A $15,000 Mac for local AI buys six years of a $200 subscription

NetworkChuck called Apple's M5 Ultra "a local AI powerhouse." 512GB RAM, 4.3x faster compute, 1.2TB/s bandwidth.

Run the best model that fits on that hardware flat out for a year and you'd produce a few hundred dollars worth of tokens, at a worse quant than the hosted version. That same model runs at $0.08 per million input tokens on any inference provider today.

$15,000 for that Mac buys 6 years of a $200/month Claude Max plan instead. In 6 years the Mac is scrap metal. The subscription gets a smarter model every quarter for the same $200.

Local wins only if data legally cannot leave the building, or you're already burning 5-6 subscriptions worth of tokens a day. That's the whole list.

Everything else is $15,000 to hedge against subscriptions getting pricier someday.