Sequelograph

A real event, read carefully. Then a fiction that does not pretend to be the news.

The record, 16 Sep 2026

Gartner forecasts $2.67 trillion in AI spending for 2026

On September 16, Gartner forecast $2.670 trillion in worldwide AI spending for 2026, including $1.484 trillion in AI infrastructure. Gartner's category includes AI-optimized cloud services, servers, network fabric, processors and devices, so it is broader than spending on data-center power alone; these figures are forecasts, not a tally of completed purchases.

Read the source at Gartner

The record ends here. Everything below this line is invented.

What if a city could schedule AI workloads like water through a reservoir?

Part one

The Switchyard Hour

At 18:02 in 2035, the county grid map turned amber around the east feeder. Tala rolled the dispatch wheel back a notch and watched the transformer trace climb anyway. Across town, a training campus had promised to shed twelve megawatts on request. Its server meter had dropped by ten. The feeder meter had dropped by two. Behind the campus's pale windows, cooling towers kept their steady white breath.

Tala had twenty minutes before the feeder crossed its operating limit. The town seed library sat on the same branch, in a low brick building where her mother kept jars of local corn and bean seed in a refrigerated room. Her mother had called at lunch to say the compressor was making a new rattle. If the library lost power, it had a backup, but the old generator needed a person to start it. Tala's brother was the only one with the key, and he was driving home from the coast. She had promised their mother she would keep the lights steady through the evening.

The campus's contract called the training load flexible. Tala had believed the word. She had also believed a server meter described the whole facility. She asked the campus operator to pause the next training batch, then pulled the feeder's live data into a clean window. The computing load fell again, this time by three megawatts. The campus still drew almost as much as before. The chilled-water pumps were filling their tank back to the top.

"You sold me twelve," Tala said.

"We gave you twelve at the IT meter," the operator replied. "The cooling plant is holding temperature. We can't shut that off with people in the building."

Tala put the two readings on the same clock. The campus meter sampled every second; the feeder meter averaged over five minutes and included the pumps on a neighboring substation. She checked the campus's battery schedule, then the feeder drawing. The operator had reported a real drop, but it was not the drop the contract promised. She asked for the chiller's next restart time and heard a keyboard clatter on the other end of the line. The load would rise again in eleven minutes, exactly when the dispatch model expected the branch to be cooling.

The answer was reasonable. The load sheet was not. It treated the compute racks as though they sat alone in a field. Tala pulled up the facility diagram. The training jobs used most of the power at the racks, but the campus also needed chillers, pumps, batteries and network equipment. Some of those loads could move in time; others had to stay. She had been asking for a number at the wrong boundary.

Her mother's seed room pinged the local meter: temperature stable, compressor on. Tala opened the substation thermal model. The amber band would clear if she could shave six more megawatts before the next dispatch interval. The campus could not simply stop cooling, and the utility could not tell a seed refrigerator to wait. She sent the operator a narrower request: hold the training batch, stop charging the battery, and coast the cooling loop on the chilled-water tank for twelve minutes. Keep all inference traffic and the safety systems live.

The campus operator paused to check the water temperature. Tala used the quiet to inspect the old dispatch table. Its model had multiplied the IT-meter reduction by a fixed efficiency factor, as if every saved watt also saved the same share of cooling. But the cooling system had its own momentum. The water in the tank could carry the cold for a while; the pumps could restart later, when the grid had room. She moved the next set of background jobs to the night window, when the feeder usually ran cool.

The operator agreed. A row of small blue bars disappeared from the load chart. The feeder line bent down, slowly, and held. Tala watched the temperature estimate fall beneath the limit. She asked her mother to read the seed-room probe.

"Still at four degrees," her mother said. "And I can hear the compressor again."

A pause followed, filled by the seed room's hum over the call. Her mother had packed those jars with neighbors, each label marked in a different hand: blue beans from a farm that no longer flooded, corn saved from the school garden, peas from a kitchen window. Tala could not send the grid a photograph of them, but she could see every jar when the room's temperature alarm flashed.

Tala laughed once, quietly. The room had bought them another hour. She saved the dispatch rule: training could pause, cooling could coast only while the tank stayed below its limit, and the campus would report the entire facility meter rather than one rack. A data center's load includes the machines that keep the machines cool. The sentence looked plain on the screen, but it changed the unit Tala would ask for the next time someone offered flexibility: megawatts at the point where the utility actually measured them, for a stated duration, under a stated cooling condition.

Then she opened the county reserve map. Eleven campuses had offered flexible capacity. Their blue bars stacked neatly beside one another, as if each could pause at once. Tala looked at the timestamps: every bar used the same modeled cooling reserve, the same tank volume, the same quiet hour. On the map, the eleven promises added up to a reserve the grid had never actually seen.

What’s real

  • Gartner forecast $2.670 trillion in worldwide AI spending for 2026.
  • Its forecast assigns $1.484 trillion to AI infrastructure.
  • Gartner's infrastructure category includes servers, networks, processors and devices.

What’s invented

  • Tala, her family, and the county seed library are fictional.
  • The training campus and its grid contract are invented.
  • The reserve map's double-counted cooling capacity is speculative.

Part two: A Reserve You Can Measure

Tala has one tested twelve-minute cut and eleven untested promises; she must learn what the grid can count on.

Part two isn’t on sale yet. Check back soon.

Filed under technology, data centers, power grids, energy.

More records

How it works

Each record starts with one real science event, with a link to where it was reported.

A short story then asks what happens next. Part one is free to read.

Part two costs a few dollars, paid once. You keep it.