128GB Local AI Machines in Korea: Prices for Mini PCs, GB10, Mac Studio and Multi-GPU

How much does a local AI machine with 128GB of memory cost in Korea? Even at the same 128GB, finished products range from about KRW 5 million to nearly KRW 12 million, and the build-it-from-cards route runs from six used RTX 3090s at about KRW 10.65 million to four RTX 5090s at about KRW 39.9 million (board and power supply not included). This article rewrites a YouTube video and separates every number into official data, measurements reported by media and bloggers, Korean retailer listed prices, calculations, and our own measurement. Prices are as checked on 2026-11-02.

The conclusion in 3 lines

• The cheapest route within the range we checked is a Strix Halo mini PC. The GMKtec EVO-X2 128GB configuration is KRW 5,017,900 (1TB) and KRW 5,288,400 (2TB) as listed on the manufacturer's Korean store.
• Among new 128GB finished products confirmed on sale in Korea, the one with the highest official memory bandwidth is the Mac Studio M5 Max (40-core GPU) 128GB configuration: KRW 8,710,000 and 614GB/s. However, reported speeds were measured under different conditions, so this is not a ranking.
• Some people do not need 128GB. If a model file plus its context memory comes to around 20GiB, the file fits even on a machine without 128GB.

This article uses five kinds of numbers

The scope is fixed machines that sit on a desk or at home. Laptops use different criteria and are left out. Finished products and desktops you assemble yourself are kept separate, and because the missing parts differ, they are not ranked against each other.

Do you really need 128GB? Start with file size

The question that stood out most in comments on overseas videos was "so how much does it cost?", and the video organizes Korean won prices first. Before that, it looks at model file sizes. gpt-oss-120b is 59.0GiB, Qwen3.5-122B-A10B is 71.3GiB, and Qwen3.8 Flash-Next at roughly 4-bit is 87.2GiB, all sizes that can load on a 128GB-class machine. Qwen3.8-27B at 4-bit is 16.4GiB, so it does not need a 128GB machine. GLM-5.3-Flash and DeepSeek V4 Flash fit in the 80 to 90GiB range when shrunk to roughly 1-bit, but we could not check their quality. DeepSeek V4.1 Flash is over 200GB and Kimi K3 is over 400GB, so even 128GB is not enough.

You cannot use all 128GB. For Strix Halo, many sources say the Linux default is about 50%, around 120GiB if raised, and up to 96GB on Windows (reported by others). We could not find an official figure for the Mac or GB10; assuming 75% for the Mac gives about 96GiB (calculation). The total requirement with a 32K-token context and buffers is about 61.1GiB for gpt-oss-120b, about 73GiB for Qwen3.5-122B-A10B, and about 89.0GiB for Flash-Next (calculation). All fit within 96GiB, but the last one is tight.

Five ways to get 128GB, and the Korean price table

Three are finished products (Strix Halo mini PCs, the GB10 family and Jetson Thor, Mac Studio) and two are assembled from parts (multiple GPUs, and server or desktop RAM plus a GPU with RAM offload).

Strix Halo mini PCs. GMKtec has a Korean store, and its prices are listed by the manufacturer, so the basis is the firmest.

Other manufacturers appear as overseas-seller listings, and we could not confirm domestic after-sales service. The trap is that the same product shows different prices on the list page and the detail page.

The GB10 family and Jetson AGX Thor. Official bandwidth is 273GB/s for both. Storage capacity differs between products, so this is not a like-for-like comparison.

Among price-comparison detail values, excluding overseas import-agent listings, the lowest in the GB10 family is the Lenovo PGX at KRW 8,933,639. The GX10, also 1TB, is listed at KRW 9,000,000 (1-year warranty) and KRW 9,500,000 (1+2-year warranty) depending on the warranty configuration, and the evidence grades differ. The Jetson AGX Thor is a developer kit for robotics and edge use. One seller lists it at KRW 9,000,000 and another at KRW 8,700,000 before VAT. An overseas-listed GX10 showed KRW 7,443,070 on the list but KRW 8,363,000 or more on the detail page, with a note that after-sales service may not be available. The KRW 1,000,000-range listings that look like 128GB RAM kits are the lowest option price, so we did not use them.

Mac Studio. According to the Apple Korea configurator, the M5 Max (40-core GPU) with 128GB and a 512GB SSD is KRW 8,710,000 (KRW 68,047 per GB of memory, calculation), with 614GB/s of bandwidth (the 32-core GPU configuration is 460GB/s). The M5 Ultra offers 96GB and then 256GB, so there is no 128GB configuration (96GB is KRW 9,490,000). Used M1 Ultra 128GB machines appear at asking prices of KRW 5,900,000 to KRW 6,500,000, but these are private sales.

Multiple GPUs. Counting card prices alone when filling 128GB with the sum of card memory: six used RTX 3090 24GB cards (144GB) are about KRW 10,650,000 (median asking price KRW 1,775,000 per card, 17 listings), four Radeon AI PRO R9700 cards are KRW 11,639,960, and four RTX 5090 cards are KRW 39,893,760. Five 3090s give only 120GB, which falls short, so it takes six. The board, power supply and case are not included, and the card power total alone for six 3090s is 2,100W. The used prices are median asking prices without warranty, and we could not check slots or noise.

