Lenovo ThinkStation PGX Review
July 8, 2026The Lenovo ThinkStation PGX (SHOP HERE) is not really a workstation in the traditional sense, but a workstation nonetheless. This is the age of AI, right? Released in October of 2025, this is an optimized solution for AI developers and researchers, data scientists, practitioners, students and application engineers. It’s designed specifically to address AI development and is powered by an NVIDIA GB10 Grace Blackwell superchip offering a full petaflop of AI performance.
Don’t be fooled by its size. The Lenovo ThinkStation PGX can tackle large generative AI models of up to 200 billion parameters. If that’s not enough to tackle the problem, you can connect another Lenovo ThinkStation PGX, using the integrated QSFP ports. It can be used as a stand-alone AI workstation, integrated with your existing ThinkPad or ThinkStation or as mentioned, part of a 2-node cluster. When connected to another PGX, you can attack larger AI models of up to 405 billion parameters. Pre-installed and configured on the system is the NVIDIA DGX Operating System, which is a modified Ubuntu Linux operating system. Also, pre-installed is the NVIDIA AI software stack, which includes PyTorch, Jupyter Notebooks, and CUDA 13. With this little unit, developers can prototype and fine-tune models locally, keeping IP on premise, and develop and deploy Agentic AI systems, without even reaching for the cloud.
The basic specs for this system include a 4TB NVMe M.2 drive, 128GB of unified memory, and a system-on-a-chip NVIDIA GB10 Grace Blackwell superchip with a 20 core ARM CPU. NVIDIA’s Superchip offers the same characteristics as NVIDIA’s enterprise-class Datacenter Grace CPU. With ARM architecture, it includes 10x Cortex-X925 cores and 10x Cortex-A725 cores. Then you have the integrated NVIDIA Blackwell Architecture GPU with 6,144 CUDA Cores, 5th generation tensor cores, 4th generation RT Cores, plus a dedicated Video encoder and decoder. When combined, it offers 1,000 TOPS or trillions of Operations Per Second.
This chip will deliver up to a petabyte of AI performance at FP4 precision. FP4 is an ultra-low precision numerical, floating-point format that uses significantly less memory, bandwidth, and energy. FP8 is better for more precision but cannot match the Speed of FP4. It also has NVIDIA’s NVLink-C2C technology for a connection between the CPU and GPU processors that is up to 5 times faster than PCIe 5.0 bandwidth.
There’s no setup required, it’s plug-and-play for developing and experimenting with AI right out of the box. Businesses, students, and developers now have AI at their fingertips at a reasonable price point. Is this a workstation? No, not really. You can game on it with some slight modifications, but it can’t run Windows applications, at least not natively. It’s a testing ground for AI development and experimentation. You may want to brush up on some basic Ubuntu Linux commands for this system. You can connect a monitor, mouse and keyboard using the ports in the back or connect it to your workstation or laptop either directly or through a switch. Included with the system are 2x USB-C to USB-A dongles. You can also connect it to a Bluetooth device.
At this point most of us are familiar with using AI tools as they have infiltrated every aspect of our lives. If you want to look behind the curtain and try creating your own AI environments, or train your own AI for specific tasks, then this little unit is definitely for you. Simplicity, Scalability, and empowering people to experiment with AI without the hassle of setting your system up from scratch. PyTorch and Jupyter are already pre-installed, plus the NVIDIA application library. You don’t have to wonder if updates will be compatible with your PGX, because updates to the AI software library and other software assets will be compatible. That right there, removes a layer of complexity, and frustration at the same time.
Several other companies have also released similar small footprint AI development systems featuring the NVIDIA GB10 Grace Blackwell Superchip, paired with up to 128GB of unified memory. NVIDIA would be one of them with their DGX Spark platform. On a typical workstation you would have DRAM controlled by the CPU, VRAM controlled by the GPU, and then additional attached components that might use those assets after commands are filtered through the CPU.
With unified memory, the CPU and GPU draw from a coherent memory pool, independently. Unified memory reduces the number of commands required to access that memory, as well as reducing latency. It also reduces potential errors from stale data. It may be only milliseconds to your system but it does add up. This enables all devices that might be working on the data to be using the most current data. Crucial for heterogeneous, high-performance computing, like in AI.
The PGX can also draw on Lenovo’s full-stack, validated Hybrid AI platforms when developers want more powerful appliances to train their AI models or take it to the next level. Moving AI models from the PGX to any accelerated cloud or data center infrastructure requires virtually no code changes. When developers are ready to supercharge their AI applications, Lenovo offers additional tools validated by NVIDIA. They cover the gamut from entry-level, smaller single node deployments with Hybrid AI 221 platforms, to developing large-scale inference use cases with Hybrid AI 289.
Hybrid AI 221 refers to 2x processors paired with 2x GPUs and 1x high-performance network adapter. Get it? 221. Hybrid AI 289 also incorporates a dual processor server, plus 8x GPUs, and 9x high-performance network adapters. Lenovo offers several more powerful backend server options compatible with the PGX. Like the Lenovo ThinkSystem SC777 V4, designed specifically for AI. It’s housed in the ThinkSystem N1380 enclosure, which can support up to 8x water-cooled SC777 V4 blades. Each blade has 2x NVIDIA Grace processors and memory, working with NVIDIA’s GB200 or G200 GPUs. It makes it even easier to bring that initial development you started on the PGX to fruition, in a timelier manner.
The front panel is generously perforated with little hexagonal holes for air circulation. A vent on the bottom of the chassis provides for additional air flow. Around back, there are a number of ports including; 1x USB-C for power, then 3x USB-C ports offering 20Gb/s transfer speeds with DisplayPort 2.1 support, then a single HDMI 2.1a port with multichannel audio. An RJ-45 port is included for basic network communications at 10GbE. Lastly dual Quad Small Form-factor ports for NVIDIA ConnectX-7. Those QSFP ports allow you to connect to other ThinkStation PGX units or connect to a network switch at up to 200Gb/s. This system also supports WiFi 7 and Bluetooth 5.3.
With its small size, just saying chassis seems a little overkill. This unit is only 1.13 liters in volume, which is about the same as having two water bottles out of a 24-pack placed side-by-side. Those are only 500ml each but the two together are a full liter. There’s really not much to do regarding upgrades either. It’s pre-configured and pre-loaded with software. As mentioned, the only thing you might want to do is brush up on your basic Linux Ubuntu commands. Once you plug it in and turn it on, you can either press the Windows logo on the keyboard, which won’t launch Windows but will open the activities menu on the upper left of the display. Honestly, we’re not familiar with Ubuntu Linux nor building AI models. The interface is nice and well organized but this would take a little time for us to dial in.
If you’re a data scientist, a student in computer science, or just a hobbyist who likes to tinker, then this Lenovo ThinkStation PGX supercomputer, mini titan is a worthy challenge that can yield some impressive results. It can give you a foothold on AI and potentially enable you to develop the foundations of something like Skynet, and help usher in our AI overlords faster. If you would like more information on this platform, or any other system, contact IT Creations! We have tons of inventory in stock for servers, workstations, components, and more.
