The Interconnection Queue Is the New Mempool
A crypto publication ran a technology brief this week that contained zero crypto. No ticker. No chain. No protocol. No address. No block height. The claim was that hyperscale data center operators are shifting expansion overseas because of rising costs and local opposition. Five assertions. No numbers. No counterparties. No dates. No sourcing.
I read it the way I read an unaudited contract submission. There is no state, no caller, no invariant, no test suite. It does not compile to a decision.
But the emptiness is itself the finding. The story the brief declined to tell is the most consequential convergence event in this industry since the merge: the collision of the proof-of-work hashrate fleet with the AI compute buildout, both bidding for the same scarce input. Firm electricity at an interconnectable site. A crypto outlet covered an energy story and missed the crypto inside it.
Context: Every Data Center Is a Power Contract With a Building Attached
Compute is replaceable. The interconnect is not. That single distinction has governed mining economics since 2021.
China's mining ban in mid-2021 removed more than half of global hashrate in a matter of weeks. What survived was not the fastest fleet or the best-operated fleet. What survived was the set of operators holding executed interconnection agreements and energized substations in jurisdictions that would still let them run. The fleet relocated to Texas, to Alberta, to upstate New York hydro, to Kazakhstan, to Paraguay, to the UAE. The United States share of global hashrate climbed from a rounding error to roughly a third or more of the network.
The physical asset that mattered through all of that was never the ASIC. An ASIC is a depreciating box with a two-to-four-year useful life and a resale market measured in fractions of cost. The asset is the site: the land, the transformer, the switchgear, and above all the queue position and the power purchase agreement behind it.
Then the subsidy halved. The April 2024 halving cut the block subsidy from 6.25 BTC to 3.125 BTC. For an operator running older-generation hardware at 25 to 30 joules per terahash and an all-in power cost around $0.06 to $0.08 per kilowatt-hour, that halving did not compress margins. It removed them for a meaningful slice of the installed fleet.
At the same moment, the demand curve on the other side of the fence changed shape entirely. AI training racks went from 5 to 15 kilowatts to 40 to 130 kilowatts per rack. Air cooling stopped working at that density. Liquid cooling, direct-to-chip and immersion, became the default rather than the experiment. Power procurement stopped being a line item in a facilities budget and became the whole business.
Two industries that look nothing alike now want exactly the same thing: a substation, a megawatt, a signed long-term power agreement, and a position in a queue. That is the context the brief skipped. It described the migration as cost-driven and friction-driven. The actual drivers are three. The AI demand curve. Jurisdictional de-risking. Sovereign cloud mandates. Cost and local opposition are second-order variables dressed up as causes.
Core: The Queue Is a Ledger
The interconnection queue is a ledger. It records every request to inject power into, or withdraw power from, a transmission network. It has entries, timestamps, ordering rules, and a settlement process. It also has something most ledgers do not: a multi-year settlement finality window.
In PJM, the largest organized market in North America, requests have waited years to clear study cycles, and some have waited seven years or more. MISO is not materially better. ERCOT is faster because Texas chose a connect-and-manage regime that lets generators and large loads energize ahead of full transmission upgrades, absorbing the congestion risk themselves. That is why Texas won the mining fleet. It was not ideology. It was settlement latency.
Here is the audit question I would put to any operator pitching an expansion plan. What is your queue position, and what is the energization date attached to it? Compare that to a block confirmation. You can broadcast a transaction. Broadcasting does not mean you are in the block. Paying a study deposit does not mean you have a megawatt. In a normalized grid, the binding constraint is not capital. Capital is abundant and cheap. Interconnectable megawatts are neither.
I stress-tested a structurally identical problem in 2020. I wrote a Python script that sampled 10,000 random liquidity events against the Compound V1 interest rate model and measured the tail. The conclusion was not that the market breaks. The conclusion was that the model breaks first, and the market finds out afterwards. The same ordering holds here. The capacity model breaks before the capex budget does. Capital will still be available on the morning the grid operator slips an energization date by eighteen months. The model will already be wrong.
Core: The Conversion Arithmetic Nobody Puts in the Deck
Take a 100 megawatt mining facility running a mixed fleet at an average 25 joules per terahash. At 25 kilowatts per petahash, 100 megawatts equals roughly 4 exahashes per second of installed capacity.
