Entergy Mississippi just told its single largest prospective customer that it might not be able to keep the lights on. The customer is AWS. The message was a disconnection warning.
Strip the press-release language and you have a rare event. Utilities do not threaten to cut off a $10 billion anchor tenant unless something structural has already broken. Power companies compete for data centers the way launchpads once competed for token listings. They offer speed, tax abatements, priority interconnection. They do not, as a rule, publish the possibility of pulling the plug.
So when one does, I stop reading the narrative and start reading the numbers. A disconnection warning is not an engineering notice. It is a settlement notice. Someone is about to be handed a bill, and the utility is making sure it is not the one holding it. The warning is the artifact. The cost allocation is the crime scene.
Here is the background the 60-word news item leaves out.
The category that matters is "large load." In U.S. utility regulation, a large load is any customer drawing roughly 100 MW or more. Historically that meant a smelter or a refinery. Since 2023 it means two things: AI data centers and, in the same regulatory bucket, crypto mining operations. The two are now fused in the eyes of every state commission. When a utility files a large-load tariff, it is writing rules that will govern both.
Why crypto media picked up a Mississippi utility filing tells you more than the filing itself. The editorial instinct was not random. It reflects a real 2024–2025 convergence: miners and hyperscalers are competing for the same constrained electrons, and regulators have stopped distinguishing between them. The equipment does not care whether the compute is hashing SHA-256 or running inference. Neither does the substation.
Now the supply side. The U.S. interconnection queue holds roughly 2,600 GW of projects waiting for grid access, with an average wait near five years. Gas turbine lead times have stretched from about 18 months to 2027–2028. Transformer lead times run two to three years. Data centers build in 18–30 months. Power plants build in four to five. The load always arrives before the supply, and that gap is not a market inefficiency — it is the defining constraint of the decade.
Mississippi used to be a power-surplus, low-rate state. That surplus is the product being sold to AWS. The disconnection warning is the moment the state discovered its inventory was smaller than its promises.
Let me be forensic about what "disconnection" actually encodes.
A utility issues a public disconnection warning for one of three reasons. First, genuine capacity deficit: the effective generating capacity, not the nameplate number, cannot cover the customer's peak. Second, interconnection sequencing failure: the transmission or substation upgrades required cannot be completed inside the customer's energization schedule. Third, negotiation leverage: a rate case or large-load tariff is stuck, and the utility wants both the regulator and the customer to feel the ceiling.
All three share a single root cause. The utility under-forecast load growth, committed capacity it did not have, and is now trying to reprice the commitment after the fact. This is the same failure mode I documented in the 2020 Compound governance gap — a system designed around a theoretical load profile, exposed the moment real capital moved through it. I simulated a governance attack on the cETH contract then and found a 12-second window with no slippage protection. The protocol did not fail because it was badly coded. It failed because the model of who would show up was wrong. Governance is just a slower attack vector, and load forecasting is governance.
Here is the piece everyone skips: who pays for the new generation?
The utility wants the data center to carry the incremental cost through a large-load tariff, so residential ratepayers are not subsidizing a trillion-dollar company. The data center wants cheap, firm, reliable power or it relocates. The regulator is terrified of stranded cost — the scenario where the utility builds generation for the load, the load leaves, and the bill lands on households in a state with one of the highest poverty rates in the country.
That three-way standoff is the entire story. The disconnection warning is the utility's nuclear option inside it. When a power company publicly threatens to cut its biggest new customer, the rational read is not "engineering accident." The rational read is "the contract is not signed."
I have spent years tracing who exits before the collapse. In the 2022 Terra/Luna cascade, the signal was not the depeg — it was three wallet clusters moving hours before the peg broke. I mapped $40 billion of collapse through wallet clusters and found the insiders who left early. The event was a predatory execution, not a market accident. Silence in the logs is the loudest scream. Here, the loudest scream is a utility saying out loud what its models already told it: it cannot guarantee the electrons it sold.
And notice what the item refuses to quantify. How many megawatts short? Which rate case number? What generation is being proposed — and is it gas? That last question is the ESG trap. If Entergy backfills AI load with new combined-cycle gas, AWS's Scope 2 emissions rise against its own net-zero pledge. The "carbon neutral" story quietly requires a fossil buildout. Trace the hash, ignore the hype.
The bulls are right about one thing, and it is the thing utilities fear most.
The correct answer to a constrained grid is not to wait for the grid. It is to bypass it. Behind-the-meter generation, co-location with existing nuclear, dedicated PPAs — the Talen–Amazon Susquehanna co-location fight proved the model is real and the regulator knows it. FERC's 2024 ruling against expanding co-location did not kill the strategy; it confirmed the strategy was threatening enough to litigate.
The blind spot runs the other way. Decentralized power sounds resilient, but co-location simply moves the single point of failure from the utility substation to a private interconnect agreement. The miner or hyperscaler that builds its own gas plant has not escaped the grid. It has become the grid, with none of the public oversight and all of the same fuel risk. The 2021 BAYC metadata exploit taught the same lesson in a different medium: the artwork lived on-chain, the image lived on one server. Ownership was a promise with a central point of failure behind it. Immutability is a promise, not a feature — and so is a power purchase agreement.
The grid is a ledger. Every megawatt is a line item, and every line item settles eventually. Entergy just wrote an entry that most utilities only whisper in closed rate cases: the load was sold before the capacity existed.
Watch two numbers. The Mississippi Public Service Commission's next large-load tariff docket, and the fuel mix of whatever generation Entergy proposes to close the gap. One tells you who pays. The other tells you whether "carbon neutral compute" was ever real.
When the ledger clears, count who is still holding the position.