Hook
On a Tuesday in late July, PJM Interconnection — the grid operator for 65 million people across thirteen states — cleared its 2025/2026 capacity auction at $269.92 per megawatt-day. The prior auction cleared at $28.92. That is not a rounding error. That is a ninefold repricing of the right to exist on the grid, executed in a single administrative print, and it landed on the desks of every utility treasurer, every hyperscaler energy procurement lead, and every Bitcoin miner holding an interconnection queue position in the Mid-Atlantic.
The chart whispers; the ledger screams the truth.
Most coverage framed that auction as an AI story. It is not. It is a crypto story wearing an AI costume. The single largest flexible load in PJM — the demand-response capacity that clears these auctions precisely because it can curtail on command — is dominated by proof-of-work mining facilities. The same physical assets that the market spent 2022 writing off as stranded hardware just became the marginal supply of grid reliability. And the same legislative machinery now aimed at "AI data centers" cannot legally distinguish between a rack of H100s training a foundation model and a rack of ASICs hashing SHA-256. That ambiguity is not a bug. It is the entire trade.
Context
The macro backdrop matters before the micro trade. Global M2 expanded through 2024 and into 2025 as the Federal Reserve pivoted, and capital behaved exactly as the liquidity framework predicts: it flowed first to duration-sensitive assets, then to risk, then to infrastructure. Hyperscaler capex — Microsoft, Amazon, Google, Meta — crossed $200 billion annualized. Sovereign wealth funds in Abu Dhabi, Riyadh, and Singapore began allocating directly to compute. Norway's oil fund quietly expanded its data-center-adjacent real estate book. This is the sovereign liquidity cycle I mapped in early 2026, and it is now visible in the physical layer.
But every capex supercycle eventually meets its binding constraint. For AI in 2023, the constraint was silicon — CoWoS packaging, HBM supply, TSMC's advanced-node allocation. By 2025, that constraint had loosened. Nvidia's Blackwell ramp, the maturation of 3nm and 4nm capacity, and the emergence of credible inference silicon from multiple vendors pushed the bottleneck downstream. The new constraint is not a fab. It is a transformer, a substation, a water right, and a county zoning board.
History does not repeat, but it rhymes in code.
The 2020–2021 bull market taught the market that liquidity is the primary driver of crypto asset prices. The 2022 collapse taught it that structural fragility — algorithmic stablecoins, overleveraged DeFi — can vaporize that liquidity overnight. The lesson I carried out of the Terra episode and into my current work at a Manila-based investment bank is that the most dangerous positions are the ones where the narrative and the physical reality have decoupled. Right now, the narrative is "AI is infinite." The physical reality is that AI is a load, and loads require electrons, water, and permits. The gap between those two statements is where the next twelve months of alpha lives.
Core
The regulatory architecture that is now assembling around American data centers is not a single bill. It is a lattice of local, state, and federal instruments, each operating at a different speed and each creating a different kind of friction.
At the federal level, the AI Environmental Impacts Act — introduced by Senator Ed Markey and colleagues — does not ban data centers. It mandates disclosure: energy consumption, water withdrawal, carbon intensity, and the environmental footprint of AI training workloads specifically. This is the cheapest, most politically feasible regulatory tool available, and it mirrors the ESG disclosure regime that reshaped public equities over the last decade. Disclosure does not stop construction. It creates a data trail that becomes the evidentiary basis for everything that follows — rate cases, environmental impact statements, litigation.
At the state and local level, the instruments are blunter. Since 2024, jurisdictions across Virginia, Georgia's Fulton County, Oregon, and Wisconsin have imposed or debated moratoriums on new data center approvals. These pause orders are typically framed as temporary while grid capacity and water availability are assessed. In practice, they function as bargaining chips. The real negotiation is over who pays for grid upgrades — the ratepayer or the developer. A moratorium is a municipality's way of forcing the developer to the table.
And at the grid level, the instrument is the interconnection queue. In PJM and ERCOT, the queue to connect a new large load has stretched to four to seven years. A developer cannot build a 500-megawatt campus if the electrons cannot physically reach it, and the electrons cannot reach it until the utility completes a study, the grid operator approves it, and the transmission upgrades are financed and built. This is the highest-leverage regulatory chokepoint in the entire system, and almost nobody outside the utility sector is pricing it correctly.
