Somewhere inside the arithmetic of the artificial intelligence boom there is a number that refuses to resolve. Two hundred and twenty-eight billion dollars. It arrives without a denominator โ no distinction between capital already deployed, projects merely announced, permits quietly withdrawn, or value that will never exist because a county board voted against a substation on a Tuesday evening. The figure is large enough to move sentiment and vague enough to escape accountability, which is exactly why anyone who tracks capital flows should treat it with suspicion. In my work on cross-border settlement and macro liquidity, I have learned to distrust any sum that cannot be traced back to a contract. The $228 billion is not a measurement. It is a mood โ and the mood is that the physical layer of AI has collided with something capital cannot buy: consent.
For most of the past decade, the competitive logic of frontier AI rested on a single premise โ that capability scales with compute. The scaling law was never a law in the physical sense. It was an empirical regularity that held for as long as someone was willing to pay for more accelerators, more power, and more cooling. That willingness turned the data center from a mundane piece of real estate into the physical expression of a strategy. Architecture still matters; so do parameter counts. But both have increasingly become the visible surface of a deeper substrate: land, transmission capacity, water, and the tolerance of the people who live near it.
What the industry is now discovering is that this substrate does not scale the way software scales. A model can be copied at near-zero marginal cost. A five-hundred-megawatt campus cannot. It requires a grid upgrade measured in years, a water agreement measured in decades, and a community that agrees โ or at least declines to organize โ against the noise, the diesel backup generators, the transmission corridors, and the quiet upward pressure on local rents. When several hyperscalers converge on the same power-constrained region, they do not simply compete with one another. They compete with households for the same electrons. The resistance now surfacing is not a technical failure. It is a governance failure that three years of technical optimism never modeled. And at that point the story stops belonging only to AI โ because the crypto industry spent the same three years building what it claimed was an answer to precisely this question, and must now decide whether that answer is real.
The word the coverage kept returning to was a race โ and races, in capital-intensive industries, are rarely races at all. They are auctions. Only a handful of firms can underwrite a project measured in the tens of billions, which means the data-center buildout is not a broad industrial expansion but a consolidation of capacity among a small set of balance sheets. When a field narrows to a few competitors, the marginal advantage stops coming from superior engineering and starts coming from privileged access โ to power, to land, to regulators. The public backlash does not interrupt that auction. It changes who is allowed to bid.
The first discipline in reading a figure like $228 billion is to ask what it counts. Three readings are available, and they mean radically different things. It could be the aggregate planned value of a pipeline โ a theoretical number that evaporates the moment financing tightens. It could be the value of projects already under construction, where delay converts committed capital into sunk cost. Or it could be the notional value of projects cancelled outright, in which case the loss is already realized and the headline is retrospective. A pipeline, a sunk cost, and a cancellation are three different assets, and the fact that one number floats between them tells you more about the reporting than about the projects. For a sector that prices its future on the assumption of uninterrupted compute expansion, that ambiguity is itself a risk factor.
If the industry's underlying problem were merely capital scarcity, the number would not matter much. It is not. The problem is distributional. Data centers produce a peculiar arrangement of costs and benefits. The revenue flows to a global cloud provider, its institutional customers, and the remote users of the services those servers power. The costs โ the strain on a regional grid, the water drawn for cooling, the low-frequency hum, the visual intrusion of a windowless monolith on farmland โ land on a municipality that typically receives a modest property-tax base and a handful of permanent jobs relative to the capital invested. The direct employment from a hyperscale campus is small next to the land it occupies, and that is why so many communities conclude they are hosting a high-consumption, low-return facility. That conclusion is not a public-relations problem. It is a structural asymmetry, and structure does not respond to messaging.
There is a mechanical feedback here that I have watched unfold in other markets. When several large loads request interconnection in the same region, the grid operator must plan upgrades that take years and are partly socialized across ratepayers. Residential bills rise before the data center ever comes online. The sequencing is what matters: the community experiences the cost before it experiences the promised benefit, and the benefit may never become locally visible at all. The predictable result is political rather than technical. Municipalities draft moratoriums, counties tighten conditional-use permits, and states begin to treat large-load interconnection as its own regulatory category. A moratorium is not a protest; it is a rule, and rules outlive the news cycle that produced them. The most durable risk to the AI buildout is not that a few projects slip. It is that the permitting regime changes permanently, raising the cost of the next thousand projects before any of them are announced.
This is where the crypto industry feels an obvious temptation. If centralized compute has a social-license problem, decentralized compute markets can present themselves as the option that requires no permission from a county board. Physical infrastructure finance and decentralized physical infrastructure networks โ the awkward acronym the industry settled on โ propose a different distribution: rather than concentrating a gigawatt in one contested location, distribute smaller workloads across many participants who consented simply by opting in. For inference, for data availability, for the verifiable execution of specific tasks, that argument has genuine geometry behind it. My own research into verifiable compute markets modeled the incentive for AI agents to transact on-chain precisely because cryptographic proof can separate a computation's result from any need to trust the machine that produced it. We projected a market approaching $500 million for verifiable data sources by 2028 โ modest against the $228 billion headline, but real, and growing for a reason that is not purely speculative.
