Own Your AI is a nine-part Top Tool Stack series on running your own local AI instead of renting it from the giants. This is Part 8 of 9. See the whole series →
By now you can run a model, pick a tool, choose a model and do it safely. This one is for the person who has caught the bug and wants to build something more serious, from spending nothing to building a genuine home AI machine. We will be honest about the costs and the diminishing returns at every tier, because this is a hobby where it is very easy to spend a lot of money chasing a little more quality.
Tier 0: use what you already own ($0)
Genuinely, start here. A modern laptop or a recent Apple Silicon Mac runs small and mid-size models today, as covered in Part 6. Spend a few weeks running local AI on your existing machine before buying anything. Most people discover their current hardware covers more of their needs than they expected, and the ones who do outgrow it end up knowing exactly what they actually need. That is money saved and regret avoided.
Tier 1: one good graphics card (the sweet spot)
The single most cost-effective upgrade for local AI is memory on a graphics card (VRAM), because VRAM is what lets you run bigger, better models at speed. The enduring community favourite is a used 24GB card (the previous-generation NVIDIA 3090 has been the darling for exactly this reason: lots of VRAM per pound on the second-hand market). A single 24GB card moves you from “8B models” to “comfortable 70B-class models”, which is the biggest quality jump you will feel. For most enthusiasts, this is where the journey should stop, because the returns after this get expensive fast.
Tier 2: the Apple Silicon shortcut
Worth calling out separately, because it breaks the usual rules. Apple’s M-series chips share memory between the processor and graphics, so a Mac with 64GB or 128GB of unified memory can load very large models without a rack of graphics cards, quietly and with modest power draw. It is not the cheapest route per gigabyte, but for a silent, tidy, low-power machine that punches far above its weight, it is arguably the most sensible “serious” option for a non-tinkerer.
Tier 3: the multi-GPU rig (here be dragons)
This is PewDiePie territory: several graphics cards in one machine, hundreds of gigabytes of VRAM, running the very largest open models and experiments like his “Council” (several models voting on the best answer). It is genuinely thrilling and genuinely a money pit. The honest warnings:
- Cost. Estimates of Felix’s own rig ran from around $20,000 to $40,000. You do not need anything remotely like that, and for personal use, you almost certainly should not.
- Power and heat. Multiple high-end cards draw serious electricity and dump serious heat into your home. This is the tier where your energy bill and your cooling stop being an afterthought, which is precisely what next week’s article tackles.
- Diminishing returns. Going from nothing to a 24GB card transforms your experience. Going from one card to eight mostly lets you run models that are still not quite frontier quality, for a lot of money and noise.
The grounded recommendation
For the overwhelming majority of people, the right build is Tier 0 then, if you outgrow it, Tier 1: use what you have, and if you genuinely need more, add a single high-VRAM card or buy a high-memory Mac. That covers everything most people will ever ask of a local AI. Tier 3 is a wonderful hobby if the building itself is the fun, as it clearly is for Felix, but it is a hobby, not a requirement, and nobody needs a five-figure rack to own their AI.
Whatever tier you land on, the same question applies: how do you run it without a punishing electricity bill or a guilty conscience? That is the finale. Next week, the honest guide to making your DIY AI as clean, cheap and efficient as it can realistically be.
The AI tool stack actually worth paying for
One email a week. The tools, models and moves that matter, minus the hype and the horseshit filter set to maximum. Free.