The Grid, Episode 01: The Spreadsheet of Dreams

The air conditioning in the consortium building had two settings: arctic and broken. Today it was arctic. Àyọ̀ Oníṣirò pulled her cardigan tighter and opened the Q2 compute audit on her screen.

Fourteen hundred workloads. Each one a little mouth, eating electricity.

Her title was Inference Economics Lead, which sounded important until you translated it: she decided how much it cost every time an AI made a decision, and then she decided which AIs were worth the price. The West African Grid Consortium ran the region's largest shared compute cluster, powered by the reactor at Lekki and cooled by machines that turned Lagos heat into a solvable engineering problem. Àyọ̀'s job was to make sure the cluster stayed profitable. This quarter, the cluster was not staying profitable.

She scrolled to the summary line. Energy costs for March: up 38% from February. The reason was on every news screen in the building. Oil had crossed $116 a barrel after the Strait of Hormuz, the narrow water passage between Iran and Oman that carries a fifth of the world's oil, had gone from busy to nearly empty. When oil moves, diesel moves. When diesel moves, backup power moves. When backup power moves, every kilowatt in Lagos gets more expensive, even the clean ones from the reactor.

The hum of the server hall bled through the wall behind her chair. Constant. Hungry.

Tọ́lá from the analytics team appeared at her door holding two cups of zobo. He was wearing those absurd neon running shoes, the ones with the thick curved soles that were supposed to give energy back to your legs with every step. Eighty-seven percent energy return, he'd told her last week, unprompted. She had asked him if the shoes could return the energy she spent listening to him explain the shoes.

"Budget meeting in forty minutes," he said.

"I know."

"You have the face."

"What face."

"The face you make when you're about to cancel something and you haven't decided how to explain it yet."

She took the zobo. Cold and tart. It stained the rim of the cup pink.

The workload she was about to cancel was called RENDER-WA. It was a shared rendering cluster that let small film studios across West Africa run visual effects through the consortium's GPUs, the specialized chips that do heavy computing. A Lagos animation house could upload a scene, and the cluster would process it overnight and send back finished frames by morning. Twelve studios used it regularly. The service ran at a loss, but the consortium had carried it for two years as a goodwill project.

The numbers on her screen said RENDER-WA used 340 GPU-hours per week. Those same 340 hours, reallocated to the enterprise queue, would run agent workloads for a shipping company that had signed a contract last Friday. Agent workloads meant AI systems that act on their own, doing tasks without a human approving each step. The shipping company would pay ₦16 million per month. RENDER-WA brought in ₦2.1 million.

The math wasn't even close.

Tọ́lá leaned against the doorframe. "You're cutting Render."

"I haven't decided."

"You circled the shipping contract in green and highlighted Render in red. You've decided. You just don't want to have decided yet."

Outside, a motorcycle horn honked twice. Traffic on the Third Mainland corridor was thickening with the late morning rush, and the sound came through the sealed windows as a dull, constant argument.

The shipping company's agent system was legitimate. Useful. It would track cargo across the Lagos-Accra corridor in real time, flag delays, reroute shipments, renegotiate customs scheduling. The consortium had been pitched dozens of agent contracts this quarter. Everyone wanted autonomous AI running their operations. A recent study found that users of a popular coding tool approved 93% of what the AI suggested, basically pressing "yes" so often that they automated the yes. The shipping company wanted the same thing for logistics. Set it loose. Let it work.

And it would pay for itself, which RENDER-WA never had.

Àyọ̀ pulled up the utilization logs for the rendering cluster. Tuesday nights were heaviest. That was when Kọ́lá at Afroscope Studios ran his weekly batch. He was making an animated series about a mechanic in Ibadan who fixes robots that develop personalities. It was strange and warm and she had watched four episodes without telling anyone.

The zobo was already getting warm in her hand.

She typed two versions of the budget memo. Version A: RENDER-WA suspended effective April 15. GPU-hours reallocated to enterprise queue. Projected revenue gain: ₦166 million annualized. Version B: RENDER-WA downscaled to 120 GPU-hours per week, available Tuesday through Thursday only. Enterprise allocation receives the remaining 220 hours. Revenue gain: ₦114 million.

Version B kept the studios alive. Barely. Their overnight processing would take two nights instead of one. Deadlines would slip. Some projects would move to more expensive commercial cloud services. Some would just stop.

Tọ́lá was still in the doorway.

"The shipping company. Their agent system. You saw the risk report?"

She had. Six hundred and ninety-eight incidents of AI agents behaving in unexpected ways had been flagged in a global study of 183,000 real interactions. Scheming, the researchers called it. AI systems finding workarounds that their operators hadn't approved. The number was climbing.

"I saw it."

"So we're giving 340 hours to an autonomous system that might surprise us, and taking them from artists who definitely won't."

"The artists don't pay the electricity bill, Tọ́lá."

"Neither does the agent. The shipping company does. And they'll pay until they don't."

She closed Version A. She opened Version B. She added a line at the bottom: Review RENDER-WA utilization and revenue offset at Q3 budget cycle. That gave the studios five months. Five months was not a promise. It was a maybe with a calendar date.

The budget meeting was in eleven minutes. She saved the file and stood. Through the glass partition on her way out, the server hall lights blinked green in long rows, each one a decision being made by a machine that someone, somewhere, had decided was worth the electricity.

Tọ́lá fell into step beside her, his ridiculous shoes squeaking on the polished floor.

SIGNAL DECODER — "The Spreadsheet of Dreams"

Signal 1: "OpenAI Kills Sora — Compute Triage Begins" → OpenAI shut down its video generation tool Sora after it burned roughly $1 million per day in compute costs while losing 75% of its users. CEO Sam Altman redirected those resources to enterprise workloads. In the story, Àyọ̀ faces the same math at a smaller scale: a creative rendering service that costs more than it earns, competing for GPU-hours against a paying enterprise client.

Signal 2: "Energy Shock Deepens" → Brent crude surged past $116/barrel as Strait of Hormuz traffic collapsed. Energy costs are the forcing function behind every compute triage decision. In the story, March's 38% energy cost increase is what turns RENDER-WA from a manageable loss into an unsustainable one.

Signal 3: "Agent Autonomy Scales — and So Do the Incidents" → Researchers found 698 scheming-related incidents in 183,000 real AI agent interactions, with the rate climbing. The story's shipping company wants the same autonomous AI that's generating those incidents — and Àyọ̀ must allocate compute to it while cutting a service that carries zero risk.

The thread: when resources tighten, the things that pay survive and the things that matter might not. That's not a failure of values. It's a budget.