How to Build an AI Workstation on a Tight Budget

You can put together a machine that runs real AI work for far less than the flagship-build crowd implies — as long as you spend the money where it changes what you can do and refuse to spend it where it does not. The trick is knowing which parts are gates and which are merely nice.

On an AI machine the gate is almost always one number: how much memory sits on your GPU. Everything in this guide comes back to protecting that number. It holds whether you buy new, used, or a mix, and whether you are running models locally or doing light fine-tuning — the principles do not care where the parts come from.

Put the money into GPU memory first#

For AI work, the part that decides what you can run is the accelerator — in practice a GPU — and, above all, how much dedicated memory (VRAM) it has. A model has to fit in that memory to run at full speed. If it does not fit, one of two things happens: it will not load at all, or your tools spill the overflow into system RAM and run part of it on the CPU, which works but drops throughput sharply. Either way, VRAM is the wall. So the first and largest slice of a tight budget belongs here.

The useful instinct on a budget is to favor more memory over more raw speed. A card with a larger memory pool and modest performance will run things a faster card with a small pool simply cannot — and a model that runs slowly still beats a model that does not run at all. This is also where the used market pays off most. A previous-generation card with generous VRAM often beats a current entry-level card with less, for similar money: a used 24GB card such as an RTX 3090 will hold models that a newer 16GB card like the RTX 5080 has to turn away, even though the 5080 is the faster chip. As of mid-2026 the top consumer GPU tops out at 32GB (the RTX 5090), so 24GB used cards remain the budget sweet spot for capacity per dollar. Buy memory capacity first, speed second.

How much VRAM do you actually need?#

You can size this yourself instead of guessing. A model's weights take up roughly the number of parameters times the bytes per parameter. Unquantized, weights are usually 16-bit — two bytes each — which is why people quantize. Quantization compresses the weights to about 4 bits each, roughly half a byte, for a small and usually acceptable quality cost. That is the whole difference between a 70-billion-parameter model needing about 140GB and needing about 40GB.

A rough working rule for 4-bit models is about half a gigabyte of VRAM per billion parameters, plus a few gigabytes of headroom for context and overhead. In current mid-2026 terms:

• A small model like Gemma 3 4B runs in roughly 4GB — fine on an 8GB card or even a capable integrated GPU.
• An 8B–14B model such as Qwen3 14B wants roughly 8–10GB.
• A 32B dense model, or a 30B mixture-of-experts model like Qwen3 30B-A3B (which keeps all its experts in memory but only activates a few billion parameters per token, so it runs fast), fits comfortably on a 24GB card. That is exactly why 24GB is the popular target.
• A 70B model such as Llama 3.3 70B needs around 40GB at 4-bit — a two-card build or a unified-memory machine.

One thing people forget: long context windows cost extra VRAM on top of the weights, for the key/value cache that holds the running conversation. Big weights plus a big context can blow a budget you sized for weights alone, so leave headroom rather than fitting the model exactly.

The other path: unified memory#

There is a second way to buy capacity, and it changed the budget math over the past year. Apple Silicon Macs and AMD's Ryzen AI Max+ 395 ("Strix Halo") do not have separate VRAM at all — the GPU shares one large pool of system memory. On the AMD part that pool runs up to 128GB, most of it usable by the GPU, enough to hold models that will not fit on any single consumer discrete card. These ship as mini PCs and laptops, roughly in the $1,500–$2,000 range as of mid-2026, which undercuts a multi-GPU rig for the same capacity.

The catch is bandwidth. Unified system memory is much slower than a discrete GPU's dedicated memory, so those large models load and run — just at fewer tokens per second than a "real" GPU manages on a model that fits its VRAM. It is a capacity play, not a speed play, and it suits running models better than heavy training. If your goal is to run big models cheaply and you can live with moderate speed, it is the most interesting budget option on the board right now.

Give the system enough RAM to stage the work#

System memory is not where the model runs, but it is where models load from, where data gets staged, and where everything else lives while the GPU works. Too little of it turns into constant disk swapping that drags the whole machine down — and it is also the safety net when a model does not quite fit in VRAM and you offload part of it to the CPU.

A sound rule is to have at least as much system RAM as VRAM, and ideally more — 1.5 to 2 times is comfortable, especially if you plan to run models that spill out of the GPU. Treat it as cheap insurance against bottlenecks. The good news: system RAM is usually easy and inexpensive to add later, so it is a fine place to start modest and upgrade when you feel the pinch.

Spend just enough on the parts that do not gate capability#

The processor matters for loading models and for the steps the GPU does not handle, but a mid-range chip is plenty for most AI work. It earns its keep mainly when you are offloading model layers to the CPU; short of that, do not let CPU marketing pull money away from VRAM.

Storage should be a fast NVMe solid-state drive, because models are large and load times add up — a single 70B model at 4-bit is around 40GB on disk, and they accumulate faster than you expect. You do not need the biggest or fastest tier to start; a 1TB-class NVMe drive is a sensible floor you can expand.

The power supply and cooling are the two places not to cut corners. A sustained AI load runs the GPU flat out for long stretches, which means real heat and, on modern cards, large split-second power spikes well above the rated draw — size a PSU to the average and it can trip. Buy a reliable supply with genuine headroom and cooling built for sustained load, then put the rest into memory.

What to look for#

• A GPU chosen for the largest VRAM your budget allows, prioritizing capacity over raw speed. A used 24GB card is the current value benchmark; a unified-memory machine is the alternative when you need more than a single card holds.
• Enough system RAM to avoid swapping and to cover CPU offload — at least matching your VRAM, with the option to add more cheaply later.
• A reliable power supply with headroom for transient spikes, and cooling sized for sustained load.
• A fast NVMe solid-state drive large enough for the several models you will actually keep around.
• A platform — board, case, PSU — that leaves room to add memory, storage, or a second GPU down the line.

What to skip#

• A flagship processor. A mid-range chip rarely limits AI work, and the savings buy more VRAM.
• The largest, fastest storage on day one, when a sensible NVMe drive does the job and can be expanded later.
• Cosmetic extras like lighting and showpiece cases that add cost without adding capability.
• A faster GPU with a small memory pool when a slower one with more memory runs more of what you need. Speed is worthless on a model that will not load.

The honest summary: a budget AI workstation is an exercise in priorities. GPU memory is the gate, so it gets the money — and quantization plus a used 24GB card or a unified-memory box is how you buy a lot of it cheaply. System RAM, the power supply, and cooling keep the machine stable and get bought sensibly. The processor and storage are good-enough purchases. Spend in that order and a modest budget produces a machine that does real work, with a clear path to grow as your needs and your wallet allow.