I bought my AMD Strix Halo system (128GB RAM) in 2025 for about $1800 at the time. So was very lucky timing.

However I would say realistically its only been in last 6 months that it has been "paying for itself" due to local model improvements. Qwen3-Coder-Next was the real local model I found that was usable... and now Qwen-3.6 and Gemma4 have both made it for "routine coding" task can be done locally.

For doing really entailed work I still use SOTA models, but my job is mainly sys-admin so lots of python, ansible, bash, sql scripts, etc....and for those things local models are very very helpful.

I've even moved my accounting system to beancount (beancount and ledger are both text-based accounting systemsm so perfect for an a LLM)....I simply fed the CSV files from my bank into local model and asked it to create a parser to parse those from csv to beancount format...

The real advance of local models really makes me question is these multi-billion dollar data centers are going to be used to the degree they imagine....I think its very possible that local models get "good enough" for 90% of usecases