Anthropic, the AI darling currently saddled with eye-watering cloud bills, has found itself backed into a fiscal corner, forced to confront the reality that scaling a trillion-parameter model costs slightly more than free pizza and kombucha on tap. In a move born from necessity, innovation, or possibly sheer panic, Anthropic has turned to Craigslist, relentlessly hoarding old, cracked-screen iPhones to keep Claude from sputtering offline.

“Initially, we figured we’d just spin up a couple thousand GPUs in AWS,” confessed a visibly tired Anthropic engineer, nervously refreshing a Craigslist tab labeled “used phones, any condition.” “But it turns out that’s really expensive. Like, ‘VCs call you at midnight’ expensive.”

Anthropic’s new distributed iPhone computing cluster, proudly named “Project ShatteredGlass,” reportedly consists entirely of used iPhone 7s and XS models sourced primarily from sketchy parking-lot meetups and awkward door-to-door handoffs in suburban cul-de-sacs. With roughly 50% having functional screens and an impressive 40% retaining battery life beyond a swift sneeze, the makeshift cluster is now proudly running Claude inference jobs at near-2014 levels of performance.

“We found the sweet spot was buying from sellers who listed their phone condition as ‘it works, kinda,’” explained Anthropic’s VP of Compute Strategy while sending payment via Venmo to user BigMike128. “Look, some are skeptical. But honestly, once you jailbreak, root, and rip the screen off, you’ve essentially got yourself an Nvidia A100, just slower, hotter, and prone to spontaneous resets.”

The developer community reaction was swift, ranging from bemusement to mild horror. Reddit threads exploded with reports of increased iPhone resale values, while Stack Overflow questions titled “Why is Anthropic’s API returning ‘Battery Critically Low’?” and “ClaudeGPT keeps complaining about lack of storage, please help” surged in popularity.

Some users have embraced the chaos, proudly contributing their battered devices to Anthropic’s cause. “Honestly, it’s kind of validating,” commented one Redditor. “My cracked iPhone 8, which I previously used solely as a coaster, is now running cutting-edge AI workloads. Well, at least until the battery explodes.”

Anthropic’s head of engineering remains defiant and optimistic. “Sure, they’re slower than actual GPUs. Sure, our entire cluster has about as much RAM as a mid-tier gaming PC. But what else can we do? Cut back on training? Move to Oracle Cloud?” He laughed nervously, shaking his head. “Nah, we’re not that desperate.”