in a token that had no claim on it. The utility token was the artifact. The utility was the marketing. I wrote a long, unfashionable piece about the utility token fallacy and predicted a correction for anything without a real user โ and I was right about the correction, roughly eighteen months early, which in this industry is indistinguishable from being wrong. What I learned was not that I could see the future. What I learned was that the industry habitually mistakes a capability for a solved problem, and the unsolved part is almost always human.
In 2020 I went deep into Uniswap and the yield farming complex, and I stopped writing about code and started writing about social contracts. An automated market maker is not a market. It is an agreement between strangers about how to behave when the price moves, enforced by code and subsidized by a token. The mechanism worked. The trust was the fragile part.
In 2021 I pivoted to soulbound tokens โ reputation and identity as non-transferable credentials โ because I had become convinced that the interesting frontier was not ownership of assets but proof of participation. That essay travelled further than anything I had written, which taught me something uncomfortable: the market rewards the narrative that arrives before the infrastructure, not the one that arrives with it.
In 2022 I stopped writing entirely for a while. I published something raw called The Cost of Belief about what it does to a person to hold a thesis through a ninety percent drawdown, and the response told me that the readers wanted the struggle more than they wanted the model. I stopped pretending the analysis was separable from the analyst.
And in 2025 I spent the year translating regulatory and technical structure for institutional readers โ compliant decentralization, as I framed it โ which taught me the single most useful thing I know about infrastructure narratives: institutions do not adopt your rails. They adopt your rails only when the rails are the cheapest way to satisfy a constraint they already have.
Now look at the data center story through that lens and it becomes legible. The constraint is not compute. The constraint is a site with power, water, and permission. And every crypto narrative currently being sold as "AI-adjacent" is really a claim about who gets to satisfy that constraint.
Compute Is Electricity With a Latency Budget
The first thing to internalize is that a data center is not a building full of computers. It is a machine for converting electrical power into calculation, with a thermal problem attached. That is the entire physics. Everything else โ the land, the fiber, the cooling loop, the transformer yard โ exists to serve that conversion.
This matters because it relocates the bottleneck from a place where capital has leverage to a place where it does not. A hyperscale AI campus draws power on the order of hundreds of megawatts. For scale: that is not a factory. That is a mid-sized city's worth of load, arriving on a single interconnection request, in a county that may have spent the last two decades with flat or declining demand and an aging distribution network.
Now add the queue. In most developed markets, connecting a load of that size to transmission is not a purchase; it is a study, then a network upgrade, then an agreement, then construction of shared infrastructure that the developer may or may not fully fund. Transformers of the relevant class have been on extended lead times across the industry for years. The line between a signed power agreement and energization can be measured in years, not quarters.
That queue is the real asset. Not the land. Not the shell. The scarce thing is a position in the interconnection queue at a node where the grid can actually absorb you, and a county that will not sue you for making noise at three in the morning.
Which is where the community enters, and where the economics get genuinely hard. The costs a data center imposes are real and local: strain on the local grid that can show up in rates, water consumption for cooling in places where water is contested, diesel backup generators that cycle during grid stress, visual and acoustic intrusion, and upward pressure on land values that pushes out the people who were there first. The benefits are real but diffuse: some construction jobs, a small permanent staff, tax revenue that frequently arrives discounted through abatement agreements.
The asymmetry is the whole story. The cost is concentrated on the people who live there; the benefit is concentrated on shareholders and remote users who will never see the site. Industry data on direct employment per megawatt is unflattering โ a modern AI campus employs a fraction of what a traditional industrial facility of comparable land area employs. That gap is not a communications problem. It is a structural grievance, and it will express itself whether or not anyone writes a press release about it.
I have one relevant piece of first-hand experience. In the course of auditing mining operations for a client in 2023 โ siting, load profiles, host agreements โ I sat through two municipal sessions in jurisdictions where a facility was being proposed. I went in expecting procedural opposition and found something more durable: a specific, well-researched list of concerns, presented by residents who had done more homework than the developer's presentation had. The developer had brought renderings. The residents had brought the substation's load history. That asymmetry told me more about the next five years of infrastructure than any quarterly report I have read since.
The Halving Ledger, and Who Owns the Interconnection
Here is where this stops being an AI story and becomes a crypto story, and it starts with arithmetic that nobody disputes.