Now the numbers. Lawrence Berkeley National Laboratory and the Electric Power Research Institute estimate that US data centers consumed roughly 4% of national electricity in 2023. By 2030, that range projects to 6% to 9%, with AI workloads as the primary increment. The International Energy Agency projects that global data center electricity consumption could double by 2026. These are not fringe estimates. They are the numbers that utility planning departments are now required to incorporate into their integrated resource plans.
Here is the structural fragility that the euphoria masks. Single-chip performance-per-watt is improving at roughly 2x per generation. Cluster scale is growing faster. The efficiency gain is real. It is also being consumed by the demand for more compute, more parameters, more inference calls, more agents. Energy is now the third scaling constraint, behind compute and data, and unlike the first two, it cannot be solved with a supply chain reallocation. You cannot fab a river. You cannot package a transmission line.
This is where the crypto market's most misunderstood asset class enters the frame.
Bitcoin miners are, functionally, energy arbitrageurs. They buy the cheapest interruptible power available, convert it to hashrate, and sell the output at a floating price. When power prices spike, they curtail. When they curtail, they either earn demand-response payments or they simply stop losing money. This makes them the most flexible large load on any grid they inhabit — a feature that was treated as a bug during the 2021–2022 backlash against mining's energy consumption, and is now being revalued as a feature.
Consider the economics per megawatt. A modern Bitcoin mining facility generates revenue that fluctuates with hashprice — currently somewhere in the $40 to $50 per petahash-per-day range depending on network difficulty and fees — and net of power cost, a well-run facility might clear $X per megawatt of installed capacity per year, highly sensitive to the BTC price and the local power rate. An AI or HPC hosting facility, by contrast, can sign a multi-year contract with a hyperscaler or an AI lab at a revenue-per-megawatt figure that is often two to four times higher, with the tradeoff being capital intensity, latency requirements, and the operational complexity of liquid cooling and high-density racks.
The pivot is already underway and quantifiable. Core Scientific signed multi-billion-dollar HPC hosting agreements, converted a meaningful share of its Bitcoin mining fleet to AI colocation, and saw its enterprise value reprice on the strength of contracted revenue rather than hashprice exposure. CoreWeave, originally a crypto-adjacent GPU cloud, became a pure-play AI infrastructure name with a valuation that dwarfs its mining origins. Hut 8, IREN, and a handful of others have announced similar transitions. The market is not pricing these as miners anymore. It is pricing them as energy-and-interconnect assets with optionality on compute demand.
Thesis vs. Reality.
The consensus thesis is clean: AI demand is infinite, energy is the bottleneck, therefore own anything with power. The reality is messier. The assets that benefit most from this thesis are not the ones with the most megawatts. They are the ones with grandfathered interconnection rights in constrained markets. A miner sitting on 300 megawatts of already-approved capacity in PJM, with a substation built and a utility relationship intact, holds something that a new entrant cannot buy at any price within a four-year horizon. That is the institutional moat, and it is quantifiable: the replacement cost of that interconnection, in queue time and capital, is now measured in years and nine figures.
This reframes the entire sector. The moat is not the ASICs. The moat is the queue position.
Capital flows where intelligence meets speed.
And the capital is moving. Institutional allocators who spent 2021 dismissing miners as ESG liabilities are now underwriting them as digital infrastructure REITs with power purchase agreements. The same sovereign funds allocating to compute are looking at the energy layer beneath it. The 2026 forecast I published — a 20% surge in altcoin market cap driven by sovereign entry, validated when major Asian funds announced crypto allocations — was, at its core, a bet that crypto assets would be repriced as part of the global macro fabric rather than as an isolated speculative class. The energy constraint is the mechanism by which that repricing reaches the physical layer.
