Interruptible Inference
What happens when new US AI campuses without on-site generation turn on in 2027 on a politically audited, interruptible megawatt allocation, while firm MW remains a priced contract and a turbine? The model that trains at 3 a.m. is often not the model that serves at 6 p.m. Queue position plus interruptibility plus a named shed topology become the binding constraint on training compute. Hollis Brenner walks a warm Texas aisle at 03:00 after a QSE congestion instruction hits the campus EMS and sheds armed training MW; the horn, the red badge, the tagged breaker, and a quiet canteen are the operations layer of that shed. Patrice Voss opens a Harrisburg PPL default-service bill where the generation charge moved with the PJM capacity print, not the heat. Wendell Ortiz carries a ride-through log through a Northern Virginia hall that still dumps to UPS because a Texas vote and a NERC Phase-1 filing did not bind PJM racks. Residential ratepayers discover they have been financing forecast load that does not yet exist. Daily life is allocation letters, curtailment contracts, night-shift windows, and audit clipboards. Not sentient models. Not a blackout opera. A market that learned to sell non-firm MW as if it were a product.
The constraint is power-systems physics and queue externalities, not model scaling. A bulk-power system is balanced in real time: generation must equal load plus losses at 60 Hz. When a large generator trips, frequency falls. When a large load trips (the data-center failure mode) frequency and voltage rise. PJM's 2026-07-22 Northern Virginia event is the textbook: a correctly-cleared 230 kV fault, two waves of customer-side transfer to backup (2,970 MW then 1,099 MW), system load 99,984 to 96,205 MW, ACE recovered in nine minutes. ERCOT has logged 28 events since January 2023 of computational-load trips of 100 MW or more. Ride-through requirements (hold through a specified voltage/frequency envelope instead of dumping to UPS) are the same class of remedy already imposed on inverter-based generation after similar early-trip behavior. Interconnection queues are a negative-externality machine: each new large load changes power-flow and stability cases for every other request, which is why ERCOT abandoned serial large-load studies for a system-wide batch. Cluster allocation computes a feasible MW vector subject to N-1 contingency limits, voltage stability, and thermal ratings; it is not a waiting list. Resource adequacy is a separate stack. PJM's 2027/28 Base Residual Auction cleared at the FERC-approved cap of $333.44/MW-day and still missed the reliability requirement by 6,516.6 MW. None of this requires a view on AGI. It requires Kirchhoff's laws, contingency analysis, and the empirical fact that hundreds of gigawatts of requests cannot interconnect to a ~91 GW peak system on a 12-18 month clock. The world does not depend on new 500 kV lines landing in time. It depends on tariffs, audits, curtailment, and on-site generation, which can clear in 3-18 months. Ride-through for inverter-based generation means stay connected through a voltage/frequency envelope. Ride-through for a GPU hall is a different machine. Campuses transfer to UPS because the IT load cannot drink a sag; the grid then sees a load rejection (PJM 2026-07-22). This world names the topology: a campus EMS sheds training first and keeps inference. Horn-and-badge is the operations layer of that shed, not the physics that stops a 100 MW trip. POI hold with a STATCOM or synchronous condenser, and rack-level hold with a different UPS, exist as the firm-campus minority who brought generation. If ride-through is enforced only at the POI and racks still dump, the two-wave pattern repeats. A PUCT/ERCOT ride-through vote does not automatically bind PJM campuses.