The April 2024 halving cut the block subsidy to 3.125 bitcoin. Miner revenue per unit of hash collapsed, and it did not recover through price alone for most of the fleet. The cost curve did what cost curves do: it squeezed out the operators with expensive power and old machines, and it rewarded the ones with the cheapest electrons and the best capital access.
I have held a specific view on this for a while, and I will show it rather than state it. Watch what the surviving public miners did with their balance sheets in the eighteen months after the halving. They stopped describing themselves as bitcoin miners and started describing themselves as energy infrastructure companies. They signed hosting agreements. They announced HPC and AI colocation conversions. They monetized one asset that almost no AI company has: an energized site, with an interconnection agreement already executed, at a node with existing transmission headroom, in a jurisdiction that already said yes to them once.
Read that last clause again, because it is the entire thesis of this piece. In a world where permission is the binding constraint, a previously permitted site is not a consolation prize. It is a monopoly on time.
There is a cost to this pivot, and I do not think the industry has priced it honestly. The decentralization argument for proof-of-work rested, in its strongest form, on the claim that hashing power was geographically and organizationally distributed โ that no small set of operators could coordinate to censor or reorder. What we are watching is that fleet being repurposed toward workloads that are, by their nature, contracted, permissioned, and concentrated. The hardware does not disappear; it changes hands and changes purpose. The largest operators convert because only they can carry the capital cost of a conversion, and the smaller ones sell rigs or sell power contracts to whoever will buy them.
I hold no illusions about what this means. The hash rate concentrating into a handful of pools was already hollowing out the decentralization consensus long before anyone signed an AI hosting deal. The pivot does not create that concentration. It monetizes it. The consensus was already hollow; now it has a price per megawatt, and the price is high.
For an investor, the implication is uncomfortable and useful: the most direct listed exposure to the AI compute buildout may not be an AI company at all. It may be a miner with a substation in a county that has already granted its permits, retooling a shell it already built. That trade has largely been discovered. What has not been discovered โ and this is where I would look โ is the second derivative: the companies that supply the conversion. Thermal management, high-density power distribution, substation equipment, and the engineering firms that can turn a mining hall into a liquid-cooled inference hall without tearing down the slab.
The DePIN Falsification Test
Now the part of the market that will object to everything above, because it has a token that solves exactly this problem on paper.
Decentralized physical infrastructure networks โ the compute-flavored ones in particular โ have spent several years arguing that the answer to centralized hyperscale is a marketplace of distributed capacity. Idle GPUs, networked and matched to demand, priced by auction, owned by no one. I think the pitch is elegant. I also think it survives contact with the workload far less well than its market capitalization implied at the peak, and I want to give you a falsification test rather than an opinion.
Ask four questions of any decentralized compute network, and require a number, not a narrative.
What is the fabric? Frontier training and serious inference depend on high-bandwidth interconnect inside a node and a low-latency, high-throughput network between nodes. Consumer cards strung across home internet do not have that. The workloads they can serve are real โ small-model inference, rendering, batch jobs, fine-tuning of modest models โ but they are not the workloads that consume megawatts.
What is the utilization? Distributed supply that is available intermittently, at unpredictable quality, with no service-level guarantee, is not interchangeable with a contracted reservation. Utilization on these networks has historically been a fraction of nominal capacity, and nominal capacity is what gets reported.
Who eats the latency and egress cost? Moving data to a distributed pool is frequently more expensive than the compute saved, particularly for workloads with large datasets or stateful requirements. This cost is structural, not a bootstrapping problem.
And the fourth: what does the token actually hold? If the answer is "governance over a protocol that facilitates rentals," then the token is a claim on activity fees, not on capacity โ and fee-based claims on thin activity are the most brutally repriced assets in every crypto bear market I have lived through.
I am not saying the category is worthless. I am saying the category's addressable market is smaller and more specific than the token market assumed, and that in a bear market the only thing that reprices faster than a narrative is a narrative with a service-level problem. The honest version of the DePIN compute pitch is not "we beat the hyperscalers." It is "we serve the demand that is too small, too bursty, or too price-sensitive to sign a three-year reservation." That is a real business. It is just not a gigawatt business, and the valuations were built on the gigawatt version.
The Consent Ledger โ Where Identity Infrastructure Finally Has a Customer
Now let me return to something I care about more than the trade, and where I think a genuine, defensible use case has quietly appeared.
