In the queue of PJM Interconnection, 200 GW of data center capacity sits waiting. That is more than the entire peak demand of France. But the applications are hollow: no binding power purchase agreement, no financial guarantee—just a placeholder on a spreadsheet. The code behind them is empty. In the code, I found the ghost of the architect.
I have seen this pattern before. In 2017, I audited smart contracts for a project that promised a decentralized exchange but held no liquidity. The blockchain was clean; the intent was not. Now, the same ghost haunts the US electricity grid, disguised as AI demand. The narrative is seductive: AI will consume 10% of global electricity by 2030, and every data center announcement is a bet on the future. But the atoms of power do not care about narrative. They care about signatures, cash, and timelines.
Context: The Historical Narrative Cycle
The current AI electricity boom echoes the 2021 crypto mining frenzy. Miners announced gigawatt-scale facilities, signed land options, and queued for interconnection—only to vanish when Bitcoin’s price dropped or ASIC efficiency improved. According to the US Energy Information Administration, over 40% of crypto mining interconnection requests in ERCOT were withdrawn or abandoned between 2022 and 2023. The same cycle is now repeating, but with a twist: the load is labeled “AI” instead of “mining.”
Protocols differ, but the pattern is identical. A developer announces a $1 billion data center, secures media coverage, and reserves capacity on the grid. The grid operator, bound by regulation, must plan for that load. It builds new transmission lines, procures gas turbines, and delays the retirement of coal plants. Meanwhile, the real project—the one with a signed PPA and a working GPU cluster—waits in line behind the ghost. The queue becomes a cemetery of intentions.
Core: The Mechanism of Phantom Load
Based on my experience auditing DeFi protocols during the 2020 liquidity mining boom, I learned that token incentives create centralization risks. The same principle applies here: capacity announcements create narrative incentives. A company that announces a 500 MW data center without a binding PPA gains a temporary stock price boost, deters competitors from entering the same substation, and locks up a grid resource that is fundamentally finite. The cost is externalized to the public—ratepayers fund the stranded assets.
Technical analysis of grid interconnection data reveals a stark gap. In PJM, the queue contains over 200 GW of projects, of which only 30% have signed any form of PPA. Of those, fewer than 10% have made financial deposits that exceed the cost of a new server rack. The rest are placeholders. The grid operator’s planning model assumes all of them will materialize, because that is how the system is built: first-come, first-served, with no penalty for withdrawal.
This is not a problem of too much AI demand. It is a problem of zero-cost signaling. When the pool empties, only the intent remains. The real intent is not to build, but to occupy. The occupancy is a form of capital—a new resource that can be traded, sold, or used to extract subsidies. I have seen the same mechanism in crypto: projects that mint tokens without a product, farm liquidity without a protocol, and then exit. The audit is not a check; it is a confession. The grid queue is a confession of the industry’s speculative nature.
Contrarian: The Blind Spot of the Phantom Narrative
The conventional wisdom is that phantom projects are a nuisance that will self-correct when the AI bubble bursts. But the contrarian angle is more uncomfortable: these ghosts are a feature, not a bug. They serve the interests of incumbents who control the limited resource of grid capacity. Large cloud providers and energy companies can afford to lock up capacity for years, waiting for the right moment to build. Smaller AI labs and startups cannot. The phantom load becomes a barrier to entry, a moat built not by technology but by bureaucracy.
I recall a conversation with a grid planner in the Midwest. He told me, “We have 50 GW of data center requests, but I know that 40 GW will never break ground. Yet I have to plan for all 50 because the rules say so.” The rules are the problem. They reward announcement over execution. Until the grid operators introduce a financial commitment mechanism—a performance bond, a deposit that scales with capacity—the phantom load will persist.
Moreover, the carbon cost is real. Each ghost project delays the retirement of a coal plant or the construction of a renewable farm. The stranded assets are not just financial; they are emissions that stay in the atmosphere for decades. The AI industry’s narrative of efficiency and progress is undercut by the infrastructure it leaves behind, like a mining rig that never mined a single block.
Takeaway: The Next Narrative
The phantom load problem is not a reason to abandon AI infrastructure. It is a reason to demand transparency. What if every interconnection request were recorded on a public, immutable ledger—a blockchain for grid capacity? A PPA could be a smart contract, with penalties for withdrawal. The queue would become a market, not a lottery.
I have spent years bridging the gap between code and narrative. The next narrative is not about how much electricity AI will consume. It is about how we verify that the demand is real. When the pool empties, only the intent remains. The intent is to build a future that is accountable to atoms, not just stories. The ghost can be exorcised, but only if we audit the queue itself.
