Skip to content
Goodsoil Dev

Guide 7 of 7

Local Models vs. Cloud APIs

1 min readWhy I pay per token instead of buying a GPU and heating my flat.
Local hardware versus cloud APIsA local machine carries upfront hardware cost, electricity, and heat, while a cloud API is pay-as-you-go with no hardware to own and a model that keeps improving.YOUR OWN HARDWARELocal Machine• Large upfront GPU + RAM cost• Electricity, heat & cooling, always• Noise, fan wear & maintenance• Capability fixed at purchasePrivate, offline, one-time buildVSSOMEONE ELSE'S DATA CENTRECloud API• Pay only for the tokens you use• No hardware, power bill, or heat• Nothing to repair or replace• Model improves without new hardwareElastic, always current, zero setup

Running a model on your own machine means prompts stay home, no provider account, experiments continue when an API is down. Real benefits — ones I gave up anyway. The choice isn’t “free vs paid”. It’s buying and running a small AI data centre versus renting someone else’s, which they keep upgrading.

Why I rent instead of own

The buy-in is high

A model worth using for code or agents wants workstation hardware: big-memory GPU, lots of RAM, fast disk, a PSU with headroom. Paid up front, before the first useful token — locked to one configuration while providers spread new accelerators across thousands of users.

The meter keeps running

An always-ready box idles on power; inference pushes it harder. In Singapore the air-con removing that heat is on the bill too. Plus noise, fan wear, updates, the occasional dead part.

The gap moves

Local models handle summarizing, Q&A, and private drafts well. But size isn’t skill — training, tuning, context handling, tool use, and the surrounding software all count, and providers keep shipping stronger options. Test on your own tasks, not on labels.

Compare on five lines

  • Privacy: must data stay on your disk?
  • Offline: must it work with no connection?
  • Quality: how often may it be wrong before it’s useless?
  • Full cost: hardware + power + your maintenance hours, vs the invoice?
  • Upkeep: who updates, fixes, replaces?

Privacy, offline, steady predictable volume: local earns its keep. Changing models with zero maintenance: cloud wins.

My answer

I pay per token. No idle hardware between projects, no heat, no repair queue. Privacy or offline rules can flip the call — for how I build, the invoice is cheaper than the data centre.