Budget Build Monthly — September 2026

As of 2026-10-07 (edition prices: 2026-09-04).

First edition. From here on, once a month we re-price the three build tiers against the weekly price snapshot and call out what changed. All prices below are as of 2026-09-04 with sources; benchmark references are sourced rows only on our board.

Where these stand now (2026-10-07): the market moved hard in September. The 3060 is $495, the used 3090 $1,300 (up 24%) and the 5090 $9,000 — see the 2026-09-28 snapshot. The tier shapes below still hold; the dollar figures are September history.

The tiers

~$800 — 12GB entry

• GPU: RTX 3060 12GB, $477 new (Newegg, 2026-09-04; $495 in the 2026-09-28 snapshot). Still the cheapest 12GB you can buy, four years after launch.
• Runs: 7B–8B Q8_0 interactive, 14B Q4_K_M interactive, small MoE with CPU spill.
• Full list: the $800 entry build.

~$1,500 — 24GB sweet spot

• GPU: used RTX 3090 24GB, ~$1,050 (eBay sold average, May 2026 — the number to beat; the September market has been climbing, so the weekly snapshot is what you budget from).
• Runs: 32B Q4/Q5 interactive; board reference: 81.9 tok/s on Qwen3.8-27B Q4_K_XL (sourced, llama.cpp) — interactive for most people.
• Full list: the $1,500 sweet-spot build.

~$3,000 — 48GB step-up

• GPU: used RTX A6000 48GB, ~$2,500 (eBay); new $4,650.
• The 48GB buys 70B-class dense models at Q4 on one card — no NVLink, no 12VHPWR mod, no dual-power-supply headache.
• Cheaper alternative for the same 48GB: the $2,500 dual-3090 70B build — two used 24GB cards beat the A6000 on price, resale and aggregate bandwidth; the A6000 wins on simplicity.
• Software side: run 70B on used RTX 3090s.

~$5,800 — 32GB flagship

• GPU: RTX 5090 32GB, ~$5,000 (US street average, Sept 2026 — street price, not MSRP; the number moves weekly).
• The speed ceiling on the board: Qwen3.8-27B at 435.1 tok/s (IQ4_XS, 131K context) on the sourced side, plus a sourced 48B-A3B MoE row at 280.8 tok/s (Kimi-Linear-48B-A3B, IQ4_XS, llama.cpp) — the build for people whose 24GB tier ran out of room and who want NVFP4 speed, not 70B capacity. A 4,546.8 tok/s NVFP4/sglang row exists but is still partial, so we do not quote it as a board result.
• Full list: the $5,800 flagship build.

Value picks of the month

These are the rows that don't fit a tier but change what's worth buying this month:

This month's board note

The board now carries one of our own measurements (self-measured): RTX 4090 D 24GB, Qwen3.8-27B at IQ3_S, 67.4 tok/s single-stream with MTP speculative decoding (n-max=2) and 256K context headroom — see the full row. Side-by-side with the sourced 3090 row (81.9 tok/s at Q4_K_XL, llama.cpp, no MTP), the takeaway: quant choice and spec decode move numbers as much as the card generation does. Use the RTX 3090 vs RTX 5090 comparison to check pairs against sourced data before you trust a marketing claim.

Method

• Prices: only sourced numbers from the weekly snapshot; each price carries its source and as-of date. No market-size claims, no "analyst says".
• Benchmarks: only rows on the board with engine and version, quant, context and attribution. Our own measurements (self-measured) are marked self and carry the same fields as community submissions.
• Next edition: late October, against the mid-October snapshot.

Sources

• Prices (2026-09-04 edition, the snapshot this review compares against): weekly price snapshots · 2026-09-28 update: current snapshot
• Board rows quoted above (sourced only): benchmark board · RTX 5090 · RTX 3090 · RTX 3060
• Our own measurements (self-measured) — 67.4 tok/s, IQ3_S: benchmarks
• Tier detail: $800 entry build · $1,500 sweet spot · $9,800-class flagship