Speed is considered separately from price

Memory bandwidth sets the ceiling, not the speed itself. Strix Halo is 256, GB10 and Jetson Thor are 273, and the Mac Studio M5 Max is 614GB/s (official). Graphics cards are per-card values, not the combined speed of several cards.

gpt-oss-120b reads about 2.77GB per token, so the ceiling is about 92.5 tokens/s at 256 and about 222 tokens/s at 614 (calculation). The published reports are below, and because all conditions differ they are not a ranking.

The reported values were 29 to 56% of the calculated ceiling, and the Mac got lower as the context got longer. Higher bandwidth points in the faster direction, but it is not a formula that predicts the number, and because conditions differ we cannot say which is faster or by how much.

The reality of RAM offload and used 3090s

RAM offload splits a model between system RAM and card memory, so the RAM side is slower. Desktop DDR5-5600 dual-channel bandwidth is 89.6GB/s (calculation), below the 936GB/s of a single RTX 3090. On the same Strix Halo, a large dense model was reported at 3 tokens/s and gpt-oss-120b, which has few active parameters, in the 50s. One blog reported 23.85 tok/s running gpt-oss-120b with an RTX 4070 12GB and 64GB of dual-channel DDR5-6000, but it is a single report, so treat it as reference only.

For RAM prices, by price-comparison listing, server DDR5 128GB (two 64GB modules) is KRW 8,930,000, more than a single EVO-X2. Four Samsung desktop DDR5 32GB modules are KRW 3,163,880. DDR4 server RAM, two 64GB modules at KRW 2,880,000, is cheaper, but the boards that take it have a theoretical bandwidth of 85 to 94GB/s, about a third of Strix Halo, and adding an example board brings the total to KRW 4,538,510 or more.

Here is one measurement of ours, listed separately with its conditions.

We ran Qwen3.8 Flash-Next shrunk to Q2_0 (a 66.4GB file) with Strata 0.1.43.1 on an RTX 3060 12GB server and got 36.4 to 43.9 tok/s for prose generation with thinking on (measured 2026-10-13, median of 3 runs, three settings 43.9 / 39.0 / 36.4). The server's container RAM limit was 42.6GiB. We did not measure speed on a 128GB-RAM desktop; the claim goes only as far as saying this file fits in 128GB. If you already have a desktop, the parts to replace are RAM and a card: four Samsung DDR5 32GB modules at KRW 3,163,880 plus a used RTX 3060 12GB at a median asking price of KRW 350,000 comes to about KRW 3,513,880, without the board, CPU, power supply or case (calculation).

Three-year total cost, subscriptions and electricity

Dividing by price alone, a KRW 29,000 per month subscription (ChatGPT Plus, Google AI Pro) equals 182 months for the EVO-X2 2TB, 300 months for the Mac Studio and 308 months for the Lenovo PGX. A KRW 159,000 per month subscription (the lowest ChatGPT Pro tier) equals 33, 55 and 56 months respectively (all calculation, and it does not assume equal quality). To recover only the hardware cost within three years (36 months), the monthly subscription fee would have to be at least KRW 146,900 for the EVO-X2 2TB, KRW 241,944 for the Mac Studio and KRW 248,157 for the Lenovo PGX. Claude is priced in dollars, so it was left out of this comparison (the exchange rate was not confirmed).

Electricity is an upper-bound calculation using the manufacturer's maximum power (adapter rating and similar), 8 hours a day for 30 days. Finished products come to about KRW 8,000 to 21,000 a month, and multiple cards come to about KRW 43,000 to 200,000 a month for card power alone. As a three-year total, the mini PC is KRW 5,288,400 for the hardware plus KRW 300,000 to 720,000 in electricity, about KRW 5.59 million to 6.01 million, while 36 months of subscription is KRW 1,044,000 at KRW 29,000 a month and KRW 5,724,000 at KRW 159,000 a month. This does not assume equal performance or limits.

Criteria by situation, and what we could not confirm

These are the criteria the video lays out (they are neither a purchase recommendation nor a ranking).

• If budget comes first, the lower-priced option within the range we checked is the Strix Halo mini PC (GMKtec EVO-X2, KRW 5,017,900 for 1TB and KRW 5,288,400 for 2TB).
• If the file plus context size is around 20GiB, you may not need 128GB, and at 60 to 90GiB it is a size a 128GB-class machine can load.
• For speed (bandwidth), among new 128GB finished products confirmed on sale in Korea, the Mac Studio M5 Max has the highest at 614GB/s (GB10 and Thor 273, Strix Halo 256). Reported speeds were measured under different conditions, so this is not a ranking.
• For DIY builds we could not check boards, power supplies or noise, and for RAM offload our measured 36 to 44 tok/s is not a 128GB-RAM value.

What we could not confirm: noise and real-world power draw (the only noise figure is one manufacturer claim of 32dB at 140W for Beelink), domestic after-sales service, customs and warranty (outside the GMKtec Korean store), whether the GX10 and server RAM listings are new and what their specs are, speed data for the Jetson Thor, and how much of GB10's memory is usable. We also could not find Korean sales evidence for Framework Desktop and Corsair (Minisforum AI X1 Pro was excluded because its maximum is 96GB).

Reference date and limits

• The video was published 2026-11-02, and prices were checked on 2026-11-02 (Korean retailer listed prices). This article reflects the video's publication date, and prices may have changed since.
• This is a comparison within the range we checked, not every product. Finished products, RAM only, and cards only are not ranked against each other because the missing parts differ.
• Prices mix official store and configurator values with price-comparison and used asking-price listings, and the type of each value is marked in the text.