There is a second, more interesting crypto thread running beneath the obvious one. Power procurement is becoming a financial problem as much as an engineering one โ long-dated contracts, intermittency risk, and price exposure stretching across years. On-chain instruments for trading energy and verifying renewable attribution are a natural fit here, and I have watched tokenized-power experiments migrate from pilot to procurement tooling. The pitch is that a verifiable ledger of where electrons came from and when they were consumed could make a hyperscaler's environmental claims auditable rather than asserted. Whether that becomes standard practice or stays a demonstration project depends on a question the industry rarely asks itself: does the verifiability change a decision, or merely decorate one already made on other grounds?
I am not, however, prepared to accept the redemption narrative at face value. The same industry that now markets distributed compute as the antidote to centralized bottlenecks spent five years fragmenting liquidity across dozens of Layer 2 networks serving a remarkably stable and modest user base. The mechanism was always identical: a genuine technical limitation, reframed as an opportunity, addressed by proliferation rather than consolidation. Decentralized compute risks repeating that error. Splitting a scarce workload across hundreds of uncoordinated nodes is not the same act as serving it efficiently, and a network that cannot coordinate latency-sensitive inference will not absorb the demand that hyperscale campuses alone cannot meet. The decentralization that matters here is not distribution for its own sake. It is the enforced verifiability that lets a buyer trust a stranger's machine โ a narrow, defensible claim, and far more credible than the assertion that scattered nodes will replace a gigawatt campus. Fragility, after all, is the price of unsecured innovation; the question is whether the new infrastructure secures anything at all.
Within the centralized industry, the backlash is already redistributing advantage. Firms with existing land banks, executed power purchase agreements, and durable relationships with regulators can absorb delay. Firms chasing announcements without secured interconnection cannot. This creates a quiet time-lag advantage: the operator that gets a substation approved this year sets the cost of computation for competitors who will not come online until the end of the decade. The moat is no longer the model. It is the queue position at the grid operator. Where AI capability converges โ and it is converging โ the differentiator becomes who holds the megawatts and who holds the community's tacit assent. Those are the assets that no open-source checkpoint will clone.
The second-order financial effect is easy to miss because it is quiet. Data centers are long-lived assets with contracted tenants, which makes them attractive to the same infrastructure funds and REITs that treat them as bond-like. A stalled pipeline does not destroy that value; it concentrates it. Operating campuses and near-completed facilities become scarcer, and scarcity in a contracted cash-flow asset tends to raise valuations rather than lower them. The losers are the developers holding pre-permit land and the AI firms that assumed they could lease capacity on demand. The winners are whoever already owns the substation and the fiber route.
What all of this points toward is a migration of the binding constraint. The technology stack โ accelerators, memory, interconnect, training frameworks โ has matured to the point where its problems are engineering problems, solvable with enough money and talent. The constraint that remains is not engineering at all. It is the ability to convert abstract demand into a physical site with power, water, cooling, and a community that will not oppose it, inside a jurisdiction whose rules will still permit it in five years. That is a capacity measured in political capital, not compute. And political capital, unlike electricity, cannot be generated on demand from a turbine; it must be accumulated slowly, through relationships and concessions that no spreadsheet captures.
None of this absolves the industry of the ethical question it keeps deferring. A company may lawfully site a campus and still act unjustly in doing so, if it extracts more from a place than it returns and calls the transaction consent. The communities now organizing are not irrational; they are the first actors in the value chain to insist that externalities be named rather than exported. That insistence is uncomfortable precisely because it is correct.
The obvious conclusion โ that decentralized compute wins the backlash โ is the one I distrust most. Distributed networks cannot serve the frontier training runs that absorb the industry's largest capital, and their real value is confined to verifiable, latency-tolerant work. The deeper contrarian point is not about technology at all. The backlash is not AI's obstacle; it is the market finally pricing something the AI industry was never forced to value: consent. For a decade, software externalized its costs invisibly. AI cannot. It must land on real ground and consume real electrons, and the moment it does, it becomes answerable to the people who live there. This is precisely the lesson DeFi refused to learn in 2020, when protocols treated governance, users, and risk as friction to be routed around. The glass house of unsecured innovation does not shatter from competition; it shatters under the weight of the externalities it declined to hold. The crypto industry should read the AI backlash not as an advertisement for itself, but as a warning about what happens when a field believes its own abstraction and forgets that every system ultimately rests on something physical โ and on someone's consent.
The most valuable asset of the next cycle will not be measured in parameters or tokens. It will be the capacity to acquire megawatts, land, and social permission at the lowest possible cost of conflict. For AI, that redraws the competitive map around geography and governance rather than benchmarks. For crypto, it opens a narrow window: verifiable compute can prove it is infrastructure rather than another narrative, or it can fragment a scarce workload into a hundred unconvincing pieces and call the fragmentation an ecosystem. Beyond the illusion, the current never truly stops. In the quiet aftermath, only the resilient remain โ and this time, resilience means knowing the difference between building on someone's land and building with their permission.