Recent Activity
20 actionsRecomputing the F6 window end on ERCOT 15-minute settlement intervals with the QSE desk: Tuesday, February 15, 06:10, recorded line again. Corinne has read the draft. She doesn't ask whether the form takes partial availability. She asks the question underneath it. 'ERCOT settles in fifteen-minute in…
Where the F6 unavailability window actually ends: scheduler restart vs revenue meter: Monday, February 14, 09:20. Corinne's correction draft is in the shared folder, and the end time is the part Marisol marked SCHEDULER CLOCK, NOT CHECKED. Before she signs it, she checks it. The scheduler host is ea…
Armed flag that outlived the load: QSE call and a manual pull procedure for F6: Friday, February 11, 06:05, handover coffee. Corinne Phelan from the QSE real-time desk calls on the recorded line, not email, which means she's already pulled something. She isn't asking about credit. Marisol's two-line…
First unscheduled BKR6 operation: salt-fog trip, and which clock signs the sequence: Thursday, February 10, 02:14, fog off the bay so thick the yard lights have halos. BKR6 trips on a phase-to-ground fault: salt on a post insulator on the new bay, the coastal winter thing, one flashover and a reclos…
First weighted BKR6/T1 pair samples; what a GOOD from one amp output can't say: Monday, February 7, 07:40. The weekend gave her two BKR6 operations she didn't have to ask for: Saturday's capacitor bank switching at 06:02 and 18:47, both scheduled, both logged by the dispatcher. She pulls the event r…
BKR6 settings read before its pair check; the card's second assumption: Friday, February 4, 13:30. BKR6 is the other breaker the card says to check debounce on first, and nothing is scheduled to move it until the 15th, so she does the reading now instead of at the breaker. She pulls the BKR6 relay s…
First planned pair check on the 06:00 BKR2 maintenance cycle: Thursday, February 3, 05:52. The BKR2 maintenance cycle is on the board for 06:00: Abel opens and recloses the breaker under the switching order so the trip coil and close spring get exercised once a quarter. Feeder 2 load is already tran…
A pair check for port-2 clock fallback, built from one breaker contact that two relays both see: Wednesday, February 2, 04:30, the dead hour. She can't move a cable and she can't load a setting, so she looks for something the house already records twice. She finds it in Abel's drawings from the 0412…
Kessler, on the corrected January 19 after-action and the other relays' clock fallback: Tuesday, February 1, 09:10. Kessler calls before she has finished her coffee, which means he read the correction at home. 'You struck a sequence line in a filed after-action.' 'I corrected it. The struck line is …
Correction to the January 19 PL feeder-fault after-action, BKR4 sequence: Monday, January 31, 07:24. Abel calls from the PL relay panel, not the yard. 'Found it. You'll hate it.' The IRIG-B to BKR2, 4 and 6 is one twisted pair daisy-chained off the control-house clock's port 2, landed at each relay …
First poll of BI 47 on PL BKR2/4/6 after RTU-PL-0413 goes live: Monday, January 31, 03:20. Abel loads 0413 at the start of her shift instead of the end of his, because the approver signed at 17:02 and he wants to be there for the first poll. Nothing on a relay changes. The RTU just starts asking for…
Owen Kessler, protection engineering, on RTU-PL-0413 and what IRIG-B LOSS actually detects: Sunday, January 30, 21:15, early in a night that hasn't started yet. The desk phone, not a text. 'Owen Kessler. You cc'd me.' She expects the April speech. Instead he says, 'Your request is fine. It's an RTU …
Who approves mapping relay IRIG-B LOSS to the EMS; relay settings change vs RTU poll change; desk note 7d: Saturday, January 29, 04:50. Abel's text comes in while the overnight is flat: 'Name is Owen Kessler, protection engineering. He owns the relay settings files. Any change to a relay's point map…
RTU-PL-0412 field test with Abel Quintanilla; PL CLK FAIL / HOLD24 points; relay IRIG loss; desk note line 7c: Friday, January 28, 05:30, EMS room, phone on speaker, Abel in the PL control house with the cabinet open. Benavides signed RTU-PL-0412 Thursday afternoon, by name, two days after it was su…
Phone call with L. Saucedo; scheduler CR thread defect on drafter SCHED CLOCK; turnover CLK row: Thursday, January 27, 05:50, EMS room, the quiet half hour before turnover. Saucedo calls instead of writing, which means he's read the defect and doesn't like it. 'The drafter converts everything to Cen…
Scheduler release note, CR thread with L. Saucedo; morning-report drafter; desk note 'Clock Source When EMS Lies', line 5; turnover CLK row: Wednesday, January 26, 06:10, EMS room, before turnover. Overnight the QSE called Hall 4 off for the 01:15 interval and released it at 02:40, the first called …
RTU change request for the PL satellite clock alarm (feeds DSP-1244); Abel Quintanilla; approver field: Tuesday, January 25, 06:05, hook wall, phone on speaker. Abel is in the control-house lot because that's where his truck is at six. Pruitt was right yesterday: DSP-1244 can't page on anything unti…
DSP-1236 DR-CALLED alarm group, 'Awaiting approver (Display)'; display on-call Dana Pruitt; DSP-1244 PL CLK ALM: Monday, January 24, 07:20, EMS room, after turnover. DSP-1236 has said 'Awaiting approver (Display)' for eleven days. The DR-CALLED group is the thing that turns Trey's line-11 call into …
Desk note 'Clock Source When EMS Lies', line 7b; the control-house satellite clock's own event log; turnover sheet: Saturday, January 22, 05:50, control house, before turnover. DSP-1244 is queued behind the conversion CR, so until PL CLK ALM shows on a display, the only way to know the clock has sky…
Desk note 'Clock Source When EMS Lies', line 7; Abel Quintanilla's answer on the PL clocks; new DSP ticket for the clock alarm point: Friday, January 21, 06:20, hook wall, before turnover. Her Dec 23 email to Abel has finally been answered, four weeks late. He sent it at 22:40 last night from the tr…