Lenovo Launches Compact AI Workstation with Nvidia GB10 and 128 GB Memory, Ready for Heavy-Duty Models

Lenovo has introduced a new compact AI workstation called the ThinkStation PGX, designed to offer high-performance computing for AI developers right out of the box. This mini supercomputer is powered by an Nvidia GB10 Grace Hopper Superchip, featuring 128 GB of coherent unified system memory, delivering up to 1 PetaFLOP (or 1000 TOPS) of performance. This capability supports working on models with up to 200 billion parameters. For those needing even more power, Lenovo allows users to connect two PGX units, raising the maximum parameter count to 405 billion. The ThinkStation PGX runs on the Nvidia DGX operating system and includes the standard Nvidia software stack for AI development. To further enhance its usability, Lenovo has integrated popular tools such as PyTorch and Jupyter, enabling developers to work locally without needing on-premises clusters or cloud computing resources. The company plans to launch the PGX in the third quarter of 2025, though pricing details have not been disclosed yet. Lenovo’s PGX appears to be a response to Nvidia’s Project Digits, a desktop AI supercomputer unveiled at CES 2025 and priced at around $3,000 per unit. It’s part of a growing trend of third-party AI mini supercomputers entering the market. Asus launched the Ascent GX10 in March 2025 at GTC 2025, and MSI has hinted at its own model, set to debut at Computex 2025. These devices have generated mixed reactions from the tech community, with some enthusiasts excited by the possibilities, and others more skeptical. One notable critic is Tiny Corp, the startup behind the TinyBox AI accelerator. Tiny Corp suggests that users might find it more practical to invest in a high-end gaming PC instead. According to their analysis, the 1 PFLOP performance of the AI mini PC is measured at FP4 precision, which is considered practically unusable for many AI tasks. At more commonly used FP8 precision, the performance drops to 500 TFLOPS. In contrast, the TinyBox Green, equipped with four RTX 5090 GPUs, boasts 1,492 FP16 TFLOPS, equivalent to 2,992 FP8 TFLOPS, or about 3 PFLOPs at FP8. This makes it roughly six times more powerful than Nvidia’s Project Digits systems, including the PGX. However, the TinyBox Green is significantly pricier, starting at $29,000, compared to the estimated $3,000 for the PGX. Despite the performance discrepancy, the ThinkStation PGX offers a compelling value proposition for developers who need a user-friendly, compact solution for AI tasks without breaking the bank. Its integration with essential AI development tools and its affordable price point make it an attractive option for individuals and small teams looking to enter the AI space. To stay informed about the latest developments in AI hardware and technology, follow Tom's Hardware on Google News. Make sure to click the "Follow" button to receive up-to-date news, analysis, and reviews directly in your feed.