Lenovo ThinkStation PGX SFF, NVIDIA GB10 Grace Blackwell ( ARM , system-on-chip (SOC)), 128GB, 4TB SSD, Thunder Grey

★★★★★ 4.8 (6)

The NVIDIA GB10 Grace Blackwell Superchip with a 20-core ARM CPU and integrated Blackwell GPU delivers 128GB of DDR5 memory and a 4TB NVMe SSD in a 1.20kg SFF chassis. Its 240W PSU and Wi-Fi 7 connectivity make it a compact, low-power AI development station, though its AI LLM score is weak. Best for developers who need a portable, quiet workstation for prototyping and running small AI models, not for heavy LLM training.

CPU NVIDIA Grace Blackwell GB10 Processor
RAM 128 GB
Storage 4 TB
GPU Integrated NVIDIA Blackwell GPU
Form Factor sff
psu w 240
OS NVIDIA DGX OS
Lenovo ThinkStation PGX SFF, NVIDIA GB10 Grace Blackwell ( ARM , system-on-chip (SOC)), 128GB, 4TB SSD, Thunder Grey desktop
31 Overall Score

No current retailer price

Also available in:

Snapshot

The 30-Second Version

128GB of DDR5 and a 4TB SSD in a 1.2kg box make this a memory monster, but the GPU is one of the weakest we've seen at the 8th percentile. It's built for AI prototyping on NVIDIA DGX OS, not general use. Prices swing by $1,650 in the US, so shop carefully.

Pros & Cons

Pros

  • 128GB DDR5 is the best RAM config in our entire database 100th
  • 4TB NVMe SSD puts it in the top 2% for storage 98th
  • Weighs just 1.2kg, absurdly portable for a workstation 73rd
  • Wi-Fi 7 and 4x USB-C keep connectivity current 70th
  • Runs cool and quiet under multi-day stress loads per owners

Cons

  • GPU sits at the 8th percentile, a real weak spot
  • Port selection lands at the 39th percentile, below average
  • CPU is middle of the pack at the 50th percentile
  • Runs NVIDIA DGX OS, not standard Windows or Linux
  • No HDMI 2.1a beyond 4K@60Hz, so no high-refresh displays

What owners think

The Word on the Street

4.8/5 (6 reviews)
👍 Owners consistently say it runs cooler and looks cleaner than the DGX Spark, with identical software compatibility.
👍 Multiple buyers report it handles multi-day stress workloads and basic LLM inference without issues.
🤔 A few reviewers note it's a solid dev box but only suitable for basic inference with common models, not heavy lifting.

How owner sentiment changed over time

Exclusive

Based on when customers actually wrote their reviews - so you can see whether early praise held up.

Owner sentiment has cooled since launch
1★2★3★4★5★Q4 '25: 5.0★ · 1 reviewQ2 '26: 4.0★ · 1 review11Q4 '25Q2 '26
Avg ratingHappy (4-5★)Unhappy (1-2★)Bar height = number of reviews

Based on 2 dated customer reviews, grouped by calendar quarter. Period analysis is in English.

The proof

Performance

The 128GB of unified DDR5 is the star of the show. That's the absolute best we've seen in this category, and it means you can load models that would choke most desktops. The 4TB NVMe SSD is also a standout, giving you room for multiple large model files without juggling external drives. The 20-core ARM CPU sits right at the 50th percentile, which is about average for this class, but that's not really the point. The GB10 Superchip is designed around memory bandwidth and AI workloads, not raw CPU throughput.

The integrated Blackwell GPU is where things get rough. At the 8th percentile, it's one of the weakest GPUs we've tracked in this category. That's not a dealbreaker for its intended use, since the Superchip's unified memory architecture handles AI inference differently than a discrete gaming card, but it does mean this machine won't flex for gaming, video editing, or anything that leans on traditional GPU muscle. The 240W PSU tells the same story: efficient, but not built for expansion.

Performance Percentiles

CPU 50.3
GPU 7.7
RAM 99.7
Ports 38.8
Storage 98.1
Reliability 70
Social Proof 73.4

Specifications

Full Specifications

Processor

CPU NVIDIA Grace Blackwell GB10 Processor
Cores 20

Graphics

GPU Integrated NVIDIA Blackwell GPU
Type Discrete

Memory & Storage

RAM 128 GB
RAM Generation DDR5
Storage 4 TB
Storage Type NVMe SSD

Build

Form Factor sff
PSU 240
Weight 1.2 kg / 2.6 lbs

Connectivity

USB-C Ports 4
HDMI HDMI® 2.1a (supports resolution up to 4K@60Hz)
Wi-Fi Wi-Fi 7
Bluetooth Bluetooth 5.3
Ethernet Integrated Ethernet

System

OS NVIDIA DGX OS

vs Competition

Against the Apple Mac Studio M5 Max, the ThinkStation PGX wins on RAM capacity, since 128GB unified memory beats what Apple offers at this price point, but the Mac Studio's GPU absolutely crushes the PGX's integrated Blackwell for graphics and creative work. The Dell Desktop ECT1250 and HP OMEN TPC-W062-MT are traditional desktops with discrete GPUs that will run circles around the PGX for gaming and general productivity, but neither can touch the 128GB RAM figure. The ASUS ROG GM700 and CLX SET TGMSRTU5204BM are gaming-focused machines that make the PGX look silly on GPU performance, but they're built for a completely different job. If you need CUDA-adjacent AI prototyping in a tiny, quiet box, the PGX has no direct rival here. If you need anything else, the competitors win.

