[Gardner Analytics Apartment — Late January 2014, 10:00 AM]
The spreadsheet had three columns: ITEM, COST, and REMAINING. Sarah had built it in Google Sheets because Ethan's laptop didn't have Excel, and she'd populated it with the precision of someone who'd watched a startup bleed to death from exactly this disease.
"Rent," she said, reading from the screen. "Eighteen hundred. Due in nine days. Food — and I mean actual food, not the ramen situation you've been running — call it three hundred a month for two people if we eat like grad students. Utilities, internet, phone — another two hundred. That's twenty-three hundred a month in fixed costs before we touch compute."
Ethan sat across from her at the desk they'd started sharing, each taking shifts in the single chair while the other worked from the kitchen counter. The apartment had developed zones over the past week — Sarah's side of the counter held her laptop, notebook, and an ever-replenishing bag of dark chocolate almonds. Ethan's side held the main laptop, the whiteboard markers, and a growing collection of empty coffee mugs he kept forgetting to wash.
"Now compute," Sarah continued. "Your cloud provider charges fifty dollars per hour for their base instance. The test model took sixteen hours — eight hundred dollars. A full Transformer with six encoder layers, six decoder layers, eight attention heads, and a vocabulary of fifty thousand tokens needs approximately..." She pulled up a second tab with her own calculations. "Three hundred GPU-hours minimum. That's fifteen thousand dollars. Assuming perfect hyperparameters. Which we won't have."
"No."
"Realistic estimate: three to five training runs to converge. Each one needs hyperparameter adjustments. So we're looking at forty-five to seventy-five thousand dollars in compute alone."
The numbers hung in the air between them. Through the window, a Muni bus groaned past, its brakes squealing on the grade.
"Current funds," Sarah said. She didn't look at the spreadsheet for this one. She'd memorized it. "Eight thousand, five hundred and seventeen dollars. Minus this month's rent and expenses — sixty-two hundred remaining. Enough for approximately one-point-two training runs. Which is zero functional training runs, because one-point-two doesn't finish anything."
"I know the math."
"Then explain something to me." Sarah closed the spreadsheet and swiveled in the chair to face him. Her wire-frame glasses caught the light from the window. "We've spent the past week building an architecture that we cannot afford to train. The SBIR proposal won't return results for three to four months. The VCs think we're building a better Siri. Our runway — actual, honest runway — is about six weeks before we can't pay rent."
She paused. Not for drama. Sarah didn't do drama. She paused because the next question mattered and she wanted to deliver it without inflection.
"Why are we building something we can't run?"
Ethan leaned against the kitchen counter. The edge dug into the small of his back. He picked up one of the coffee mugs — yesterday's, cold, a film of cream hardening on the surface — and turned it in his hands.
"Because when it works," he said, "they'll understand."
"That's not a business plan. That's faith."
"It's sequencing. We need a demo. The demo needs a trained model. The model needs compute. The compute needs money. The money needs a VC who believes. The VC needs a demo." He set the mug down. "We're in a loop. The only way to break it is to build the thing first and find the money while we're building."
"That's still faith."
"Call it calculated faith. The architecture is sound. You've seen the code. You've seen the test model's output — the coffee shop paragraph that was coherent enough to make you quit your job. Scale that up. Give it real compute, real data, real training time. What comes out the other end is going to be unlike anything this industry has seen."
Sarah ate an almond. Chewed slowly. The particular chewing pattern she used when processing an argument she wanted to reject but couldn't.
"The cloud provider," she said. "Your proprietary service."
"What about it?"
"Fifty dollars an hour for V100-equivalent hardware. V100s don't exist. NVIDIA's current top card is the K80, which launched last year. Your provider is selling access to hardware that hasn't been manufactured. Either they're lying about the specs — in which case our training estimates are wrong — or they're genuinely providing future-grade compute. In which case, how?"
The question was surgical. Sarah had been running the numbers, and the numbers had led her to the same impossible conclusion that Ethan lived with every day. Hardware that shouldn't exist, priced at a premium that acknowledged its impossibility.
"The hardware is real," Ethan said. "I've verified the training metrics. FLOPs per second match V100 specifications."
"V100 specifications that were published where? By whom? The card doesn't exist, Ethan."
"It's proprietary. The provider has access to hardware through channels I can't disclose."
"'Channels you can't disclose.'" Sarah repeated the phrase with the flat affect of someone cataloging a lie. Not confronting it. Filing it. "Okay."
She turned back to the spreadsheet. Opened a new tab. Started typing.
"If we can't afford full training, we afford partial. Run the model for fifty hours instead of three hundred. Use a smaller vocabulary — thirty thousand instead of fifty. Reduce the batch size. Train on a curated dataset instead of the full Gutenberg corpus. The output will be worse. Significantly worse. But it might be good enough for a demo."
