Updated October 2026
Build a local AI box: memory, storage, power, and the all-in-one alternative
The parts that make a local AI PC work after you've picked the graphics card, and the small boxes that skip the build entirely.
As an Amazon Associate I earn from qualifying purchases. Product links on this page go to Amazon or the Microsoft Store; both are affiliate links, and the price you pay is the same.
The picks
96GB DDR5 kit (2 x 48GB, 6000MT/s)
96GB on two sticks. Lets you run models bigger than your VRAM by spilling to system memory. Slow, but it runs.
Two-DIMM kits hold rated speed; four-DIMM kits often don't. 96GB is the sweet spot on two sticks; 128GB (2 x 64GB) is the ceiling.
Samsung 990 PRO 4TB NVMe SSD (PCIe 4.0)
Models are 5 to 40GB each and you'll keep a dozen. 4TB means never deleting one to make room for another.
Same speed as the 2TB, twice the room. Cheaper than adding a second drive later.
Corsair RM1000x 1000W 80+ Gold ATX 3.1 power supply
The part everyone underspends on. 1000W, ATX 3.1, native 12V-2x6 connector. No adapters.
Transient spikes on high-end cards trip underspecced supplies. ATX 3.1 units are rated for them.
GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB)
128GB shared memory in a small box. Loads 70B models for less than one 5090 costs; runs them at less than half a Mac Studio's bandwidth.
Up to 96GB can be assigned to the GPU on Windows, more on Linux. That loads 70B models on a machine that costs less than one 5090. 256GB/s of bandwidth means they run, not race.
NVIDIA DGX Spark (GB10, 128GB)
NVIDIA's own small box with 128GB and the full CUDA stack. For fine-tuning, not just running.
If you're fine-tuning, not just running, this is the smallest machine that does it the way the datacenter does.
Microsoft Surface RTX Spark Dev Box (128GB)
The DGX Spark idea on Windows, ready to code out of the box. Costs more than the DGX Spark; pick it only if you want Windows. Ships late November.
The same silicon as the laptop at 140W instead of 45 to 80W, so it should hold its speed under sustained load. If you don't need the screen, this or the DGX Spark is the better buy.
System memory
Your GPU’s VRAM is the fast pool; system RAM is the overflow. Ollama and llama.cpp will split a model across both when it doesn’t fit, and that’s the difference between “runs slowly” and “doesn’t run.” 64GB is the floor; 96GB on two sticks is the sweet spot because two-DIMM kits hold their rated speed and four-DIMM kits often don’t. 128GB on two 64GB sticks is the ceiling for a consumer board. Beyond that you’re into workstation boards and registered memory, and at that point the all-in-one boxes below are the better buy.
Storage
A fast NVMe matters more than you’d think. Model load time is disk-bound: the same 40GB model loads in about 5 seconds from a good NVMe and over a minute from a SATA SSD. PCIe 4.0 is plenty; PCIe 5.0 drives cost more and run hotter for no gain here. Buy capacity. 2TB is the minimum, 4TB is the one you won’t regret, and 8TB is for the person who never deletes a model and wants the second M.2 slot left free.
Power supply
High-end cards draw in spikes, not a steady line, and a supply rated for the average will trip on the peaks. The ATX 3.1 standard exists for exactly this. 850W covers one 5070 Ti or 5080-class card. 1000W covers one 4090 or 5090 with headroom. Two big cards means 1600W with two native 12V-2x6 connectors, not one connector and a splitter. Whatever the size, use the native cable and skip the adapter.
CPU and motherboard
Almost doesn’t matter for inference. Any current mid-range CPU is fine. What matters on the board: enough PCIe slots and physical clearance for a three-slot card, and two M.2 slots so you can add storage later. If you plan on two GPUs, check the slot spacing before you buy.
Or skip the build
Small boxes with 128GB of memory shared between CPU and GPU now do most of what a tower does for running models:
Strix Halo boxes (AMD Ryzen AI Max+ 395). The GMKtec EVO-X2 is the pick; the Beelink GTR9 Pro is the same silicon with better networking, and the Framework Desktop is the same chip in a nicer case sold only from framework.com. All of them let you assign up to 96GB to the GPU on Windows and more on Linux, which loads 70B models on a machine that costs less than a single 5090. Bandwidth is 256GB/s, less than half a Mac Studio M5 Max’s 614GB/s, so those models run at under half the speed; fine for chat, slow for batch work. ROCm support for this chip has been good.
NVIDIA DGX Spark (GB10) is the same idea with NVIDIA’s full software stack on top. If you want to fine-tune models, not only run them, and you want the datacenter tooling on your desk, this is the one. It costs more and it earns it only if you’ll use that stack.
Surface RTX Spark Dev Box is the same NVIDIA platform class in a Windows box, with VS Code, Git, WSL, Python and GitHub Copilot preinstalled. It costs more than the DGX Spark; pick it only if you want Windows rather than NVIDIA’s Linux-based DGX OS. It ships late November 2026, and the Surface guide has the details.
All of them are quiet, fit on a shelf, and need no power supply guide.