How much PSU do I need for 24/7 inference?

Size for sustained load plus transient headroom, not for a gaming average. Inference holds the GPU near its power limit for minutes or hours, so the PSU runs hot continuously rather than in bursts. For a single 150-250W card (Arc B580, RTX 3060, RTX 5060 Ti) a quality 650-750W 80+ Gold unit is enough. For a 300-350W card (RTX 4070 Ti Super, used RTX 3090) go 850W. For a 575W RTX 5090 go 1000W or higher, ATX 3.1 with a native 12V-2x6 cable. ATX 3.0 and 3.1 units are specified to survive 180 percent of rated wattage for 1 millisecond, which is what stops modern GPU spikes from tripping over-current protection. Verified August 2026 listings put 650-850W Gold units around $80-120, so this is one of the few parts the shortage did not ruin.

Is 64GB of system RAM enough for local AI in 2026?

Yes, for the models most people actually run. 64GB holds gpt-oss 120B at Q4 (about 62GB), Llama 4 Scout at Q4 (about 58GB), and Laguna XS 2.1 at Q8 (about 36GB) with room for context. It is not enough for 200B-class dense models or for holding two large models resident at once. In August 2026 there is a second argument for 64GB: DDR5 costs roughly four times its 2025 price, so buying 128GB now costs several hundred dollars more than it did a year ago for the same parts.

B650 or X670E for a single-GPU AI build?

B650 or B650E is the right call for a single-GPU, one-or-two-NVMe build. X670E adds a second chipset die, which roughly doubles chipset-attached PCIe lanes and gives more M.2 and USB connectivity. That matters if you plan a second GPU, several fast drives, or capture and passthrough hardware. It does not change inference speed on a single card. Verified August 2026 listings put decent B650 boards at about $140-215, and that segment was largely unaffected by the memory shortage.

What does a 64GB local AI rig cost in August 2026?

Roughly $1,700-2,100 for a budget build on an Intel Arc B580 12GB, about $2,250-3,050 for the mainstream build on an RTX 5060 Ti 16GB, and about $2,700-3,600 if you use a used RTX 3090 24GB. Memory and storage drive most of that: a 64GB DDR5 kit alone is about $680-880 and a 2TB NVMe roughly $180-330. All figures are August 2026 street ranges and move week to week; the CPU, cooler and case allowance in these totals is an estimate rather than a verified quote.

← All guides

Hardware August 11, 2026

64GB Local AI Rig: The Complete 2026 Parts List (August 2026 Prices)

A 64GB system-RAM, single-GPU tower is still the most flexible way to run local models — you get CUDA, you can swap the GPU later, and you are not locked to one vendor's memory ladder. But August 2026 is a bad moment to buy RAM. DDR5 has roughly quadrupled since 2025, and a 64GB (2x32) kit now runs about $680-880. Here is the full parts list, every choice explained, with prices as of August 2026 and an honest read on what you should buy now versus wait on.

Building a 64GB OpenClaw rig?

See our AI training options. We'll spec the build against the models you actually plan to run, then set up OpenClaw on it.

Read this first — August 2026 is a bad time to buy RAM

DDR5 32GB kits went from roughly $80-120 in 2025 to $375-589 today. A 64GB kit is $680-880.
NVMe is up about 115 percent. A 1TB Gen4 drive is $90-165, not $50.
NVIDIA has raised GeForce prices three times in 2026. Published MSRPs are not prices.
Boards, PSUs and coolers are basically unaffected. The shortage is memory and flash.
No relief is expected before 2027. If you already own DDR5, keep it.

Bottom Line (August 2026)

CPU: AMD Ryzen 9 9950X (16C/32T, Zen 5) if you offload experts to CPU. Ryzen 7 9700X if you do not.
Board: any solid B650 / B650E, 4 DIMM slots, 2 M.2. About $140-215. X670E only if a second GPU is coming.
RAM: 64GB as 2x32GB, never 4x16GB. About $680-880. This is the single most inflated line on the list.
GPU: Arc B580 12GB (~$305) to start, RTX 5060 Ti 16GB ($589-805) for the mainstream pick, used RTX 3090 24GB ($1,000-1,300) if you want 24GB VRAM.
PSU: 750W 80+ Gold for small cards, 850W ATX 3.1 for the 300-350W class (3090, 4070 Ti Super). A 575W RTX 5090 needs 1000W+ — 850W is not enough for that path. Around $80-120 at 650-850W. Size for sustained load, not gaming averages.
Storage: 2TB NVMe minimum — models are 15-70GB each. Roughly $180-330.
Total: about $1,700-2,100 budget, $2,250-3,050 mainstream, $2,700-3,600 with a 3090.

