Lenovo ThinkStation PGX SFF, NVIDIA GB10 Grace Blackwell ( ARM , system-on-chip (SOC)), 128GB, 1TB SSD, Thunder Grey

The ThinkStation PGX packs 128GB of DDR5, the absolute best in our database, into a 1.2kg SFF chassis built around NVIDIA's GB10 Grace Blackwell Superchip. The GPU is one of the worst we've seen at the 8th percentile, and the CPU is average. It's a specialized AI dev box with a 4.8 out of 5 rating from early buyers, but the USD 1,650 price spread means you need to shop carefully.

Let's talk about what this machine actually does well. The 128GB of DDR5 is the absolute best right now, and it's paired with a 20-core ARM CPU that sits right around the middle of the pack at the 51st percentile. That's not a typo. The CPU is average. The magic is in the memory bandwidth and the fact that the whole system is a system-on-chip, which means the CPU and GPU share that massive memory pool directly. For AI inference work, that's huge. You can load a 70-billion parameter model into memory and actually run it without swapping to disk.

The storage is a 1TB NVMe SSD, which lands in the 69th percentile. Solid, but not a standout. The GPU is where things get rough. The integrated Blackwell GPU sits in the 8th percentile, which is a real letdown for anything outside AI workloads. The 240W power supply tells you everything you need to know. This isn't a machine built to push pixels or render 3D scenes. It's built to run neural networks efficiently. The Wi-Fi 7 and HDMI 2.1a support are nice touches, but the port selection is underwhelming at the 39th percentile. Four USB-C ports and an Ethernet jack is a pretty lean setup for a workstation.

The obvious comparison is the Apple Mac Studio M5 Max. The Mac Studio has a much stronger GPU and better port selection, but it tops out at less RAM and uses a different architecture. For AI inference specifically, the PGX's 128GB of unified memory gives it an edge for large models. The Dell Desktop ECT1250 and HP OMEN GT16-0364 are traditional workstations with discrete GPUs that will crush the PGX in graphics and rendering tasks, but they can't touch the memory capacity. The CLX SET TGMSETRTU5204BM and MSI Aegis ZS2 are gaming-focused machines that make the PGX look silly in GPU benchmarks. The PGX is in a weird spot. It's not a great general workstation, but it's one of the best AI dev boxes you can buy right now if memory is your bottleneck.

Here's where things get interesting. The price spread on this thing is wild. In the US, you're looking at USD 5,079 to USD 6,729 across three listings. That's a USD 1,650 difference for the same machine. The best deal is at the low end, obviously, and if you're paying the high end you're getting taken for a ride. In Germany, the spread is even worse. EUR 5,592 to EUR 7,699 across two listings, a EUR 2,107 gap. For a machine with an average CPU and a bottom-tier GPU, the value proposition hinges entirely on that 128GB of RAM and the Grace Blackwell architecture. If you need local AI inference with massive memory, it's worth it. If you don't, you're paying workstation money for a very specialized box.

The headline number here is the RAM. 128GB of DDR5 puts this thing in the 100th percentile of our database, which is a fancy way of saying no other machine in this category ships with more memory right now. That's the whole pitch, really. The ThinkStation PGX is a tiny 1.2kg box built around NVIDIA's GB10 Grace Blackwell Superchip, and it's aimed squarely at developers who want to run large language models locally without renting cloud time. The 20-core ARM CPU and integrated Blackwell GPU are fine, but the memory is the star of the show.

But here's the catch. The GPU lands in the 8th percentile, which is one of the worst we've seen for a workstation in this price bracket. The integrated Blackwell chip is designed for AI inference, not gaming or traditional rendering workloads. So if you're expecting Mac Studio levels of graphics performance, you'll be disappointed. This is a specialized tool, not a general-purpose powerhouse. The 4.8 out of 5 rating from six early buyers suggests the people who get it, really get it. But it's not for everyone.

Skip this if you need a general-purpose workstation. The GPU is in the 8th percentile, which is one of the worst we've seen, and the CPU is dead average at the 51st percentile. If you're doing video editing, 3D rendering, or gaming, the Mac Studio M5 Max or a traditional workstation like the Dell Desktop ECT1250 will serve you much better. The port selection is also underwhelming at the 39th percentile, with just four USB-C ports and no legacy connections. And if you're not specifically working with AI models that need massive memory, you're paying a premium for 128GB of RAM you'll never fully use. The AI and LLM workload score of 15.4 out of 100 is also a red flag if you're expecting heavy training performance.

The ThinkStation PGX is a specialized tool that does one thing exceptionally well: running AI models locally with a massive memory pool. The 128GB of DDR5 is best-in-class, and the Grace Blackwell architecture is genuinely interesting for developers. But the 8th percentile GPU and average CPU mean this is not a general-purpose workstation. The 4.8 out of 5 rating from early adopters tells you the target audience is happy. If you're an AI developer who needs to run large models without cloud costs, this is a compelling option, especially at the lower end of the price range. If you need a workstation for anything else, look elsewhere.