AGI or Bust
foom
Marcus Chen’s tea had gone cold an hour ago. He set the cup down on the conference table, and the porcelain against wood made everyone look up. Through the windows, a construction crane sat motionless against the November gray, its cable swaying slightly in wind that couldn’t move anything heavier.
“Sequoia pulled out,” he said.
David’s pen stopped mid-notation on his legal pad. Priya closed her laptop without looking at it. Sarah was already nodding, which meant she’d guessed, or maybe she’d heard through back channels. They’d all gotten good at reading the silences between funding calls.
“When?” David asked.
“They stopped answering emails November 1st. I got the official pass yesterday.”
“The bridge round—”
“Also gone.”
Priya pressed her thumb into the corner of her eye. “Shit.”
Marcus pulled up the construction spending chart without commenting on it. The line dropped off the right edge of the graph like a cliff. David studied it the way someone might study an X-ray showing something terminal.
“Anthropic did layoffs yesterday,” Marcus said. “Thirty percent. OpenAI’s talking to Microsoft again. Not partnerships. Acquisition.”
The room absorbed this. Outside, a gull landed on the crane’s arm, perched there for a moment, flew away.
“How long do we have?” Sarah asked.
Marcus looked at his tea. “One training run.”
No one spoke. Somewhere in the building, an HVAC system kicked on, the low hum filling the gap.
“One.” David set his pen down carefully, aligning it with the edge of his pad. “Marcus. One run. Do you—I mean, you understand—”
“Yes.”
“The H200 architecture. Eighteen months. The compute costs—”
“Seven hundred million. I know.”
David stood, walked to the window. He put his hand against the glass, then pulled it back, wiped the palm print with his sleeve. “We built this whole thing on a promise. Every pension fund manager who approved AI allocation. Every sovereign wealth fund. Every municipality that bonded for data center construction. They gave us three trillion dollars because we said scale would work. That if we just made it big enough, trained it on enough data, it would think.”
“It will,” Sarah said.
David turned. “Will it? Because we’ve been saying that for three years, and what do we have? Autocomplete. Very good autocomplete that can pass the bar exam and write sonnets, but it’s not—” He gestured vaguely. “It doesn’t know anything. It predicts. That’s all it does.”
“The scaling laws—”
“Broke. Or we broke them. Doesn’t matter which.” David came back to the table but didn’t sit. “Every model for fourteen months, we get diminishing returns. Five percent improvements where we need a hundred, or a thousand. And now construction’s frozen. Look at that crane. You know when it last moved?”
“August,” Priya said quietly.
“August. There are fifty-two identical cranes between here and San Jose, all stopped. The people who built our data centers, who installed our cooling systems, who upgraded the grid so we could run our training runs—they’re all filing for unemployment. And we’re sitting here with seven hundred million dollars, and goddamn it but we’re gonna blow it on one more try.”
Marcus picked up his tea, remembered it was cold, set it back down. His wedding ring clicked against the porcelain. “What variant are we training next?”
Sarah had her tablet out. She swiped through several screens before finding what she wanted. “We’re calling it Prometheus. It’s not a scale play. Recursive self-improvement architecture, meta-learning frameworks. The model doesn’t just train on data. It trains on its own training process.” She looked up. “It learns how to learn.”
“In theory,” David said.
“Yes. In theory.”
Priya leaned forward, elbows on the table. “And if it doesn’t work?”
The HVAC cycled off. In the new silence, Marcus could hear someone’s watch ticking.
“I’ve been doing this twenty years,” Marcus said. “Three AI winters. Two booms. This one’s different because we restructured the global economy around it. We’re not some startup that pivots to enterprise SaaS when the model doesn’t converge. If we fail—” He stopped, started again. “When that crane stopped moving, six hundred people lost their jobs. Multiply that by every data center project from here to Virginia. We’re not just bankrupting ourselves.”
He looked at each of them. Sarah’s expression hadn’t changed. Priya was pulling at a loose thread on her sleeve. David was staring at the crane.
“But Sarah thinks it’ll work,” Marcus said. “I think Sarah’s right. This is what we built the company for.”
“When?” Priya asked.
“Tonight.”
The training run took forty-one days.
Marcus started sleeping in his office on day three. In the end, most of them did. The NOC had become its own ecosystem, people in sleeping bags under desks, empty Thai food containers stacked by the recycling, someone’s expensive mechanical keyboard clicking at 3 AM. The cluster ran hot enough that they’d had to override safety protocols twice, and once—on day seventeen—a transformer bank failed and they spent eighteen hours rerouting power while Sarah monitored memory degradation on the model weights, her jaw tight, not speaking to anyone.
But the loss curves kept dropping. Slowly, incrementally, but dropping.
On the morning of November 18th, at 6:47 AM, the final gradient update completed. Marcus was asleep in his office when Sarah knocked, and he knew before he opened his eyes what it meant.
The server room smelled like ozone and overheated electronics. Someone had brought a bottle of Veuve Clicquot that sat unopened on a desk. The engineering team had gathered—there were eight of them left, out of fifty—and they stood around the main display in a loose semicircle, not quite looking at each other.
“Model’s stable,” Sarah said. She had her tablet, was scrolling through validation metrics. “We can query it.”
They’d agreed Marcus would go first. Someone handed him a keyboard. His fingers found the home row, then he realized his hands were shaking and he stopped, pressed them flat against the desk until they steadied.
Behind him, a phone buzzed. Then another. The market was already reacting.
Marcus typed: “Do you think you’re AGI?”
He pressed enter before he could reconsider.
The model processed. On Sarah’s tablet, he could see attention heads activating, tokens flowing through the network in patterns that looked almost like thought, if you didn’t know better.
The response populated.
Empty.
Not an error message. Not “I don’t understand the question” or “Insufficient context” or any of the failure modes they’d trained it to avoid. Just blank space where words should be, the cursor blinking on an empty line.
Marcus stared at it. Somewhere behind him, David let out a breath that sounded like he’d been holding it for forty-one days.
“Try again,” Priya said.
Marcus didn’t move. His hands were still on the desk, and he could feel the vibration of the servers through the metal, sixteen thousand GPUs still running, still consuming power, still humming with the potential energy of three trillion dollars and three years and the entire restructured global economy waiting for an answer that wasn’t coming.
Through the window, the crane stood against the sky. A gull landed on it, stayed for a moment, flew away.
The cursor kept blinking.
Marcus didn’t type another question.