Everything recommended across the guides, in one place. Budget tier first in each group.
As an Amazon Associate I earn from qualifying purchases. Product links on this page go to Amazon, Xotic PC or the Microsoft Store; all are affiliate links, and the price you pay is the same.
The cheapest way into Apple's unified memory that's actually enough: 64GB at 307GB/s runs 30B-class models comfortably and 70B models slowly. 64GB is the top option on this chip and a configure-to-order choice at apple.com; Amazon stocks the standard configurations.
Memory bandwidth on the M5 Pro is nearly double the base M6, and bandwidth is what decides tokens per second on a Mac.
128GB at up to 614GB/s loads a 70B model at Q4 with room to spare. Quiet, small, and it doubles as a real workstation. 128GB is a configure-to-order option at apple.com; Amazon stocks the standard configurations.
The smallest quiet box that runs 70B models at a speed you'd use every day. The 128GB Strix Halo boxes fit the same models at less than half the bandwidth.
The only desktop that loads a 400B-class model without a server rack. Overkill for most people, which is the point. 256GB and 512GB are configure-to-order at apple.com.
Up to 512GB of unified memory at 1.2TB/s. If you want to run the biggest open models locally, this is the one box that does it.
The same 128GB and 614GB/s as the Mac Studio M5 Max, in a laptop. The 40-core GPU option is the one with full bandwidth; the 32-core version drops to 460GB/s. 128GB is configure-to-order at apple.com.
The fastest laptop for running a 70B model in memory. The 128GB Strix Halo and RTX Spark laptops fit the same models at well under half the bandwidth. It will be loud and warm doing it, and it will do it on a plane.
16GB of VRAM is the realistic minimum for local LLMs, and in 2026 this is the cheapest new card that has it with full CUDA support. Runs 14B models fast and 30B models at tighter quantization.
The best new 16GB card you can buy under the budget line while 24GB cards sit above it. CUDA means every tool just works.
The cheapest new 16GB NVIDIA card. Slower memory than the 5070 Ti, so fewer tokens per second, but it holds the same models.
For local LLMs, holding the model matters more than speed. This is the lowest price of entry to 16GB with CUDA. Get the 16GB version; the 8GB one is useless for this.
Discontinued, but used ones sit around the budget line and it's the only way to get 24GB there. Ollama and llama.cpp support it through ROCm; some tools don't.
Only worth it if you've checked that your stack runs on ROCm. If it does, this is the cheapest 24GB card in existence right now.
Out of production since late 2024 and now a used-market card, but still the fastest 24GB you can get without paying 5090 money. 30B models at good quantization, fast image generation.
Mature drivers, huge community, and 24GB covers most of what people actually run. Buy renewed with a warranty; the price has held because local AI buyers want exactly this card.
The classic value play: 24GB of VRAM for well under what a used 4090 costs. Slower, hotter, and still the most memory per dollar in a new-or-used NVIDIA card.
For LLM inference, VRAM capacity matters more than compute. Two of these give you 48GB for less than one 5090.
32GB and the fastest memory bus on a consumer card. In 2026 it sells for roughly double its launch price, so it's the pick only when speed matters more than the invoice.
Memory bandwidth near 1.8TB/s. Tokens per second on a 32B model are in a different league from anything below it.
128GB of memory shared between CPU and GPU in a box the size of a hardback. The PC answer to a Mac Studio for loading big models, at a fraction of the price and a fraction of the 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.
A desktop-sized Grace Blackwell box with 128GB unified memory and the full NVIDIA AI software stack.
If you're fine-tuning, not just running, this is the smallest machine that does it the way the datacenter does.
The RTX Spark platform as a small desktop: 128GB unified memory, the full CUDA stack, Windows on Arm. Microsoft.com only, US, shipping 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.
A configurator, not a fixed spec. Start from the base build, pick the RTX 5070 Ti, 64GB of DDR5 and a 2TB NVMe, and it arrives built, tested and under warranty.
The cheapest route to a 16GB-VRAM machine you don't have to assemble. Same parts as a self-build, with someone else responsible when it doesn't boot.
Core Ultra 9 275HX, RTX 5080 Laptop GPU with 16GB of VRAM, 16-inch QHD+ 240Hz panel. Order it with 64GB of RAM; the 16GB default is the one thing to change.
16GB of VRAM in a laptop is the floor for running 14B models at speed and 30B models at all. Below this tier laptops top out at 8 or 12GB and stall on anything useful.
