Builds / $30k tier

The $30k question assembly diagram
THE RIG
The $30k question rig photo 1
Screenshot: Apple Store, Mac Studio 512GB configuration ($9,499)

The $30k question

The budget from the 'Mac Studios or Spark?' thread: one 512GB Mac Studio, the big models, no cloud bill.

TECHNICAL DIFFICULTY: 2 OF 4 ยท SEAT AND GO
How hard is this to assemble? two boxes, both prebuilt: seat nothing, configure software only

This answers the question that keeps showing up on X: I have 20 to 30 thousand dollars, what do I buy? The answer for running the biggest open models locally is unified memory, and the 512GB M4 Ultra is how you get 512GB in one quiet box. Every model that matters in 2026 runs here, at full context, in your office. It will not beat NVIDIA for video generation; that is what the 5090 in this build is for.

WHO THIS BUILD IS FOR

You read model releases the morning they ship and want to run them that afternoon, alone, on your own hardware. Researchers, founders, and engineers who want the frontier without sending data to a vendor.

ESTIMATED TOTAL
$16,794
prices refreshed weekly
last check: Oct 3, 2026

The parts

Mac Studio M4 Ultra 512GB
Mac Studio M4 Ultra 512GB x1 512GB 1092GB/s apple
512GB unified memory: the whole reason this build exists. GLM 5.3, DeepSeek V4 Pro, Kimi K3, Nex-N2.5 at real context.
vendor price (fixed MSRP), checked Sep 20
$9,999
NVIDIA RTX 5090 32GB
NVIDIA RTX 5090 32GB x1 32GB 1792GB/s nvidia
Output head: video and image generation, where NVIDIA still leads.
street price via Newegg, checked Oct 3
$6,795

What it runs

GLM 5.3
743B total MoE: 4-bit is about 371GB in 512GB. The full frontier model at real context, in one box.
DeepSeek V4 Pro
1.6T total MoE: 2-bit is about 400GB in 512GB. It fits, but that quant costs quality; GLM 5.3 is the honest daily model here.
DeepSeek V4.1 Flash
4-bit is about 276GB in 512GB: the fast-everything model runs comfortably here.
WHAT IT WILL NOT RUN

Kimi K3: 2.8T parameters is 700GB of weights at 2-bit, and no quant we would recommend fits in 512GB. NVIDIA-first video stacks are the other gap: the 5090 in this build exists as the video head.

POWER, NOISE, AND THE ROOM IT LIVES IN

About 300W under load: quieter than a gaming laptop. One cable, one outlet, and the setup is entirely software.

NEXT: THE $50k TIER

Four fans of consequence. A silent 512GB frontier box, or four RTX 5090s that generate video while they chat. At this tier the question is not what can it run: everything runs. The question is what you will build with a private datacenter that draws less power than a space heater.

See the $50k tier →