Consider a claim market where the collateral is not tokens but megawatts, and the settlement layer is copper wire. It clears slowly, it clears unevenly, and most of the claims never settle at all.
ERCOT โ the operator of the Texas grid โ is currently weighing interconnection requests totaling more than 474 gigawatts. Roughly 90 percent of that, about 427 GW, comes from data centers. For scale: ERCOT's all-time system peak load sits near 85 GW, against roughly 130 GW of installed capacity. The queue is therefore about five and a half times the peak the system was engineered to serve.
That is not a demand forecast. It is a queue. And queues lie under precisely the conditions that order books lie: when the cost of placing a claim approaches zero.
Context
Texas did not become the physical backbone of American computation by accident. After China's 2021 mining ban, Bitcoin hashrate migrated into the ERCOT footprint, drawn by cheap stranded gas, deregulated wholesale power, and โ critically โ the ability to be curtailed. Under ERCOT's demand-response programs, a large miner could sell power back to the grid during scarcity, monetizing its own downtime. The miner was a battery that happened to mine.
Then the pivot. Public miners began rebranding as HPC and AI hosting companies. The pitch was elegant: a megawatt of contracted power is a megawatt of contracted power, whichever silicon it feeds. Share prices followed the narrative, not the economics. Long-term hosting contracts were signed with the same developers who, one cycle earlier, had sold exposure to Bitcoin's hash price. The asset base was identical. Only the story changed.
On September 21, Governor Greg Abbott ordered a freeze on new state data-center permits pending an audit. He required developers to disclose generation plans, expected consumption, water use, cooling technology, tax subsidies, and ownership. A report to TCEQ was due October 19 โ two weeks before the November 3 election. In the second quarter, local opposition had already stalled an estimated $68 billion in data-center projects. In August, President Trump called the Texas posture "wrong"; on September 19 he announced an "AI Force" that promised not to obstruct the industry.
So within a single election cycle, a Republican governor and a Republican president publicly collided over the same filing cabinet.
Core
Start with the arithmetic, because the arithmetic is the entire story. A 474 GW queue is not evidence that 474 GW of load will ever materialize. It is evidence that filing is cheap. In interconnection queues, the deposit has historically been small relative to the optionality it purchases. The result is a structural incentive to file first and think later โ to hold position in a serial queue the way a bot holds position in a mempool, back-running every genuine entrant who arrives after you.
If you want a mental model, this is a Sybil attack on physical infrastructure. The attacker does not need to build anything. The attacker needs to occupy the scarce resource โ queue position โ at near-zero cost, crowding out legitimate projects that show up later and find the serial queue lengthened by squatters. Volume is noise; the wallet cluster is signal. The same forensic discipline that separates wash trading from organic flow applies here. Follow the entity behind each application, not the name on the filing.
That is what makes Abbott's disclosure list so revealing. Requiring developers to state generation plans and ownership is, functionally, a know-your-customer check on the queue. The state is trying to distinguish the entity that will actually pour concrete from the entity that is merely holding capacity. Behind-the-meter generation โ self-supplied gas, co-located nuclear, long-duration storage โ becomes the only honest signal of intent, because it costs real capital to arrange and cannot be faked on a spreadsheet.
Now layer in water. The mandate to disclose water use and cooling technology is not bureaucratic garnish. Texas runs under chronic drought pressure, and evaporative cooling consumes enormous volumes of water. Closed-loop liquid cooling consumes far less but demands expensive retrofits of high-density halls. If water permits bind more tightly than power permits โ and in parts of the state they may โ then the ceiling on AI density in Texas is not silicon, not capital, and not even electricity. It is the aquifer.
Here is where the story stops being about Texas. New York froze data-center permits in July. Pennsylvania tightened its permitting rules in August. This is a national trend wearing a state costume, and it collides directly with the crypto-native infrastructure narrative.
