Local AI hardware, priced this week.

Memory shortages moved GPU prices to 2–5× MSRP. StacksNRacks tracks what a local LLM rig actually costs right now, and what each dollar buys in tokens per second.

Increments updated 2026-10-22

Six reference builds from ~$1k to ~$7k, cheapest first. Speeds are estimates (see method); measured community runs arrive in the next release.

~$1.2k

Entry: one used RTX 3090

8–14B models fast, 32B at Q4

Core
1× RTX 3090 24GB (used)
Memory
24 GB usable · 936 GB/s
Price
$1,000–$1,300 GPU only, used; add a host PC
Biggest fit
Qwen 2.5 32B Q4_K_M
Estimated decode tok/s
8B32B70B
11528—

24.3 tok/s per $1k on 32B

Still the cheapest bandwidth per dollar you can buy. 24GB caps you at ~32B dense.

Price: insiderllm, GB10 vs Strix Halo vs used 3090 · checked 2026-10-22

~$2.3k

Two used 3090s: 48GB

70B at Q4, long context on 32B

Core
2× RTX 3090 24GB (used), layer split
Memory
48 GB usable · 936 GB/s
Price
$2,000–$2,600 GPUs only, used; needs a 1200W+ PSU and a board with two x8 slots
Biggest fit
Llama 3.3 70B Q4_K_M
Estimated decode tok/s
8B32B70B
1152813

12.2 tok/s per $1k on 32B

Doubles capacity, not decode speed. The cheapest way to run 70B entirely on GPU.

Price: insiderllm, GB10 vs Strix Halo vs used 3090 · checked 2026-10-22

~$3.6k

128GB unified, cheapest: Strix Halo

70B Q4 and 100B+ MoE, slowly

Core
Ryzen AI Max+ 395, 128GB LPDDR5X (e.g. Bosgame M5)
Memory
110 GB usable · 256 GB/s
Price
$3,400–$3,800 complete mini PC
Biggest fit
Llama 3.3 70B Q4_K_M
Estimated decode tok/s
8B32B70B
3184

2.2 tok/s per $1k on 32B

Huge capacity per dollar, low power. Bandwidth is the ceiling: great for MoE, slow on big dense models.

Price: insiderllm, GB10 vs Strix Halo vs used 3090 · checked 2026-10-22

~$4.6k

Fastest single card: RTX 5090

Up to 32B at very high speed, image/video gen

Core
1× RTX 5090 32GB (new, street price)
Memory
32 GB usable · 1792 GB/s
Price
$4,300–$5,000 GPU only, US street price; MSRP was $1,999
Biggest fit
Qwen 2.5 32B Q4_K_M
Estimated decode tok/s
8B32B70B
21954—

11.6 tok/s per $1k on 32B

Unbeatable speed, but at 2×+ MSRP the value is poor and 32GB does not fit 70B Q4.

Price: Tech Insider, RTX 5090 price at $4,329 · checked 2026-10-22

~$5.1k

Quiet and fast unified: Mac Studio M5 Max

70B Q4 at usable speed, 100B+ MoE

Core
Mac Studio M5 Max, 128GB unified
Memory
96 GB usable · 614 GB/s
Price
$5,000–$5,200 complete machine, Apple list price
Biggest fit
Llama 3.3 70B Q4_K_M
Estimated decode tok/s
8B32B70B
75199

3.7 tok/s per $1k on 32B

Best capacity × bandwidth in one silent box. macOS lets the GPU use ~75% of RAM by default.

Price: insiderllm, GB10 vs Strix Halo vs used 3090 · checked 2026-10-22

~$7k

CUDA in a box: DGX Spark (GB10)

CUDA dev, fine-tuning, 70B+ at modest speed

Core
NVIDIA DGX Spark, 128GB unified
Memory
119 GB usable · 273 GB/s
Price
$6,950 complete machine, after the Oct 2 price change
Biggest fit
Llama 3.3 70B Q4_K_M
Estimated decode tok/s
8B32B70B
3384

1.2 tok/s per $1k on 32B

Buy it for the CUDA stack, not for decode speed: bandwidth is similar to Strix Halo at ~2× the price.

Price: AIMultiple, DGX Spark alternatives · checked 2026-10-22

Method

Generating a token reads the whole model from memory once, so decode speed is bound by memory bandwidth: tok/s ≈ bandwidth × 0.6 ÷ model size. A model "fits" when its weights plus 3 GB for KV cache (~8k context) and buffers fit in usable GPU or unified memory. Splitting layers over two GPUs adds capacity, not decode speed.

Reference models (Q4_K_M GGUF): Llama 3.1 8B (4.9 GB), Qwen 2.5 32B (19.9 GB), Llama 3.3 70B (42.5 GB). Prompt processing, batching and MoE models behave differently; that's what measured runs are for.

Prices are dated snapshots from the linked sources, not live quotes. Check a listing before you buy.

Show us your rig

Running models locally? Share the parts, the price you paid and your llama-bench numbers. The best rigs get featured in the newsletter.

Submit a rig