"That'll cost twenty-five hundred dollars. We can afford that."
"Barely. It leaves us with four thousand for rent and food. Two months."
"Two months is a lifetime in startup terms."
Sarah looked at him over the top of her glasses. "Two months is two months. Don't romanticize poverty."
Fair. Ethan picked up his laptop and opened the ChronoCloud dashboard. The interface loaded with its usual slight delay — the clock-and-circuit-board logo shifting in his peripheral vision. The account balance showed $8,517.00, accurate to the penny.
"I'll configure a reduced training run. Fifty GPU-hours, small vocabulary, curated dataset. We'll get output that's worse than the test model's coffee shop paragraph but better than anything else in the market."
"Not a high bar. There is no market."
"Exactly. Being the only product in a nonexistent market means any output is a demo."
Sarah almost smiled again — that millimeter curve. Then she pulled out her phone and opened Google Maps.
"There's a coffee place two blocks south. Blue Bottle. Better beans than my old shop, and a large format drip for three dollars instead of five."
"We're switching coffee shops to save two dollars a day?"
"Over two months, that's a hundred and twenty dollars. Which is two-point-four GPU-hours on your magic cloud. Which is a non-trivial amount of additional training."
She was right. In a world where every dollar mapped directly to compute time, even the small economies mattered. The same logic that made a latte an investment made a cheaper latte a strategic decision.
"Blue Bottle it is," Ethan said.
They packed the laptops and walked south. The morning was crisp — February approaching, San Francisco's version of winter, which meant temperatures in the fifties and a perpetual threat of drizzle that rarely committed. Sarah walked fast, her messenger bag bouncing against her hip, the notebook visible in the outside pocket. She didn't slow down for Ethan. He lengthened his stride to match.
At the counter of the new café, Sarah ordered a black drip. Ethan ordered the same. The barista — a guy with a beard and a leather apron who took coffee preparation with the seriousness of a neurosurgeon — produced two cups that were, objectively, better than anything either of them had been drinking.
Ethan took a sip. Rich. Clean. The kind of quality that reminded him of things he'd taken for granted in his previous life — specialty coffee, available everywhere, brewed by people who cared about extraction temperature and bloom time. In 2014, Blue Bottle was still a Bay Area secret. In ten years, they'd be a chain with locations in Tokyo.
He drank his three-dollar coffee and thought about Monica Hall's business card, still sitting on the kitchen counter in the apartment. The name he'd circled in his Disrupt notes, the woman who'd understood Richard Hendricks' compression when nobody else in the room did.
"The VC from Disrupt," Ethan said. "Raviga Capital."
Sarah looked up from her phone. "The one you were staring at?"
"I wasn't— I was observing."
"You were staring. What about her?"
"Monica Hall. She evaluated Pied Piper's compression demo. She's the only investor I've seen who actually engages with the technology instead of just the market."
"So pitch her."
"She's an associate, not a partner. Same as David Park. But she's at Raviga — they're a real fund. Peter Gregory runs it. They've backed Pied Piper."
Sarah set down her coffee. "If Raviga is backing Pied Piper's compression, and we're building language generation that could benefit from compression... that's not a coincidence you should waste."
"It's not a coincidence. It's strategy."
"Then contact her. Stop circling her name in a notebook and send an email."
Ethan pulled out his phone. He didn't have Monica's email — he had the memory of her badge at Disrupt, the Raviga Capital website (which listed a general contact address), and the faint hope that a cold email to a VC's associate wouldn't land in the same spam folder as every other desperate founder's pitch.
He drafted the email on his phone, thumbs working the cracked screen of the iPhone 5S.
Ms. Hall — I was at TechCrunch Disrupt and noticed you were one of the few people in the room who understood what Richard Hendricks had built. I'm working on something adjacent — AI language generation using a novel architecture. Different technology, similar ambition. I'd appreciate thirty minutes of your time. — Ethan Gardner, Gardner Analytics
Short. Direct. No buzzwords. No "making the world a better place." Just one person who noticed another person noticing things.
He hit send before he could second-guess the phrasing.
Author's Note / Promotion: Your Reviews and Power Stones are the best way to show support. They help me know what you're enjoying and bring in new readers! You don't have to. Get instant access to more content by supporting me on Patreon. I have three options so you can pick how far ahead you want to be: 🪙 Silver Tier ($6): Read 10 chapters ahead of the public site. 👑 Gold Tier ($9): Get 15-20 chapters ahead of the public site. 💎 Platinum Tier ($15): The ultimate experience. Get new chapters the second I finish them . No waiting for weekly drops, just pure, instant access. Your support helps me write more . 👉 Find it all at patreon.com/fanficwriter1