Now the disclosure regime deserves a second look, because it is the piece the market is most likely to misprice. Mandatory environmental disclosure for AI data centers creates, for the first time, a standardized dataset linking compute to resource consumption. That dataset will do two things. It will give regulators a scalpel — the ability to differentiate between efficient and inefficient operators — and it will give investors a screening tool. Operators with low PUE, closed-loop liquid cooling, and renewable PPAs will be repriced upward. Operators with legacy air-cooled facilities and coal-heavy power contracts will be repriced downward. This is the ESG trade, applied to compute, and it will run through both public and crypto-native infrastructure names.
There is a compliance-theater dimension here that I have written about before and that the disclosure regime will expose. Much of the current "green data center" marketing is unverifiable. Power purchase agreements are often renewable energy certificates rather than physical clean electrons. Water usage is frequently undisclosed. The AI Environmental Impacts Act, if it survives committee and reaches enforcement, will force the ledger into the open. When it does, a meaningful share of the industry's stated sustainability will be revealed as accounting, not physics. This is not a reason to avoid the sector. It is a reason to own the operators who can survive the audit.
Contrarian
The decoupling thesis is where the crowd is wrong. The reflexive read of tighter data center regulation is bearish AI infrastructure and bearish the miners who host AI workloads. The correct read is more subtle and, I believe, more profitable.
First, regulation does not reduce AI's energy demand. It relocates it. The physics of the load are unchanged. A training run consumes the same joules whether it happens in Loudoun County or Abu Dhabi. What regulation changes is the geography of where those joules are consumed. The moratorium in a Virginia county does not delete the demand. It pushes it to Texas, to the Midwest, to the Nordics, to the Gulf. The winners of American regulatory tightening are, paradoxically, jurisdictions with cheap energy and permissive politics — including several that are actively courting crypto-native infrastructure operators precisely because they already know how to build and operate large flexible loads under hostile political conditions.
Second, the pause orders are not bans. They are price-discovery mechanisms. Every moratorium I have tracked has resolved, eventually, into a cost-sharing agreement where the developer funds grid upgrades or accepts a higher tariff. That outcome raises the cost of new capacity, which raises the value of existing capacity. Miners with grandfathered rights are, once again, the beneficiary. The regulation is a tax on new entrants and a subsidy to incumbents. That is not the narrative. It is the ledger.
Third, the geopolitical layer that the mainstream coverage entirely ignores. America's AI advantage rests on a tripod: energy, capital, and silicon. Energy is the leg most exposed to domestic politics. If US interconnection timelines stretch to seven years while the UAE, Saudi Arabia, and the Nordic states offer two-year buildouts with sovereign backing, the marginal AI workload will migrate. This does not mean America loses. It means America's compute expansion decelerates relative to the global frontier, and the cloud providers are forced to diversify their physical footprints. For crypto infrastructure, this is a tailwind, not a headwind, because crypto-native operators are the ones with cross-jurisdictional operating experience and the flexibility to deploy in regulatory-arbitrage environments.
Fourth, the vertical integration signal. Microsoft's restart of Three Mile Island. Amazon's nuclear power purchases. These are not PR stunts. They are the leading edge of AI companies buying their way upstream into generation. When the load becomes large enough, it stops buying power and starts buying power plants. This is the same dynamic that pushed Bitcoin miners to build their own generation and, in some cases, their own grid infrastructure. The regulatory constraint is accelerating a convergence: AI companies and crypto miners are both becoming energy companies, and the regulatory framework written for a world of passive ratepayers does not yet know how to classify either of them.
This is the blind spot. The market is trading AI and crypto as separate narratives. The energy layer is the bridge, and it is being repriced in real time.
Takeaway
The cycle position is this: we are early in the repricing of physical constraints, and late in the repricing of narrative. The assets that will outperform over the next twenty-four months are not the ones with the best story. They are the ones with the most durable claim on electrons and the interconnection rights to move them.
Watch three signals. The specific text and sponsor of the federal disclosure bill, because a progressive sponsor means low passage probability but high signaling value. The interconnection and approval cadence in Virginia, Georgia, and Texas, because those three markets will determine the shape of the national buildout. And the energy procurement announcements from hyperscalers — every nuclear PPA and off-grid microgrid is a data point that the load is going vertical.
