Lenovo launches the ThinkStation PGX, a compact desktop workstation built around Nvidia’s Grace Blackwell superchip and aimed at AI developers.
Lenovo launches the ThinkStation PGX, an eccentric workstation. The device is not equipped with an x86 chip from Intel or AMD, but features the ARM-based Nvidia GB10 Grace Blackwell Superchip. This chip has 128 GB of shared system memory and delivers up to one petaFlops of low-precision AI performance.
The desktop doesn’t run on Windows by default, but Lenovo provides it with the Nvidia DGX OS and a complete AI software stack, including PyTorch and Jupyter. This gives developers immediate access to the most important tools to start prototyping, fine-tuning, and inferencing AI applications.
Once testing is complete, Nvidia’s software stack makes it possible to transfer developed models to data center or cloud environments with minimal code adjustments.
Stronger Together
The Grace Blackwell superchip can handle generative AI models up to 200 billion parameters. For larger models up to 405 billion parameters, two devices can be linked together.
The device fills an interesting niche. Hardware-wise, the setup of the ThinkStation PGX with Nvidia GB10 is reminiscent of systems built around the new AMD Max Pro chips. These also combine GPU computing power with large amounts of shared memory. AMD has the advantage of x86 compatibility, but for developers, the integration of Nvidia’s AI stack and the cloud compatibility of the ThinkStation PGX may be a bigger asset.
With the ThinkStation PGX, Lenovo targets developers looking for a powerful but economical alternative to offload their cloud or cluster resources. The desktop offers an environment where AI applications can first be developed locally before moving to large-scale training or production environments. The ThinkStation PGX will be available from the third quarter of 2025. Lenovo has not yet shared a recommended price.