Which power is real

Bottleneck Labs reads public grid records to separate real power from paper power before money, sites, and timelines get committed. A data center site, a queue position, a financing memo, or a power-backed development pitch eventually reduces to the same question. Will the megawatts actually arrive, and when?

The answer usually appears early. It sits in filings, study reports, network-upgrade assignments, and queue postings that most people only skim. We read those records, write down which projects look likely to survive, which projects look likely to exit, and where room to connect is opening. The public record grades those calls as new postings arrive.

This matters because the constraint on the buildout has moved past chips and capital. The hard edge is the distance between promised power and real power. Anyone deciding inside that distance needs a way to check the story before it becomes an expensive commitment.

Graded in public

Numbers before adjectives, so every claim on this page can be checked against the record. PJM runs the largest electricity market in North America, serving sixty-five million people and the heaviest concentration of data centers on earth. In June 2026 it reposted the network-upgrade costs for the two hundred seventy projects surviving in its transition queue, the event that decides who keeps paying and who walks away.

Written in advance

Before the June posting, we wrote down which connection points would live and which would die. Twenty-four of twenty-four came back correct, including one corridor where we called the split on both named sides. We named which side keeps its projects and which side loses them.

Misses, published

The same test produced two wrong reads. They stay in the record, with the mechanism that fooled us named. The misses show where the read needed repair.

Being graded now

Ahead of this cycle's commitment deadline, we sealed keep-or-quit calls on all two hundred seventy projects. The exits we expect carry roughly 4,200 megawatts and $2.3 billion in assigned costs. The postings that follow will grade every one of those calls over the coming weeks.

What the record taught

Our finding runs against market instinct. When neighbors quit a crowded queue, the survivors usually get cheaper. At the June posting, one 1.3-gigawatt hub watched its assigned costs fall from roughly $280 million to $42 million as the crowd around it thinned. Most of the market assumes the opposite. We side with the record.

Published reads

These pieces show the current focus. Each one follows a power claim through the grid records that either support it, weaken it, or change the decision around it. Together, they show what gets checked before a power claim becomes a deal assumption.

• The Grid Test Behind the AI Power Story2026

  America's most famous nuclear-powered data-center deal, followed through the grid's own paperwork. The piece shows what the power story proves and the test every AI power claim has to pass.

• Position Tax2026

  A whole queue generation, measured through the cost of holding an interconnection position. It shows who paid, who survived, and what the position was really worth.

• Where the Next Failures Will Form2026

  The current queue generation, read corridor by corridor. It shows where the risk concentrates and which projects looked exposed before the next postings arrived.