Spec Lenovo ThinkStation PGX SFF Apple Mac Studio M3 Ultra Dell Desktop ECT1250 HP OMEN TPC-W062-MT ASUS ROG GM700 CLX SET TGMSETRTU5204BM
CPU NVIDIA Grace Blackwell GB10 Processor Apple M3 Ultra Intel Core Ultra 7 265 Intel Core Ultra 9 285K AMD Ryzen 7 9800X3D Intel Core i9 14900KF
RAM (GB) 128 96 32 64 32 64
Storage (GB) 4096 1024 2048 2048 2048 8192
GPU Integrated NVIDIA Blackwell GPU Apple M3 Ultra 60-core Intel UHD Graphics NVIDIA GeForce RTX 5090 NVIDIA GeForce RTX 5070 Ti NVIDIA GeForce RTX 5070
Form Factor sff sff sff tower mid-tower mid-tower
Psu W 240 - 180 - 850 850
OS NVIDIA DGX OS macOS Windows 11 Home Windows 11 Home Windows 11 Home Windows 11 Home
Compare Compare Compare Compare Compare
Product CPUGPURAMPortsStorageReliabilitySocial Proof
Lenovo ThinkStation PGX SFF 50.37.799.738.898.17073.4
Apple Mac Studio M3 Ultra Compare 99.77695.793.653.699.386.2
Dell Desktop ECT1250 Compare 88.458.781.774.582.87099.1
HP OMEN TPC-W062-MT Compare 98.197.893.899.290.17067
ASUS ROG GM700 Compare 80.593.181.796.190.137.399.6
CLX SET TGMSETRTU5204BM Compare 94.989.396.584.199.111.694.7

Price

Value & Pricing

Pricing is all over the place depending on where you look. In the US, listings range from $5,079 to $6,729 across three vendors, a spread of $1,650 for what should be the same machine. The lowest US price is the one to chase if you're buying. In Germany, the gap is even wider: €5,592 to €7,699 across two listings, a €2,107 difference. That kind of variance means you should absolutely shop around before pulling the trigger. At the low end, you're paying a premium for the Superchip and that massive RAM pool, but it's defensible for AI prototyping. At the high end, you're getting fleeced.

Read more

Overview

The headline numbers here are wild: 128GB of DDR5 and a 4TB NVMe SSD in a 1.2kg box. That RAM figure sits at the 100th percentile in our database, and the storage isn't far behind at 98th. This is a tiny workstation that packs more memory than most full-size towers we've tested. The catch is the GPU, which lands at the 8th percentile, and that tells you everything about what this machine is and isn't for.

Lenovo calls the ThinkStation PGX a workstation, but it's really a purpose-built AI dev box. The NVIDIA GB10 Grace Blackwell Superchip combines a 20-core ARM CPU with an integrated Blackwell GPU on one package. It runs NVIDIA DGX OS, not Windows or standard Linux. If you're shopping for a general-purpose desktop, keep scrolling. If you want a quiet, compact machine for prototyping LLMs and running inference, this thing is genuinely interesting.

Common Questions

Q: Can I run Windows or standard Linux on the ThinkStation PGX?

No. It ships with NVIDIA DGX OS, which is built specifically for the GB10 Grace Blackwell Superchip. Standard Windows or consumer Linux distros aren't supported, so plan your workflow around DGX OS.

Q: Is the integrated GPU good enough for gaming or video editing?

Not really. The integrated Blackwell GPU ranks at the 8th percentile in our database, which is one of the weakest we've tracked in this category. It's designed for AI inference and compute tasks, not gaming or GPU-heavy creative work.

Q: How much RAM and storage does it actually have?

It comes with 128GB of DDR5, which is the best RAM configuration in our entire database, and a 4TB NVMe SSD, which lands in the top 2% for storage. That's enough to load large language models locally without external drives.

Who Should Skip This

Skip this if you need a general-purpose desktop. The GPU is one of the weakest we've seen at the 8th percentile, so gaming, video editing, and 3D rendering are off the table. The port selection is below average at the 39th percentile, and the CPU is just middle of the pack. If you're not specifically doing AI development or inference on NVIDIA's software stack, you're paying a lot for RAM you won't fully use and a GPU that will frustrate you.

Verdict

The ThinkStation PGX is a niche machine that nails its niche. The 128GB of RAM and 4TB SSD are genuinely impressive, and owners report it handles multi-day stress loads without breaking a sweat. The 4.8/5 rating from six reviews backs that up, though the sample size is small. But the 8th percentile GPU and below-average port selection mean this is not a general-purpose workstation. Buy it for AI development and inference, not for spreadsheets or gaming. And if you do buy it, hunt down the lowest price, because the vendor spread is brutal.

Usage Scores

Overall (30.6)AI/LLM (15.9)Gaming (21.4)Portability (32)Creator (26.5)Business (35.1)Developer (33.8)Home Office (38.2)Workstation (30.4)
  • Overall30.6
  • AI/LLM15.9
  • Gaming21.4
  • Portability32
  • Creator26.5
  • Business35.1
  • Developer33.8
  • Home Office38.2
  • Workstation30.4

Similar Products