When I wrote about soulbound credentials in 2021, I made an argument I still believe: that non-transferable attestations of participation and reputation would outlast the speculative use of tokens as collectibles. What I got wrong was sequencing. I assumed the demand would come from individuals wanting portable reputation. It did not. Most people do not want a portable reputation; they want a job, a loan, or a lease, and the institution granting it does not care about your wallet.
But look at what a data center siting fight actually is. It is a negotiation over a set of specific, verifiable commitments: this much water per year, this much noise at the property line, this much tax revenue un-abated, this many local hires, this much grid capacity released on demand during peak, this much community fund, audited annually. Those commitments are currently recorded in planning documents, press releases, and sometimes nothing at all. They are almost never machine-readable, comparable across jurisdictions, or enforceable by the people who extracted them.
That is a verifiable credential problem. Not a profile picture problem.
A community benefit agreement, attested and versioned, with the obligations in a form that any resident can query, is the first application of the identity layer I have seen that has a counterparty who actually wants it. The county wants it, because it makes enforcement cheap. The developer wants it, because being able to demonstrate compliance is the cheapest way to get the next permit in the next county. The residents want it, because a promise in a press release is worth nothing and a recorded obligation with an audit trail is worth something.
And here is the honest caveat, because I have been burned by exactly this kind of enthusiasm before. On-chain attestation of an off-chain obligation is theater unless the off-chain obligation is enforceable. Recording a promise does not create a remedy. If the recording layer sits on top of a contract no court will enforce, you have built a very elegant museum of good intentions. I spent six months in 2022 learning that lesson in a different context, and I do not intend to relearn it here.
Which is why I think the interesting version is not a token. It is an attestation standard adopted by the entities that already hold enforcement power โ municipal authorities, utility regulators, and the insurers who underwrite the construction. Insurers are the ones I would watch. An insurer pricing political and permitting risk has an obvious reason to want standardized, comparable, queryable records of what a developer promised and whether it delivered. That is a customer with a budget and a loss ratio. That is not a narrative.
What Does The Token Hold?
Which brings me to the part of this market I trust least, and where I will be direct because the cost of indirection here is real money.
There is a wave of product being sold as "tokenized AI infrastructure" โ exposure to data centers, compute revenue, or energy contracts, wrapped in a digital asset. I have read a representative sample of these structures. Almost none of them give the holder a claim on the thing that matters.
Ask what the token holds. A claim on depreciation of physical equipment you cannot inspect? A revenue share on a lease with a counterparty you cannot verify? A governance right over a special purpose vehicle whose operating agreement you have not read? Or a marketing wrapper around a private credit fund that already existed, sold at a markup to people who wanted AI exposure and could not get it through a brokerage account?
Institutions do not need your public chain. They already have the rails โ the depositary, the custodian, the credit agreement, the limited partnership. What they lacked was never settlement infrastructure. It was a legally enforceable claim on a permitted asset.
That is the RWA lesson in one sentence, and the data center trade is its purest current expression. The value is in the interconnection agreement, the power purchase agreement, the water rights, the local approval, and the lease. Every one of those is a legal instrument. None of them becomes more valuable because it is represented by a token on a chain whose validator set has no jurisdiction over any of it. If anything, putting it on-chain adds a compliance surface without adding a remedy.
So the useful question when you are pitched any AI-infrastructure token is not "what chain is it on." It is: at the end of the day, what can I sue for? If the answer is vague, you are holding a story with a ticker, and stories with tickers have a known half-life.
The DA Parallel
I want to draw one structural comparison, because I think it is the cleanest way to see what is happening, and because it is a mistake I have already watched this industry make once.
For several years, a substantial amount of capital and engineering went into dedicated data availability layers for rollups. Modularity was the thesis. The argument ran that rollups would generate so much data that the settlement layer would be the bottleneck, and therefore a purpose-built DA layer would capture the value. It was a coherent argument. It was also, in practice, a bet on demand that had not been measured.
The overwhelming majority of rollups never generated data at anywhere near the volume that would justify a dedicated availability layer. The infrastructure was provisioned for a narrative trajectory. The trajectory arrived with a smaller slope, and the capacity sat there, over-engineered relative to load, chasing a market that was never as large as the diagram implied.
