I learned about Bitcoin in 2011. I started mining Bitcoin in 2012 on a laptop. By 2023 I had a 165,000 sq/ft facility printing money.
I learned about AI in 2020. I started building with it in 2021. By 2026 I had four autonomous agents running million-dollar businesses on MacMinis, connected to a DGX Spark (brain/RAG), with my main orchestrator on a BEAST Apple Pro, while running science in pulcros and various other racked hardware I can find…
The simple reality is I (am) a scientist by education and personality. I just cannot stop learning, tinkering, and breaking things. As for AI, the end-state is fundamentally clear:
You must go local AI. The economics for doing so will only get cheaper, so you can wait a bit… but it will be the inevitable place (you want to be) at.
WHY GO LOCAL AI NOW
No one can take it away - Hosted access can be revoked at any time; local weights are yours forever… and you can TRAIN TRAIN TRAIN for YOU.
You control the quality - Providers silently quantize, update, or degrade models (peak-hour throttling, KV-cache quantization) happens all the time. Self-hosted output quality stays constant, and your evals/workflows don’t break when a model is swapped under you… I cannot believe that we allow our paid service to downgrade on the fly. It’s worse than Apple iPhone planned obsolescence…
No silent model swaps - you decide when to upgrade; API models change and invalidate workflows that were tuned on them… this has been fundamentally frustrating for me as I know what (I) want the agents to do and have them move to other models mid flight is quite regarded.
“Civilizational infrastructure” - AI will soon matter more than the internet or electricity… and we’re already seeing the FREE AI FOR ALL sentiment being heralded as a utility that shouldn’t be gated by a few companies… sure. So. Let’s go local.
Nobody sees your conversations or IP - critical for organizations as people share more with AI than any other technology. My consulting has revealed to me how important good AI safety really is. LOCAL MODELS IS THE WAY!
Data is being harvested for ads/profiling - cites ChatGPT ad targeting and it looks stupid and lowers trust. Providers know more about you than any website and frankly I just can’t use any 'free’ GPT services anymore as the ads blatantly tell me they are reading everything. It’s just too much…
Data can be used against you - worst-case scenarios like insurers buying conversation data, or systematic leverage/blackmail is actually a thing. Let’s be real. Data breaches are a dime a dozen…
Again… you can wait as prices go down… but the learning/training you’ll need to optimize (should begin) today.
ECONOMIC VALUE & INCENTIVES
Money is simply on the table… the verdict is still out… but I have seen ~70% first-year savings deploying owned infrastructure vs. API/subscription spend and I have not had any of my clients from the previous 3 years go fully local… so I cannot provide data on my side from that… but for ME… I can’t be a useful data-point. I waste money for the love of science.
We have always seen that hardware appreciates in usefulness… or re-usefulness. Smaller/more efficient open models keep making the same box more capable (DGX Spark went from “useless” to running frontier-class models in a year)… and I’ve been pretty much heralding it (to my clients) as a great sandbox to learn.
Finally, open models have closed the gap in huge ways. Kimi K3 / DeepSeek V4 Flash class models are near-frontier and runnable locally, so the capability excuse is gone and the GLM5.3 is the nuts. It just works.
PETER’S PRACTICAL ON-RAMP
~$2500 easy mode: Mac Studio
~$5K entry: DGX Spark, Strix Halo, or 2x used RTX 3090
~$10K: RTX 5090 full build
Ask your ai agent about balancing bandwidth, capacity, and software support together. Tell your ai agent to consider ODS project (Apache 2.0) packages the full local stack (search, RAG, agents) and the internets for the hardware deep-dives for your specific use-case. There are (plenty) of resources for OpenClaw/Hermes (you can look back and find my walkthroughs).
This is my full brain dump on going local ai. I foresee the only variables in the near future to be lower costs, easier installs, automated setups, and all-in-one solutions to emerge for the normie and retail.
Build-Anything-In-A-Box is coming.
You’ll want to make sure it’s imminent usefulness will not come at the cost of giving (everything) away.
Best,
ps
Join my AI Workshop September 9 at 5pm EST to begin salvation.