One Grid Operator Names Its AI Load. Here Is What Sits Under the Name.

In July we wrote about a one-year ban that rests on a question no public record can answer: of everything queued up to be built, how much is real? Two weeks after that order, the operator of New York’s grid published a new edition of its interconnection queue workbook. It is, in the public queue records we have read, the only U.S. file that names AI data centers as a class. The code is DAT-AI, and the file’s own key defines it: “AI Datacenter”. It sits beside separate codes for general data centers, cryptocurrency mining, and microchip fabrication.

Here is what sits under the name, dated. In the edition of July 31, 2026, 5,290 MW across 12 requests carry the DAT-AI code — 39% of the 13,568.9 MW that NYISO publishes as classified load in its queue. The breakdown is the operator’s own table, titled MW By End-Use: general data centers 5,877 MW (43%), AI data centers 5,290 MW (39%), microchip fabrication 1,056 MW (8%), cryptocurrency mining 745 MW (6%). The operator states the scope of that total in the table itself: it excludes withdrawn and in-service load projects. The file does not say who assigns the code to a given request, so we do not call it self-reported; we report the label as the file gives it. Every figure recomputes from the operator’s published file, and the full derivation is public at /ai-load.

Two caveats travel with the number, and they are the useful part.

The first is age. 95% of the AI-datacenter class entered the queue in 2025. The cryptocurrency-mining class entered entirely between 2018 and 2021. The comparison everyone will want to make — which class is “serious” — is the one this file cannot support: a request filed in 2019 has had years to move through studies that a request filed in 2025 has not yet had. Any maturity comparison across these classes measures the age of the request, not the credibility of an industry.

The second is the signature. NYISO’s status ladder names the executed interconnection agreement explicitly — status 11, “IA Completed”. In this edition, no load request sits at 11, and the IA Tender Date column is empty on all 73 rows. Every request under the AI code sits at or before the reliability-study stage, before cost allocation. The rows the file marks as under construction carry no AI code — which, given the filing dates, is what you would expect. It is also the point: even the one file that names AI load cannot tell you which of these requests is firm.

That gap is no longer academic, because states are now acting on the totals. New York’s Executive Order 62 paused state discretionary environmental permits for data centers of 50 MW or more on July 14, and the order itself cites nearly 12 GW of data center load requests in the NYISO queue as of May 2026, with more than 8 GW entering in 2025 alone (the order); it directs studies on interconnection but does not pause it. Texas moved three weeks later from the other direction: on August 3 the Governor directed an audit of the data centers in ERCOT’s interconnection queue, citing roughly 474 GW of large-load requests, about 90% of them data centers, and ERCOT delayed its Batch Zero review while the audit proceeds (Utility Dive, Aug. 5). Two states, two instruments — a pause and an audit — and the same missing denominator under both: no public record we have read grades which large-load requests are firm.

The federal record has the same shape. FERC’s June 18 order in the PJM proceeding (Docket EL26-67-000) and its ISO-NE companion (EL26-72-000) opened proceedings on how large loads connect; neither specifies a public, graded status field for large-load requests. We read both orders in full before writing that sentence. In four of the seven queues we read, generation has carried such a field for years — application, study, executed agreement, in service — and it is what lets anyone verify the signed floor we publish. Load, the side driving the argument in Albany and Austin, still has no equivalent.

NYISO’s file shows what one operator’s disclosure can do: name the AI slice, date the megawatts under it, and publish enough structure to check the arithmetic. What it still cannot do is say which of those megawatts are firm. Until that field exists, we publish the part that can be checked — the edition, the source, and every step of the read, at /ai-load.

The reading above is tied to the edition dated July 31, 2026, and will be re-read against every new edition the operator publishes; cite it with its date. We are descriptive today: the predictive track record does not resolve until about December 2026, and our pages say so. Don’t trust — verify.