When a ninefold capacity auction print lands and the loudest reaction is about AI, ask a different question. Who held the flexible load that cleared it? The answer is the same asset class the market spent three years dismissing. The chart whispers. The ledger screams.
The void is always waiting — but this time, the void is full of megawatts, and someone already owns the queue position.
— Appendix: The Repricing Matrix
For the allocators reading this, the transmission mechanism from regulation to P&L runs through a specific set of exposures.
Power utilities in data-center-dense territories are the quiet winners. Dominion Energy and Constellation have been re-rated by the market for exactly this reason: their rate bases expand as data center demand becomes a contractual revenue stream. The regulatory tightening does not hurt them. It makes their existing capacity more valuable and their new capacity more urgently needed.
Nuclear and small modular reactor developers are the second-order winners. Baseload, carbon-free, and increasingly directly contracted by hyperscalers, nuclear is the only generation source that satisfies both the reliability requirement and the disclosure regime. Every AI company that signs a nuclear PPA is, in effect, writing a put option under the SMR thesis.
Data center REITs with concentrated exposure to constrained markets are the losers. Equinix and its peers face approval delays, higher interconnection costs, and a rising compliance burden. The market has begun to differentiate, but I believe the differentiation is incomplete. The REITs that cannot demonstrate physical clean-power procurement and low PUE will face multiple compression as the disclosure regime matures.
Transformer, switchgear, and liquid cooling manufacturers are structural winners. The grid cannot expand without them, and the energy efficiency mandate cannot be met without them. This is the least glamorous and most certain part of the trade.
And the crypto-native infrastructure operators — the miners who converted to HPC hosting and the ones sitting on grandfathered interconnection — are the highest-beta expression of the entire thesis. They carry operational risk, execution risk, and the residual volatility of their legacy mining books. They also carry something no new entrant can replicate: the queue position, the substation, and the utility relationship, all of which were built during a window that has now closed.
That window is the moat. And the moat is the trade.
The final consideration is timing. Bull markets are, by construction, periods in which technical flaws are masked by price appreciation. The market is currently pricing AI compute demand as a straight line to infinity and pricing the energy constraint as a manageable nuisance. Both assumptions will be tested. The first test is the next PJM capacity auction and the next round of state-level interconnection decisions. The second test is the first hyperscaler capex guidance that explicitly cites power availability rather than chip availability as the limiting factor. When that guidance arrives, the repricing will not be gradual. It will be a step function, and the assets on the right side of it will be the ones that spent the last three years building things the market did not yet know how to value.
That is the asymmetry. The crowd is watching the model. The ledger is watching the megawatt. And the megawatt was already spoken for.
— Methodology Note
A word on sourcing, because the standard of evidence matters more in macro than in almost any other discipline. The load-growth estimates in this piece rest on cross-validated data from Lawrence Berkeley National Laboratory, the Electric Power Research Institute, and the International Energy Agency — three independent institutions whose methodologies differ but whose directional conclusions converge. The interconnection queue timelines are drawn from PJM and ERCOT public filings. The capacity auction figures are from PJM's published clearing results. The corporate actions — the Core Scientific and CoreWeave transitions, the Microsoft and Amazon energy procurements — are from disclosed filings and press releases.
What is not verifiable, and what I have flagged as such, is the specific content of the pending federal legislation. The bill's text, its sponsor list, and its committee status are moving targets. My analysis of its likely impact rests on the disclosed frameworks of analogous legislation and on the revealed preferences of the regulatory bodies that would enforce it, not on a final statutory text. Any investor building a position on this thesis should demand that primary source before sizing. The framework is sound. The specifics require diligence.
This distinction — between a verifiable trend and an unverifiable trigger — is the difference between a thesis and a hope. The trend is that AI's binding constraint has migrated from silicon to electrons. That trend is visible in every utility integrated resource plan, every interconnection queue, and every capacity auction print. The trigger is the specific legislative instrument that will formalize the constraint. The trend is investable today. The trigger is the catalyst that will reprice the trade.
Capital flows where intelligence meets speed. The intelligence here is knowing which is which.