Hook
We didn't get a model. We didn't get a benchmark. We got a pipeline — or rather, a request to hurry one along. Somewhere in the last fortnight, a filing surfaced asking the Federal Energy Regulatory Commission to accelerate gas infrastructure for an AI datacenter in New Mexico, and the crypto commentariat scrolled past it because it had nothing to do with tokens. That reflex is the tell. Over the past seven days, the loudest signal in the compute economy wasn't a hashrate print or a rollup's blob utilization curve. It was the sight of a hyperscaler treating electrons, not parameters, as the scarce commodity. If you are still modeling AI as a software story, you are modeling the wrong layer. The binding constraint moved. It moved to the pipe in the ground.
Context
The reported facts are thin — deliberately, dangerously thin. Oracle, per Crypto Briefing, has asked FERC to fast-track a gas pipeline serving a New Mexico AI datacenter. No docket number. No capacity figure in MW. No MMcf/d of fuel. No pipe diameter, no length, no interconnect to a mainline, no counterparty. In a domain where a single missing field can invert the entire conclusion, that is not a story; it is a rumor with a press release stapled to it.
Here is what we can anchor. The United States grid interconnection queue has backlogged past 2,600 GW, with median wait times hovering near five years, per Lawrence Berkeley National Laboratory data. A frontier AI campus — think the Stargate Abilene target of roughly 1.2 GW — cannot wait five years. So the industry has converged on a workaround: bring your own power. Behind-the-meter generation. Gas turbines on your own parcel, fed by your own pipe, legally adjacent to the grid rather than dependent on it. The New Mexico case fits this pattern so cleanly that its absence of detail is itself informative.
And the crypto connective tissue is not decorative. Bitcoin miners spent the last cycle stockpiling exactly what AI now covets: energized sites, substation rights, power purchase agreements, and — critically — positions in those same interconnection queues. The AI narrative thinks it invented demand for power. It inherited it.
Core
Start with the constraint nobody wants to price. The bottleneck is no longer silicon. It is the gas turbine. GE Vernova, Siemens Energy, and Mitsubishi Power have seen order books reportedly extend into 2028 and beyond, with delivery lead times stretching from a historical 18–24 months toward three to five years. When you read "fast-track," do not read "regulatory convenience." Read "we are racing for a manufacturing slot and an energization window, and we will pay almost anything to jump the line." A fast-track filing is a queue-jumping instrument dressed as a permitting courtesy.
This is where my own audit reflexes kick in. In 2017 I spent a day tearing apart the Golem pre-sale contracts, not because the math was hard but because the gap between what the code claimed and what it did was where the risk lived. Same instinct here. "Oracle asks FERC" is a claim with a jurisdiction problem buried inside it. Under Section 7(c) of the Natural Gas Act, the certificate applicant for interstate pipeline construction is the pipeline company. A shipper — which is what Oracle would be — normally files a supporting motion or an intervention, not the application itself. So either the reporting collapsed "applicant" into "supporter," or Oracle is building through an affiliate, in which case the real question becomes whether FERC even has jurisdiction or whether this sits under New Mexico's PRC via a Hinshaw exemption. The bug wasn't in the headline. The bug was in the verb.
Then there is the physics of speed. A combined-cycle plant runs near 60% efficiency. An aeroderivative simple-cycle unit, chosen to compress a schedule, runs at 35–40%. You trade roughly twenty efficiency points and a measurable bump in heat rate and emissions for months of time-to-power. In a semi-arid state where air cooling inflates PUE, that trade compounds. Nobody in the AI narrative wants to say "we are building a worse plant on purpose," so they say "fast-track" instead. It is the same sleight of hand I watched during DeFi Summer 2020, when I modeled Uniswap V2's geometric mean pricing for two weeks and realized the real story wasn't yield — it was permissionless liquidity. Permissionless power is the 2026 sequel. Same narrative architecture, different collateral.
Now zoom out, because this is where the crypto read actually bites. Code is law, but liquidity is truth — and in 2025 the deepest liquidity is not in a pool. It is in a queue position. Bitcoin miners who survived the halving are sitting on energized interconnects that AI operators cannot conjure from scratch. That makes the miner a landlord of the scarcest asset in the economy, and the rent is denominated in gigawatts. The "AI eats crypto" framing is lazy. The accurate frame is that crypto accidentally warehoused the real estate that AI now must lease, and the repricing of that real estate is the trade.
Watch the shape of the deals. When a miner converts a site to HPC hosting, the on-chain hashrate does not vanish — it gets reallocated, and the marginal cost of the remaining hash rises. That is a supply shock wearing a hosting contract's clothing. Liquidity pools don't care about your narrative; neither do substations. Both clear at whatever price the marginal buyer will pay, and the marginal buyer right now is a model-training cluster with a term sheet and a deadline.
I saw this same pattern in 2021, when I ignored the Bored Ape floor chart and built a Resonance Index out of social-capital metrics instead. The number that predicted the peak wasn't price. It was the rate at which status anxiety outran utility. The AI power trade has an identical asymmetry: the narrative of infinite compute is pricing the queue before the turbines exist. The queue is the meme. The megawatt is the reality. When the two diverge, the meme always loses — eventually.
Layer this onto the regulatory seam. The 2025 PJM co-location disputes — Talen/Amazon and its siblings — already forced FERC to wrestle with whether behind-the-meter large loads should pay for the transmission they nominally avoid. The New Mexico filing, if it is real, is the same argument in a different jurisdiction: can a single industrial customer route around the public grid's cost-allocation logic by building private supply? Every yes is a precedent. Every precedent is a repricing of the commons. And note what is absent from the filing's public surface: emissions permits, CCUS commitments, fuel volumes, IT load, PUE. A brief that omits its own load figure is not being transparent. It is being tactical.
Contrarian
Here is the blind spot the AI-versus-crypto frame hides. Everyone is debating who gets the power. Almost nobody is asking whether the queue position itself has become a speculative instrument — a claim on future electrons traded long before a single turbine spins. That is a familiar shape. In 2022 I spent three months dissecting Terra's mechanism and wrote it up as the mathematics of delusion, because a system that needs infinite growth to stay solvent is not a system; it is a countdown. The interconnection queue has a whiff of the same structure. It rewards whoever holds the paper, not whoever delivers the megawatt, and the paper can be flipped before reality audits it.
The bug wasn't the demand forecast. The bug was assuming the queue clears at the speed the financial model assumes. Turbines are the auditor here, and they do not negotiate.
Takeaway
So watch two numbers, not the model launches: the docket, and the delivery date. If the gas capacity is real, the next domino is emissions permitting and a fuel-supply contract that will eventually touch a public utility's cost allocation — and that is where the precedent gets set for every miner-turned-host from Texas to Virginia. If it is not real, we just watched a narrative get priced before its hardware existed. Either way, the pipe is the product now. The only question is who is left holding the queue when the turbines finally ship.