The honest version: the GPU is no longer the expensive decision. Memory is. If you can wait two quarters on the RAM, do that and buy the rest now.

The Full Parts List

Prices are US street ranges as of August 2026 and move week to week. Ranges, not point prices, are the honest unit here.

Part	Recommendation	Why	Price (Aug 2026)
CPU	AMD Ryzen 9 9950X (16C/32T, Zen 5, AM5)	Zen 5 adds about 16 percent IPC over Zen 4, and the dual-channel DDR5-6000 controller delivers roughly 90 GB/s — the number that decides CPU-offload token speed. Ryzen 7 9700X is the cheaper pick if the model fits in VRAM.	Not covered by our Aug 2026 price check — verify before buying
Motherboard	ASUS TUF Gaming B650-PLUS WIFI or MSI MAG B650 Tomahawk WIFI (ATX, 4x DDR5, 2x M.2)	B650 runs a single GPU at full PCIe 4.0 x16 and takes two NVMe drives. X670E’s second chipset die only pays off with a second GPU or many fast drives.	~$140-215
RAM	64GB DDR5 as 2x32GB — 64GB kit	Two DIMMs, not four. Filling all four AM5 slots loads the memory controller harder and commonly drops you toward JEDEC 3600 MT/s with worse latency. 2x32 leaves both spare slots for a later 128GB move.	~$680-880
GPU	See the GPU section below	VRAM decides which models run entirely on the card. Everything past that spills into system RAM and slows down.	$305 to $5,000
PSU	Corsair RM850x (2024) or Seasonic Focus GX-850 ATX 3.1, 80+ Gold	Sustained inference is not gaming. The unit sits near full load for hours, so buy headroom and buy Gold or better. ATX 3.1 with a native 12V-2x6 cable handles modern transient spikes. 850W covers every GPU in the table below up to the 3090 / 4070 Ti Super class. It does not cover a 5090 — see the note under the GPU table.	~$80-120 (650-850W Gold)
Storage	2TB NVMe — SanDisk 2TB	A single Q4 70B GGUF is 40GB+; gpt-oss 120B at Q4 is about 62GB. Four serious models fill 1TB.	~$180-330 (2TB)
Second drive	Optional 4TB — SanDisk 4TB	Split OS and model library so a rebuild does not cost you 800GB of re-downloads.	Scale from the 1TB range
CPU cooler	Thermalright Peerless Assassin 120 SE (value) or Noctua NH-D15 G2 (quiet, 24/7)	Both are dual-tower air. The NH-D15 G2 uses offset AM5 mounting and held a 9800X3D at 72C under sustained Cinebench load at 24.8 dB. Check case clearance: it needs 170mm.	Not covered by our Aug 2026 price check
Case	Fractal Design North XL or Lian Li Lancool III (ATX, high airflow, 170mm cooler clearance)	Inference is a thermal marathon. You want front intake straight onto the GPU, dust filters, and room for a 3-slot card.	Not covered by our Aug 2026 price check
Monitor	Dell 27” 4K	Optional. 4K makes side-by-side terminal and editor work bearable.	—
Dock	Anker USB-C hub	Optional. Useful if the tower doubles as a laptop docking target.	—

On the linked RAM kits: the kits above are DDR5-4800 CL40. They work fine on AM5 at JEDEC speed. If you want the EXPO 6000 CL30 profile that the 90 GB/s figure assumes, buy a kit that explicitly lists it. Smaller options if you are building down: 32GB kit · 16GB kit. We would not build a 2026 AI rig on 32GB.

GPU: Pick By VRAM, Not By Frame Rate

VRAM is the only spec that changes which models you can run. Everything else changes how fast they run.

GPU	VRAM	Price (Aug 2026)	Who it is for
Intel Arc B580	12GB	~$300-310	Cheapest new 12GB card. Fine for 20B-class models at Q4. Software stack is behind CUDA.
RTX 3060 12GB	12GB	$329-460 new	CUDA on a budget. NVIDIA revived the SKU during the shortage, so it is no longer a cheap card.
RTX 5060 Ti 16GB	16GB	$589-805	The mainstream pick. 16GB runs Qwen 3.6 27B at Q4 or gpt-oss 20B at Q8 fully resident.
RTX 4060 Ti 16GB	16GB	Check current price	Prior-gen 16GB alternative. We have no verified August 2026 street price, so compare it live against the 5060 Ti.
RTX 4070 Ti Super 16GB	16GB	~$1,100-1,200	Faster 16GB, but now above its own launch MSRP. Hard to justify against a used 3090.
RTX 3090 24GB (used)	24GB	$1,000-1,300	Best VRAM per dollar if you need 24GB. Ignore the “$650 used 3090” figure still repeated everywhere — that price is two years gone.
RTX 5090 32GB	32GB	$4,300-5,000+	The $1,999 MSRP is effectively fiction. Only worth it if 32GB VRAM is the whole point of the build.