A wave of tokenized infrastructure and DePIN projects has spent two years selling the idea that physical compute can be financialized โ that a data center can be decomposed into yield-bearing claims and floated on-chain. Some of these instruments are honest. Many are not. I have spent months auditing the on-chain footprint of entities that market themselves as AI infrastructure plays, and the pattern is familiar: a treasury wallet, a handful of clustered addresses that rotate funds between themselves to manufacture the appearance of organic demand, and a whitepaper whose capacity projections cite interconnection positions that were never actually assigned.
The rug is not pulled; it was never tied.
Consider the math such projects present to retail. A token promises exposure to "AI compute demand." But the underlying physical asset sits somewhere in a 474 GW queue with a historical completion rate that no honest operator can quote with confidence. The token is a claim on a claim. Imagination is infinite, but liquidity is finite. When the physical layer fails to deliver, the financial layer does not slowly reprice. It gaps.
This is also where the disclosure of ownership matters more than the press releases admit. When projects preach decentralization but their foundation wallets, treasury multisigs, and beneficial owners resolve to a short list of traceable addresses, the "community-run" framing collapses under its own block explorer. Gas fees are the price of truth, and the truth here is that a data center LLC in Texas and a foundation wallet in a tax haven are frequently controlled by the same hands.
The $68 billion in stalled projects is the clearest quantification available. That figure is not an abstraction about sentiment. It is capital that arrived, secured land, engaged engineering, and then hit a wall built from electricity, water, and community consent. The capital was abundant. The social license was not.
And the social license problem is bipartisan. A national survey found 69 percent of adults oppose a local AI data center, with 57 percent of Republican voters opposed. This is not an environmentalist fringe. This is the base. The objection is rational: rising power bills, drawn-down water tables, noise, land, and a permanent employment footprint that is tiny relative to the construction jobs that vanish the moment a site goes live. The community bears concentrated costs and receives diffuse, mostly temporary, benefits.
Which brings us to the cost-allocation question the coverage keeps dodging. If grid upgrades driven by data-center load are socialized into residential rates, the political backlash intensifies by design. If they are borne by the developer, the marginal project dies. There is no comfortable third option, and the utility sits between the two, holding the rate case and the political risk simultaneously.
Contrarian
Now the part the bulls got right, because they did get something right, and pretending otherwise would be dishonest.
The freeze is not the death of Texas compute; it is a filter. When capacity is scarce and disclosure is mandatory, the winners are the large, capitalized, compliance-capable operators who can self-generate, sign long-duration power purchase agreements, and absorb audit costs. Smaller colocation providers get squeezed into dependence on third parties and pay a margin for the privilege. Permitting friction is a moat, and moats accrue to the strong. Anyone assuming a permit freeze means less buildout is confusing the denominator with the numerator.
The second thing the bulls got right is that demand is real. The queue is inflated by speculation, but even after you strip the placeholders, the residual footprint likely exceeds the grid's expansion speed. That asymmetry โ real demand, slow supply โ is bullish for power, not for compute.
Where the dominant narrative inverts is Trump's framing that opposing data centers hands AI to China. That is causally backwards. Logic does not bleed, but code leaves traces โ and so does physics. If American interconnection queues, water tables, and community consent slow domestic compute deployment, the objective effect is to narrow the lead America currently holds in physical AI capacity. The "only China is happy" line describes a world where the constraint does not exist. The constraint is the whole point. Framing a domestic cost-allocation dispute as geopolitics is a way to silence the locals, not a way to describe the physics.
Takeaway
Watch the December ERCOT audit, not the rhetoric. Whether it releases, caps, or rejects queued capacity determines how much of the 474 GW is real, and it will do so before any token marketing deck catches up. The grid is the new mempool: capacity is claimed, held, and cleared under competition, and the entities squatting on position get priced out the moment verification has teeth. For anyone holding a claim on AI infrastructure โ tokenized or otherwise โ the question is not whether demand is coming. It is whether the wire, the water, and the neighbor will let it land.