Against a network hashrate in the range of 600 exahashes per second, that site holds roughly 0.67 percent of global hashrate. Daily issuance plus fees at current parameters runs somewhere near 480 BTC per day, depending on fee load. At a 0.67 percent share, the site earns about 3.2 BTC per day. At $60,000 per coin, that is roughly $192,000 of daily gross revenue. Power draw is 2,400 megawatt-hours per day. At five cents per kilowatt-hour, that is $120,000 per day of power cost. Gross margin lands near $72,000 per day before labor, repairs, insurance, and financing.
Now reprice the same 100 megawatts as an HPC or AI hosting contract. Reported third-party hosting economics have run in the neighborhood of $1.2 million to $2 million per megawatt per year, depending on density, SLA tier, and counterparty credit. At the low end, 100 megawatts produces $120 million of annual revenue. Mining at the assumptions above produces roughly $70 million of annual gross revenue carrying full bitcoin price risk and full network difficulty risk.
The host contract is larger, longer, and far less volatile. This is not a strategic preference. It is arithmetic. When an operator says the market forced a pivot, the market did not force anything. The spreadsheet did.
The conversion is also not free. A mining hall is built for low-density, high-airflow, interruptible load. An AI hall is built for high-density, liquid-cooled, non-interruptible load. The retrofit cost per megawatt is real, and it is frequently underestimated in early announcements. The brief compressed all of this into the word cost and moved on. The word cost is where the entire analysis lives.
Core: The Contract Layer Nobody Reads
The mining-to-AI conversion is not a software migration. It is a renegotiation of the entire obligation stack, and the direction of that renegotiation is the opposite of what most people assume.
Mining contracts are short, flexible, and interruptible. Curtailment is a feature, not an outage. In ERCOT, a large flexible load can reduce consumption during scarcity events and get paid for it. A mining operator can shed load, ride a price spike down, and survive. Optionality is the product.
HPC and AI hosting contracts are the inverse. They are ten to fifteen years, frequently take-or-pay, with availability SLAs, uptime guarantees, and liquidated damages. The operator swaps optionality for duration. Under stress, a miner can turn off. An HPC host that turns off pays penalties and may lose the contract outright.
I have audited this shape before. In May 2022 I spent seventy-two hours tracing the Anchor and LUNA interaction sequence at the function-call level, and the finding was not that any single call was obviously fatal. The finding was that the failure was a chain of individually reasonable assumptions. Redemption assumed collateral value. Collateral value assumed oracle integrity. Oracle integrity assumed deep liquidity on a thin venue. Liquidity assumed continued deposits. Each link was defensible in isolation. The composition was not.
HPC hosting has the same skeleton. The availability SLA assumes the utility delivers. The utility arrangement assumes the grid operator does not curtail. The grid operator assumes the transmission upgrade completes on schedule. The project financing assumes the SLA holds for fifteen years. Four links. Each individually defensible. Nothing in a press release discloses what happens when the second link and the third link fail in the same quarter.
Stress tests reveal the fractures before the flood. Formal verification is the only truth in code, and it is the only truth in a power contract. Most of this sector is running neither.
Core: The Jurisdictional Arbitrage and Its Mirrors
The brief's implicit frame is that operators move abroad to escape friction. The historical record says the friction moves too.
China banned mining in 2021 and the fleet relocated. Ireland imposed an effective moratorium on new data center connections in the Dublin region from 2021 onward, driven by grid capacity, and operators relocated. The Netherlands restricted hyperscale builds in the Amsterdam and Haarlemmermeer areas, and operators relocated. Singapore paused new data center construction from 2019 to 2022, and operators relocated. Every destination subsequently developed its own constraint, because a grid is a physical object with a fixed thermal and transmission budget, and because local politics react to local load growth regardless of which flag the operator flies.
Quantify the arbitrage. A 100 megawatt site capturing a three-cent-per-kilowatt-hour discount relative to an alternative market is worth roughly $26 million per year in avoided cost. That is the entire prize. A single retroactive levy, a curtailment mandate, or a two-year interconnection delay erases it completely.
There is a second, cleaner parallel. I have argued for years that there are dozens of Layer 2 networks competing for the same small base of real users, and that this is not scaling but slicing an already scarce pool of liquidity into fragments too thin to be useful. The data center market is doing the same thing with power. Dozens of markets, dozens of announcements, one fixed envelope of interconnectable generation and transmission. Splitting demand across more jurisdictions does not create more electricity. It creates more coordination cost and more single points of failure.
