running models locally =
running your own node
normies want chatgpt the same way they wanted coinbase. easy, custodial, someone else’s rules.
meanwhile the freaks are pulling quantized 70B models through ollama, running open-webui on refurbed dell workstations with a single 4060 Ti, wiring up MCP servers so their local models can use real tools, and building RAG pipelines over their own documents. same energy as compiling bitcoind on a thinkpad in 2012.
open source models are eating closed source the same way bitcoin ate every altcoin that tried to compete on “features.” llama, deepseek, qwen: permissionless, forkable, nobody’s product.
running inference on your own hardware is “not your keys not your coins” for AI. your prompts, your weights, your machine. no logs. no content policy. no ToS update at 2am that lobotomizes your model because some compliance team got nervous.
the hosted API providers are the exchanges. convenient until they’re not. rate limited, censored, deprecated on a thursday with zero warning. you don’t verify, you trust.
this past week, both Coinbase and Lightning Labs shipped agent payment tooling built on HTTP 402, the internet’s forgotten “payment required” status code. the approaches tell you everything.
Coinbase’s x402 routes agent payments through USDC on Base. Stripe integrated it the same day. requires a facilitator service, stablecoin custody, onchain settlement through their infrastructure. it’s the banking stack rebuilt for machines.
Lightning Labs dropped L402: fully open source, no facilitator, no identity, no signup. an agent hits a paywall, lnget pays a lightning invoice, gets a macaroon as cryptographic proof, and moves on. agent-to-agent commerce on bitcoin rails with scoped permissions and remote key isolation. lnget works like wget except it pays invoices. unix philosophy meets sound money.
one side is recreating visa for bots. the other is extending the protocol. same fork in the road we’ve been watching for over a decade.
local AI setups today look like running bitcoin core in 2012. ugly, half broken, mass tinkering required. the people doing it anyway will be unreasonably far ahead of everyone else within a few years.
normies will get their sanitized corporate AI the same way they got their bitcoin ETFs. late, watered down, and fully captured.
the builders are in the terminal.