Workstation local AI build
This is the top tier in our build guides, and the first one where the machine is a tool rather than a computer that also does AI. It is for running the largest open-weight models locally, serving models to other people or applications, and training rather than just inference. It means multiple GPUs or a large unified memory system, and it brings real constraints around power, heat and noise. Most people do not need this tier; the hardware section and the tier below cover what a smaller machine does.
Target
Roughly 48 GB of VRAM and up, usually aggregated across cards, or an Apple Silicon system with 128 GB or more of unified memory.
Approximately, that opens up:
- Large open-weight models, the hundreds-of-billions parameter class, at aggressive quantization, and mid-sized models at full precision.
- Very long context windows, since the KV cache at long context can exceed the size of the weights themselves.
- Multiple models resident at once, which is what serving an application or an agent stack actually requires.
- Full fine-tuning rather than LoRA adapters only, plus large batch image and video diffusion jobs.
Aggregated VRAM is not the same as a single large pool. Splitting a model across cards costs you the interconnect, and not every engine splits well, so check support before you buy the second card.
GPU or unified memory
There is no consumer single-card answer at this tier, so you build it from the cards on our GPU pages:
- Multiple RTX 5090 cards, 32 GB of GDDR7 each, the highest capacity and memory bandwidth per slot available in consumer parts.
- Multiple RTX 4090 or RTX 3090 cards, 24 GB each, a common way to reach 48 or 72 GB without new-generation hardware.
Apple Silicon is the better shape here when capacity and quiet matter more than training throughput. A large unified memory configuration built on M3 Max class silicon, or the higher tiers above it, holds models that would need three or four consumer cards, in one box that draws a fraction of the power. It is the wrong choice if you are training seriously, where raw compute and mature tooling still favour a multi-GPU machine.
The rest of the box
- CPU and platform. This is the one tier where the platform genuinely constrains you. You need enough PCIe lanes to run several cards at reasonable width, which usually means a workstation-class CPU rather than a desktop one.
- System RAM. Match or exceed total VRAM, and more if you fine-tune. Datasets, staged weights and CPU offload all live here, and an undersized RAM pool turns every model load into a disk-bound wait.
- Storage. Fast NVMe measured in many terabytes, ideally split: a fast working drive for active models and checkpoints, bulk storage for the archive of weights you are not currently using.
- Power and cooling. Plan this first, not last. Several cards under sustained load pull serious continuous power, so check what your circuit supports, leave real supply headroom, and expect to make choices about card spacing, cooler type and where the machine physically lives.
Upgrade path
There is no tier above this one here; the next step is hardware most people rent rather than own. If it is more machine than you need, the enthusiast tier gives you most of the day-to-day capability in one quiet box.