TodayToday's edition
Local AI
AI workstations, desktop supercomputers and AI PCs for running large models locally, compared on memory, bandwidth, AI compute, measured speed and price, with the source for every figure.
Memory capacity decides the largest model a machine can hold; memory bandwidth decides how fast it writes. Rows run from the fastest memory down. A product name links to our review when one is filed.
| Product | Chip | Memory | Bandwidth | AI compute | Measured speed | Largest model | Power | Price | Availability | Form factor |
|---|---|---|---|---|---|---|---|---|---|---|
| HP ZGX Fury AI StationHP | NVIDIA GB300 Grace Blackwell Ultra Superchip (Grace 72-core Neoverse V2 + Blackwell Ultra GPU)[1] | 252 GB HBM3e GPU + 496 GB LPDDR5x CPU (4x 128 GB SOCAMM); 748 GB unified[1] | 7.1 TB/s (GPU HBM3e); 396 GB/s (CPU LPDDR5x)[1] | — | — | up to 1 trillion parameters (HP: model harness quantizes at FP4) |
| — |
| Japan list price from 28,490,000 yen incl. tax (sales from Oct 27, 2026); no US list price published[2] |
| Japan sales from Oct 27, 2026; Windows support planned from December 2026[2] |
| tower, mountable in a 5U rack with included rails; needs a dedicated 200-240 V circuit[1] |
| NVIDIA DGX StationNVIDIA (systems built by ASUS, Dell, GIGABYTE, HP, MSI, Supermicro and Exxact) | NVIDIA Grace, 72-core Arm Neoverse V2[3] | 748 GB coherent: 252 GB HBM3e (GPU) + 496 GB LPDDR5X (CPU)[3] | GPU HBM3e 7.1 TB/s; CPU LPDDR5X 396 GB/s; NVLink-C2C 900 GB/s between them[3] | 15 PFLOPS FP4 (without sparsity)[3] | Over 26,000 output tokens/s aggregate on Qwen3.6-35B-A3B NVFP4 at 512 concurrent users (24-30-token input, 1,024-token output), Gigabyte W775-V10-L01; idle never below 250 W.[4] | up to 1 trillion parameters[3] | 1,600 W total system power (20 A circuit required)[3] | No list price from most OEMs (sales inquiry only); ServeTheHome heard about $100K, up to $120-125K, and unlikely under $80-85K[5] | Shipping from ASUS, Dell, GIGABYTE, HP, MSI, Supermicro and Exxact; DGX Station for Windows expected from partners in December 2026[6] | deskside workstation tower (built by OEM partners)[3] |
|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation EditionNVIDIA (sold through PNY, TD SYNNEX and workstation makers) | GB202: 188 SMs, 24,064 CUDA cores, fifth-gen Tensor Cores, fourth-gen RT cores, 2,617 MHz boost (Workstation Edition); MIG up to 4x 24 GB[7] | 96 GB[8] | 1,792 GB/s[8] | 4,000 AI TOPS (FP4, with sparsity) (sparse)[8] | LM Studio: Mistral Small (26 GB) at 42.4 tokens/s vs 17 on an RTX 5090; Gemma 3 27B at 29 vs 5; tied with the RTX 5090 at about 81 tokens/s on DeepSeek Llama 8B Distil, where VRAM does not limit.[9] | — | 600 W total board power (Workstation Edition)[8] | $16,000 on NVIDIA's US marketplace (Workstation Edition), listed out of stock[10] | From April 2025 via PNY and TD SYNNEX; from May 2025 in BOXX, Dell, HP, Lambda and Lenovo workstations[11] | PCIe Gen 5 card, dual slot (5.4 in H x 12.0 in L)[12] |
|---|
| Apple Mac Studio (M5 Ultra)Apple | Apple M5 Ultra (up to 36-core CPU, 80-core GPU, 32-core Neural Engine)[13] | up to 512 GB unified (96 GB on 30-core M5 Ultra; 256 GB or 512 GB on 36-core/80-core M5 Ultra)[13] | 1.2 TB/s[13] | — | Qwen 3.8 27B 4-bit MLX in LM Studio: about 55 tok/s on M5 Ultra (36-core, 256 GB) vs about 40 on M3 Ultra (BGR)[14] | — | 480 W maximum continuous power[13] | $22,299 minimum with 512 GB ($6,799 36-core CPU/80-core GPU M5 Ultra config plus $15,500 memory upgrade; Oct 28, 2026)[15] | M5 Ultra models in stores since Sep 22, 2026; 512 GB option due late October 2026[16] | compact desktop (9.5 x 19.7 x 19.7 cm; 3.6 kg with M5 Ultra)[13] |