Now look at the data center buildout in 2026. Gigawatt-scale campuses are being permitted and financed against a demand curve that is extrapolated from a two-year training run of capability gains. The extrapolation may be right. But the infrastructure being committed is not modular โ it is steel, concrete, transformers, and water rights, and its useful life is measured in decades. Building twenty availability layers for demand that did not arrive was a painful lesson. Building gigawatt campuses for demand that cannot be permitted is the same mistake with a much heavier asset base.
I am aware of the counterargument, and it is not stupid: inference demand is growing, agentic workloads multiply it, and if the compute is not there, capability is stranded. All true. But note that the counterargument is entirely about demand. It says nothing about whether the supply can be sited. And the whole point of the two hundred and twenty-eight billion dollar figure โ whatever its true denominator โ is that the supply side just discovered a constraint that demand forecasts do not model.
The Contrarian Read: Opposition Is a Moat
The consensus interpretation of public resistance to data centers is that it is a risk to the AI buildout. Slower permitting, higher costs, delayed revenue, lower returns. That reading is correct as far as it goes, and it is the reading that will be in every research note by the end of the quarter.
Here is the reading I have not seen stated cleanly.
Public opposition does not kill the buildout. It concentrates it.
Consider who can survive a four-year permitting war. You need a balance sheet that can carry land and pre-construction costs for years with no revenue, a legal team that can litigate in multiple jurisdictions simultaneously, a government affairs function with real relationships at the county level, and the credibility to absorb a delay on a capex line that analysts are watching quarterly. That is a description of perhaps a handful of firms on the planet. It is not a description of a startup, and it is not a description of a mid-cap developer.
Every increment of friction in the permission process is a barrier to entry that the incumbents did not have to build and did not have to pay for. The opposition is doing the consolidating for them, voluntarily, without compensation. The firms with the largest existing footprints and the most executed interconnection agreements are being handed a moat by the people protesting outside their sites.
The second-order consequence is a repricing of what already exists. A campus that is energized, or near energization, in a jurisdiction that has already survived its own public process, becomes more valuable than a greenfield site with better economics on paper, precisely because the greenfield site has an unpriceable approval risk. That is the opposite of what an efficient market would do to assets in a supply-constrained industry โ and I would expect it to show up in infrastructure valuations long before it shows up in any token price.
The third-order consequence is where I would place a bet if I were placing one. If the binding constraint is consent, then the value migrates to whoever can manufacture consent at scale. Not a public relations firm โ a firm that can produce a genuine, defensible benefit-sharing structure, replicate it across jurisdictions, and demonstrate a compliance record. That is a service business with a moat that compounds with every successful site. It will never have a token. It will probably never be written about here. It is where the money quietly goes.
And the trap I would flag hardest is the one I have seen the industry walk into repeatedly: treating consent as a communications problem when it is a governance problem. A developer who runs a listening tour and then builds exactly what they planned has not negotiated. They have performed negotiation. Residents can tell the difference, and the tell is whether anything in the final design changed. A promise that changes nothing is not a concession, it is a delay โ and delay is the one thing a capital-intensive project cannot absorb.
The Takeaway
The narrative that is going to replace the plain "AI compute" story is not a better compute story. It is a permission story. And like every narrative transition in this industry, it will be legible in the prices long before it is legible in the headlines.
Watch four things, none of them glamorous. Watch whether the large operators start disclosing energization timelines rather than capex commitments, because energization is the only honest number in this sector. Watch whether public miners keep their power contracts or sell them, because the direction of that flow tells you who is buying the future. Watch whether any jurisdiction publishes a standardized template for community benefit agreements, because a template is the moment consent becomes a repeatable product. And watch, quietly, whether insurers begin demanding attestable compliance records before underwriting construction โ that is the signal that identity infrastructure has found a customer with a loss ratio instead of a customer with a whitepaper.
Everything else in this space is a claim on a PDF. The claims on substations are the ones that will settle.
So when you look at what you are holding โ the miner turned host, the compute network, the tokenized infrastructure yield, the AI-adjacent beta โ ask the only question that has ever mattered in a bear market. Does this thing sit on a site with an executed interconnection agreement, a signed community agreement, and a permit that has already survived a hostile public meeting?
Or does it sit on a diagram of one, drawn by someone who has never had to sit in the room where the neighbors brought the load history?