Core: Sovereign Demand and the Stablecoin Mechanism
Data sovereignty regulation creates genuine demand for compute that never crosses a border. GDPR set the template. National localization rules have extended it. That produces a buyer for sovereign cloud capacity who is not especially price-sensitive at a 20 percent premium and is extremely jurisdiction-sensitive at any discount.
The mechanism is the same one that drives stablecoin adoption in high-inflation economies, and it is worth stating precisely because the industry habitually misattributes it. The uptake of dollar-denominated tokens in Argentina, Nigeria, and Turkey is not driven by ideological affinity for permissionless systems. It is driven by the failure of the local unit of account. A user earning in a currency losing purchasing power faster than they earn it does not care about decentralization. They care about the next ninety days.
Sovereign compute demand works identically. The buyer is not purchasing cloud as a category. The buyer is purchasing jurisdictional certainty, because the cost of a data localization failure is not a 20 percent premium, it is a market exit. Demand here is a function of constraint, not narrative. Anyone modeling this market from the top down, starting from industry growth rates, will systematically overestimate and systematically misplace capacity.
Core: What Is Actually Auditable Here
For anyone positioning in a sideways tape where direction is absent and patience is the only edge, the signal is not in press releases. It is in the filings, the queues, and the contract books. Four things are checkable.
Power. Megawatts under contract, weighted average PPA tenor, and all-in dollars per kilowatt-hour including transmission and ancillary charges. A megawatt figure without a price and a tenor is not a data point. It is a marketing number.
Interconnect. Queue position, upgrade cost allocation, and the energization date as filed, not as announced. The gap between the two is where estimates go to die.
Fleet. Fleet efficiency in joules per terahash, the fraction of load that remains interruptible, and whether the underlying assets are pledged as collateral against existing debt.
Contract book. The revenue mix between self-mining and hosting, and counterparty concentration within the hosting bucket. An operator with 40 percent of revenue from a single counterparty is a structurally different instrument from an operator with 12 percent spread across four. That is measurable, and almost nobody measures it.
The block height does not lie. Neither does a queue position. Both are public. One of them gets ignored because it is not on a chart.
Contrarian: The Industry Is Encumbering Its Own Founding Claim
The consensus framing, the brief's framing, is that the migration is a cost optimization. The consensus crypto framing is that it is a straightforward bull case for miners turned AI hosts. Both miss the structural failure mode, and they miss it in the same direction.
When hashrate was the product, the market was permissionless and self-clearing. Any operator with power and capital could compete, and the difficulty adjustment rebalanced the entire network every 2,016 blocks. The mechanism was brutal and it was incorruptible. When hosting becomes the product, the market closes. You cannot permissionlessly enter a fifteen-year take-or-pay agreement with an investment-grade counterparty. You cannot permissionlessly acquire a queue position that was filed eight years ago. And there is no difficulty adjustment for interconnection queues.
The industry that spent a decade arguing that hashrate is the most honest, unforgeable, and decentralized production process in existence is now converting that fleet into a concentrated, credit-rated, contract-bound hosting business. The hashrate does not disappear. It gets encumbered. Optionality is sold to buy duration, and duration is sold to buy a lower discount rate.
Immutability is a promise, not a guarantee. This is not a bearish call on any operator. It is an observation about what is actually being sold, and it is the part of the trade that nobody is stress-testing.
The second blind spot is worse. AI outputs are unverified. In 2025 I audited a protocol where autonomous agents executed smart contracts, and I found that a linguistic tweak to a prompt could bypass access controls and reach a fund-draining path. A natural-language input defeated a formal access-control structure. I proposed a deterministic verification layer for agent outputs precisely because no such layer existed. The physical buildout is now scaling faster than the verification layer for the workloads it will run. That is a capital expenditure boom financing an unresolved trust problem, and the boom is offshore.
Takeaway: Watch the Queue, Not the Announcement
Three signals over the next two quarters, in priority order. First, energization dates slipping relative to filed schedules. Second, amendments to hosting contracts that disclose counterparty concentration. Third, curtailment volumes in ERCOT as AI load displaces flexible mining load. A sideways market rewards positioning over conviction, and positioning requires a date. Verification precedes value. So does energization.