|---|
| NVIDIA RTX Spark (N1X) PCsNVIDIA (MediaTek collaborated on the custom CPU design) | RTX Spark N1X superchip: 20-core NVIDIA Grace CPU + 6,144-core Blackwell RTX GPU (second N1X: 18-core CPU, 5,120-core GPU)[17] | Up to 128 GB LPDDR5X (20-core CPU / 6,144-core GPU version); up to 64 GB (18-core / 5,120-core laptop version)[17] | Up to 300 GB/s: LPDDR5X-9400 on a 256-bit bus (NVIDIA's product page lists no bandwidth)[18] | — | UL Procyon AI text generation (Phi 3.5): 2,703 points, preliminary, Surface Laptop Ultra 128 GB. Measured by Computerwoche.[19] | up to 120B parameters with 1M context[17] | Laptops 45-80 W; compact desktop 140 W (NVIDIA spec tables)[17] | Laptops from $2,599; pricing and configurations vary by region[20] | Laptops available now from ASUS, Dell, HP, Lenovo, Microsoft and MSI, from $2,599, 128GB configurations on shelves (2026-10-28); compact desktops 'available for sale in November' (NVIDIA, 2026-10-07)[21] | slim Windows laptops and compact desktops[17] |
|---|
| Framework Desktop (Ryzen AI Max+ PRO 495, 192 GB)Framework | AMD Ryzen AI Max+ PRO 495 (16 cores/32 threads, up to 5.2 GHz; Radeon 8065S, 40 CUs)[22] | 192 GB LPDDR5X-8533 soldered[23] | 273 GB/s (256-bit)[23] | — | — | 300B-class models at 4-bit quantization (192 GB configuration)[24] | 120 W sustained / 140 W boost; 400 W FlexATX power supply[24] | from $6,799 (DIY Edition pre-order, Oct 30, 2026)[25] | Pre-orders opened Sep 30, 2026; shipments start in November 2026[23] | 4.5 L mini-ITX desktop (96.8 x 205.5 x 226.1 mm, 3.1 kg)[24] |
|---|
| Lenovo ThinkStation PGXLenovo | NVIDIA GB10 Grace Blackwell Superchip (20-core Arm: 10x Cortex-X925 + 10x Cortex-A725)[26] | 128 GB LPDDR5X unified (256-bit)[26] | 273 GB/s[26] | — | — | up to 200B parameters (405B with two ThinkStation PGX systems)[26] | 240 W adapter[26] | about $4,100 for the 1 TB configuration (Mar 2026)[27] | — | mini PC (150 x 150 x 50.5 mm)[26] |
|---|
| MSI EdgeXpert (MS-C931)MSI | NVIDIA GB10 Grace Blackwell Superchip (Arm 20-core: 10 Cortex-X925 + 10 Cortex-A725)[28] | 64 GB or 128 GB LPDDR5x unified (256-bit)[28] | 273 GB/s[28] | 1 PFLOPS FP4 (sparse)[28] | — | up to 200B parameters on 128 GB (405B with two units, 700B with four)[28] | 240 W external power supply (48 V)[29] | No list price; MSI sells via quote (Get a quote)[28] | 128 GB on sale via quote; 64 GB version listed as coming soon (Oct 2026)[28] | mini PC (151 x 151 x 52 mm, 1.2 L, 1.2 kg)[28] |
|---|
| NVIDIA DGX SparkNVIDIA (partner GB10 systems from Acer, ASUS, Dell, GIGABYTE, HP, Lenovo and MSI) | GB10: 20-core Arm, 10 Cortex-X925 + 10 Cortex-A725[30] | 128 GB; a 64 GB configuration is sold only through OEM partners[30] | 273 GB/s[30] | Up to 1 PFLOP FP4 (theoretical, using the sparsity feature) (sparse)[30] | Qwen 3.8 27B (dense) with multi-token prediction: decode of about 20 tokens/s, with time-to-first-token staying interactive out to the full 262K context.[31] | up to 200B parameters (400B with two 128 GB units, 700B with four)[30] | GB10 TDP 140 W (chip, CPU and GPU); 240 W system power supply[30] | $6,950 (128 GB, 4 TB) on NVIDIA's enterprise marketplace, listed out of stock; 64 GB partner systems start at $4,999[32] | Shipping from the week of 2025-10-13; orders on NVIDIA.com from 2025-10-15, plus partner systems and Micro Center (U.S.)[33] | mini PC (150 x 150 x 50.5 mm, 1.2 kg)[30] |
