AS SOMEONE WHO TESTED THIS $30 SOLAR MESHTASTIC RIG FROM HIS ROOFTOP ACROSS 71 NEIGHBORING NODES, IT KILLED HIS $200 CHATGPT PRO AND $200 CLAUDE CODE MAX
three months ago he bolted a rak 4630 lora board with a 6-watt solar panel to a rooftop mount, whole rig cost him $30 in used marketplace parts
solar charges under complete cloud cover -> pi zero 2w runs quantized qwen 2.5 7b on ollama -> mesh services 71 neighboring nodes -> phone tunnels ai prompts through the mesh -> answers land locally with zero cell tower
the panel keeps the battery full even through a full day of cloud cover, mesh handshakes with 71 neighboring nodes automatically, every prompt over the rf link answers back in under 3 seconds without cellular
$200 claude code max and $200 chatgpt pro cancelled the day the solar started charging in cloud cover, a $30 rooftop rig handles what those subs charged $400 for
bookmark this and read the article below
THIS GUY BOUGHT BOXES OF RETIRED NVIDIA TESLA CARDS TO BUILD A LOCAL AI SERVER FOR LESS THAN ONE MONTH OF SUBSCRIPTIONS
these are old datacenter GPUs like the P100, M40 and P4, cards that once cost thousands but now sell used for a fraction of their original price
the best target is a 24GB Tesla M40 for roughly $130, enough VRAM to run serious local models without paying Claude, ChatGPT and transcription fees every month
add a $25 fan shroud and a $10 power adapter, drop it into an old desktop, then run Ollama for private summaries, transcription and overnight AI jobs
the entire box lands near $165 while a $412 monthly AI stack burns $4,944 a year, turning retired server hardware into a machine that pays for itself in weeks