The first Windows laptop with 128GB of unified memory next to an NVIDIA GPU: 20-core Grace CPU, 6,144-core Blackwell GPU, 15-inch 120Hz touch display. Shipping from October 16, 2026.
CUDA in a laptop that holds a 70B model. Bandwidth is unpublished (about 300GB/s derived) and the Windows-on-Arm tools are new, so it's a pick for people who want CUDA on Arm and can wait for the software to settle.
Full-tower Lian Li case with the airflow and clearance a 5090 wants. Configure the 5090, 96GB of DDR5, a 1000W ATX 3.1 supply and 4TB of NVMe.
A 5090 in a small case is a thermal problem you'll fight for years. This chassis makes that problem go away, and the builder picks the power supply so you don't melt a connector.
Workstation-class platform built for one big GPU: more memory channels, ECC options and a board laid out for sustained load rather than gaming.
For people who run models all day, not for an hour. Workstation boards hold more RAM on more channels, which is exactly what CPU offload of a 70B model needs.
Four full-length x16 slots and the power delivery to use them. Two to four GPUs in one box for models that need 48GB of VRAM or more.
Two 24GB or 32GB cards split a 70B model between them at full speed. A consumer board can't space and power two big cards; this one is built for it.
System RAM is where layers go when they don't fit in VRAM. 64GB is the practical floor for a local AI box.
Lets you run models bigger than your GPU by offloading to CPU. Slow, but it runs instead of crashing.
The minimum for an entry-level AI PC. Enough for the OS plus a model that fits in VRAM; not enough to offload anything bigger.
Keeps the entry build under the line in a year when memory is expensive. First upgrade when you outgrow it: replace, don't add, so you stay on two sticks.
Two sticks, 96GB, no stability games with four DIMMs. The right amount if you'll run 70B models with CPU offload.
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.
Two 64GB sticks, so you keep two-DIMM stability at the largest capacity a consumer board takes on two slots.
For running 70B models with most layers on the CPU. Slower than any GPU, but it runs, and it gets there on two sticks, which is the configuration that holds rated speed.
Models are big. A 70B model at Q4 is 40GB, and you'll collect a dozen. 2TB is the sane starting point.
Load time is disk-bound. A fast NVMe turns a 30-second model load into a 5-second one.
For the person who downloads every new release. 4TB holds a serious model library plus datasets.
Same speed as the 2TB, twice the room. Cheaper than adding a second drive later.
For the person who never deletes a model. 8TB holds every release you'll try this year and the datasets to fine-tune on.
One drive, one slot, no juggling. Costs more per TB than two 4TBs but keeps the second M.2 slot free.
Apple's internal storage upgrade costs more than a whole PC drive. Put a 4TB NVMe in this and get ~3GB/s over Thunderbolt for a fraction of the price.
Model load speed is disk-bound, and a Thunderbolt NVMe is fast enough that a 40GB model loads in seconds. Pair it with a 4TB drive.
Enough for one 5070 Ti or 5080-class card with headroom, and the same build quality as the 1000W. Skip it if a 4090 or 5090 is in the plan.
Most single-card builds never pull 600W at the wall. Paying for 1000W you'll never use buys nothing but a bigger fan.
A 4090 or 5090 wants headroom and a native 12V-2x6 connector. 1000W covers one big card with margin.
Transient spikes on high-end cards trip underspecced supplies. ATX 3.1 units are rated for them.
For two big cards in one box: two native 12V-2x6 connectors and the capacity to feed both under load without adapters.
Two 4090s or 5090s can spike past 1200W together. A single-connector 1000W unit with a splitter is how connectors melt.
Six cores on the AM5 socket. For inference the GPU does the work; the CPU just needs to be modern and not the bottleneck on model loading.
The cheapest AM5 chip worth buying. AM5 means a drop-in upgrade to a faster CPU later without changing the board.
What matters on the board: one full-speed x16 slot for the GPU, two M.2 slots so storage can grow, four DDR5 slots. Brand matters less than those three things.
A B650 board does everything an AI PC needs. The extra money for X670 or X870 buys USB4 and PCIe 5.0, neither of which changes tokens per second.
A mesh front and room for a 330mm, three-slot card. Glass fronts look better and run hotter; for a box that runs a GPU at full load for hours, airflow wins.
The case is the one part you keep through every upgrade. Buy for the 5090-sized card you might own in three years, not the 5060 Ti you're buying today.