|---|
| NVIDIA DGX Spark 64 GBNVIDIA (sold by Acer, ASUS, Dell, Gigabyte, HP and MSI) | NVIDIA GB10 Grace Blackwell Superchip (same chip as the 128 GB model)[34] | 64 GB LPDDR5X unified (two units pool 128 GB over ConnectX-7)[34] | 273 GB/s (256-bit; NVIDIA spec table covers both 64 GB and 128 GB)[30] | 1 PFLOPS FP4 (sparse; NVIDIA spec table covers both memory configurations)[30] | — | up to 100B parameters (200B with two clustered 64 GB units)[34] | — | from $4,999 (partner systems, Oct 2026; no Founders Edition)[34] | On sale from Oct 23, 2026 through Acer, ASUS, Dell, Gigabyte, HP and MSI only[34] | — |
|---|
| AMD Ryzen AI Halo Developer PlatformAMD | AMD Ryzen AI Max+ 395 (16 Zen 5 cores; Radeon 8060S, 40 CUs; XDNA 2 NPU)[35] | 128 GB LPDDR5x-8000 unified[35] | 256 GB/s[35] | — | gpt-oss-120b, vLLM online serving, equal ISL/OSL 256/256: 222 tok/s aggregate at batch 64 vs 701 on DGX Spark (StorageReview)[36] | up to 200B parameters (AMD's claim, as reported)[36] | 120 W[35] | $3,999 with 2 TB SSD (Jul 2026)[36] | Pre-orders from June 2026 exclusively via Micro Center; in stores from July 10, 2026[36] | mini PC (150 x 150 x 45.4 mm, under 1.2 kg)[35] |
|---|
| Framework Desktop (Ryzen AI Max+ 395)Framework | AMD Ryzen AI Max+ 395 (16 cores/32 threads, 40 graphics cores)[25] | 128 GB LPDDR5x-8000 soldered (32 GB and 64 GB options)[24] | 256 GB/s[37] | — | — | Llama 3.3 70B Q6 on one unit; full DeepSeek R1 671B across several linked units[37] | 120 W sustained / 140 W boost; 400 W FlexATX power supply[24] | $3,449 (DIY Edition, Max+ 395 with 128 GB, Oct 30, 2026)[25] | Out of stock (Max+ 395 configurations, Oct 30, 2026)[25] | 4.5 L mini-ITX desktop (96.8 x 205.5 x 226.1 mm, 3.1 kg)[24] |
|---|
| Acer Veriton GN100 AI Mini WorkstationAcer | NVIDIA GB10 Grace Blackwell Superchip (20-core Arm CPU)[38] | 128 GB LPDDR5x coherent unified[38] | — | — | gpt-oss-120b, vLLM online serving, equal ISL/OSL: 69.65 tok/s at batch 1 rising to 713.18 tok/s aggregate at batch 64 (StorageReview)[39] | up to 405B parameters with two linked GN100 units[38] | — | from $3,999 in North America (announced Sep 3, 2025; prices vary by region)[38] | — | mini PC (150 x 150 x 50.5 mm, under 1.5 kg)[38] |
|---|
| ASUS Ascent GX10ASUS | NVIDIA GB10 Grace Blackwell Superchip (20-core Arm: 10 Cortex-X925 + 10 Cortex-A725)[40] | 64 GB or 128 GB LPDDR5X-8533 unified[40] | — | — | gpt-oss-120b, vLLM online serving, equal ISL/OSL: 70 tok/s at batch 1 rising to 680 tok/s aggregate at batch 64 (StorageReview)[41] | up to 200B parameters on 128 GB (100B on 64 GB); up to four units cluster to 512 GB[42] | 240 W adapter[40] | $5,999 (128 GB, 1 TB, GX10-GG0010BN, ASUS eShop US, Oct 30, 2026)[43] | Out of stock on ASUS eShop US (Oct 30, 2026)[43] | mini PC (150 x 150 x 51 mm, 1.48 kg)[40] |
|---|
| Dell Pro Precision with GB10 (model FCM1253; launched as Dell Pro Max with GB10)Dell | NVIDIA GB10 Grace Blackwell Superchip with ConnectX-7 SuperNIC[44] | 128 GB LPDDR5x up to 8533 MT/s onboard (64 GB option also listed)[44] | — | — | gpt-oss-120b, vLLM online serving, equal ISL/OSL: 68.57 tok/s at batch 1 rising to 670.83 tok/s aggregate at batch 64 (StorageReview)[45] | — | 280 W USB-C AC adapter[44] | $6,593 (128 GB, 2 TB QLC SSD, Dell.com US, Oct 30, 2026)[44] | Orderable on Dell.com US (Oct 30, 2026)[44] | mini PC (150 x 150 mm, 51 mm peak height, 1.31 kg)[44] |