Match the PSU to the card you actually pick. The 850W unit in the parts list is sized for the middle of this table, not the top of it. The RTX 3090 (350W) and RTX 4070 Ti Super (285W) are comfortable on 850W. The RTX 5090 is 575W and NVIDIA specifies a 1000W system PSU — pair that with a 230W-PPT Ryzen 9 9950X and you are at roughly 805W of sustained draw before drives and fans, on a unit rated for 850W. That is not headroom, that is the ragged edge, and inference holds it there for hours. If you are building the 5090 path, buy a 1000W or larger ATX 3.1 unit with a native 12V-2x6 cable — we do not have a link for one yet, so shop it yourself rather than reusing the 850W pick above. The same applies to an RTX 4090 build: 850W is NVIDIA’s stated floor for that card, not a comfortable number next to a 16-core CPU.

Our pick at this tier: the RTX 5060 Ti 16GB. 16GB is the smallest VRAM that keeps a useful coding model entirely on the card, and the 64GB of system RAM behind it covers everything larger via offload.

What This Rig Actually Runs

With 16GB VRAM and 64GB system RAM, three tiers exist:

Fully on the GPU (fast, 30+ tok/s):

gpt-oss 20B at Q8 — about 13GB. The reliable tool-calling model for OpenClaw agent loops.
Qwen 3.6 27B at Q4 — about 17GB with a modest context, or Q4 with light offload.
Gemma 4 26B MoE at Q4 — multimodal, fits with room for context.

Split GPU + system RAM (usable, single-digit to low-teens tok/s):

Laguna XS 2.1 (33B total / 3B active MoE) at Q8 — about 36GB. Only 3B parameters activate per token, so MoE offload hurts far less than a dense model of the same size.
Llama 4 Scout (109B/17B MoE) at Q4 — about 58GB, 10M context window. This is the model that justifies 64GB.
gpt-oss 120B at Q4 — about 62GB. Tight, but it runs, and it is the production pick when quality beats speed.

Does not fit: dense 70B at anything above Q4, and anything in the 200B+ class. Those need 128GB or a multi-GPU box.

The pattern worth internalizing: MoE models with small active parameter counts survive CPU offload; dense models do not. A 3B-active MoE spilled into system RAM stays usable. A dense 70B in the same situation crawls.

PSU Sizing For 24/7 Inference

This is where build guides written for gamers mislead you. Gaming load is bursty. Inference load is a plateau — the GPU sits near its power limit for the length of a generation, and an agent loop can hold it there for hours.

Three rules:

Size for sustained draw, then add transient headroom. Take the GPU’s board power plus about 150W for CPU, drives and fans, then leave 30-40 percent on top.
Buy ATX 3.0 or 3.1 if the card has a 12V-2x6 connector. Those specs require the unit to ride out 180 percent of rated wattage for 1 millisecond, which is what modern GPU spikes actually look like.
Gold minimum, Titanium if the box never sleeps. At about $0.18/kWh US average, a few efficiency points compound over years of always-on operation.

Concretely: 750W for a 150-250W card, 850W for a 300-350W card, 1000W+ for a 575W RTX 5090. Verified August 2026 listings put 650-850W Gold units around $80-120, and 1200-1600W units at roughly $400-550. The PSU is not your budget problem this year.

Cooling And Case

Sustained load changes the cooling brief too. You are not chasing peak temperature in a 10-minute benchmark; you are holding a steady thermal load overnight.

Air over AIO for a 24/7 box. Fewer failure modes, no pump to die at 3am.
Front intake aimed at the GPU. The card, not the CPU, is generating most of the heat.
Check the 170mm clearance if you buy an NH-D15 G2. Many mid-towers stop at 165mm, and a tall front fan can foul the RAM.
Dust filters matter when the intake runs continuously.

Should You Build This Now?

Split the decision by part.

Buy now: motherboard, PSU, cooler, case, CPU. None of these moved much, and waiting gains you nothing.

Think hard: the RAM. 64GB at $680-880 is roughly four times the 2025 price for the same product. No analyst expects relief before 2027. If you already own a DDR5 kit, reuse it. If you can run 32GB for two more quarters, do that and add the second kit later — which is exactly why we said 2x32 in two slots rather than 4x16 in four.

Buy on need: the GPU. Prices rose three times in 2026 and may rise again, so waiting is not obviously cheaper. Buy the VRAM tier your models require and stop.

Compare against a prebuilt box before you commit — see Is a $5K local AI rig worth it? for the total-cost view.