|---|
| Dell Pro Precision with GB300 (model FCT6263; announced as Dell Pro Max with GB300)Dell | NVIDIA GB300: Grace 72-core Neoverse V2 with ConnectX-8 SuperNIC[46] | 252 GB HBM3e GPU + 496 GB LPDDR5X CPU[46] | — | — | — | — | — | $159,283 (Dell.com US default config with RTX PRO 2000 display GPU and 16 TB SSD, Oct 30, 2026)[46] | Orderable on Dell.com US (Oct 30, 2026)[46] | desktop tower (569 x 231.6 x 610.5 mm, up to 38.67 kg)[46] |
|---|
| GIGABYTE AI TOP ATOMGIGABYTE | NVIDIA Grace Blackwell GB10 Superchip[47] | 128 GB unified system memory[47] | — | — | gpt-oss-120b, vLLM online serving, equal ISL/OSL: 60.71 tok/s at batch 1 rising to 649.62 tok/s aggregate at batch 64 (StorageReview)[48] | up to 200B parameters (405B with two linked units)[47] | — | — | On the market globally from Oct 15, 2025[47] | — |
|---|
| GMKtec EVO-X2 AI Mini PCGMKtec | AMD Ryzen AI Max+ 395 (16 cores/32 threads; Radeon 8060S)[49] | 64 GB or 128 GB onboard LPDDR5X-8000 (non-upgradeable)[49] | — | — | — | Qwen3 235B on the 128 GB version[49] | 120 W sustained, 140 W peak[49] | $3,499.99 (128 GB/1 TB, GMKtec US store, Oct 30, 2026; listed against $3,999.99)[49] | — | mini PC (193 x 185.8 x 77 mm)[49] |
|---|
| HP Z2 Mini G1a WorkstationHP | up to AMD Ryzen AI Max+ PRO 395 (Radeon integrated graphics)[50] | up to 128 GB unified, up to 96 GB assignable to the GPU, up to 8000 MT/s[50] | — | — | gpt-oss-120b GGUF (63.39 GB, full GPU offload) in LM Studio: close to 40 tok/s (StorageReview)[51] | — | 300 W internal power supply[50] | — | — | mini workstation with internal PSU (five fit across a 4U rack)[50] |
|---|
| HP ZGX Nano AI Station (G1n)HP | NVIDIA GB10 Grace Blackwell Superchip (20-core Arm Grace CPU, Blackwell GPU)[52] | 128 GB coherent unified system memory[52] | — | — | gpt-oss-120b, vLLM online serving, equal ISL/OSL: 69.6 tok/s at batch 1 rising to 722.9 tok/s aggregate at batch 64 (StorageReview)[53] | up to 200B parameters (405B with two linked ZGX Nano units)[52] | — | — | — | mini PC (150 x 150 x 51 mm, height excluding feet)[52] |
|---|
| MSI XpertStation WS300MSI (NVIDIA DGX Station platform) | NVIDIA GB300 Grace Blackwell Ultra superchip (DGX Station design)[54] | 252 GB HBM3e GPU + 496 GB LPDDR5X CPU (748 GB coherent)[54] | — | — | Noise about 37-40 dBA at lower loads and 46-49 dBA at full load; GPU idle ~148 W (1.3 kW cap), Grace module idle ~78 W (300 W cap), system idle never below 250 W. Measured by ServeTheHome[55] | — | 1.6kW system limit (1600W PSU); GPU 1.3kW cap, Grace module 300W cap; measured idle never below 250W (ServeTheHome)[56] | ServeTheHome calls it a $100K+ system; no MSI list price given[56] | — | liquid-cooled desktop tower (245.7 x 529.0 x 570.4 mm)[54] |
|---|
| Tenstorrent TT-QuietBox 2 (Blackhole)Tenstorrent | AMD Ryzen 7 9700X host + four Tenstorrent Blackhole Tensix processors on two PCIe cards (480 Tensix cores)[57] | 128 GB GDDR6 on the accelerators + 256 GB DDR5-5600 system memory[57] | — | — | — | — | 1,650 W 80 PLUS Platinum power supply[57] | $9,999 (Oct 30, 2026)[57] | In stock, ships in 2 weeks (Oct 30, 2026)[57] | liquid-cooled desktop workstation[57] |
|---|
Italic values are reported by an outlet or measured by an independent lab; the rest come from the maker. Measured speeds use different models, quantizations and tools, so read each with its source rather than against the others. Voltgrid News does not test hardware itself.