Big AI Performance in a Tiny Box: Lenovo’s Latest Innovation
Lenovo has announced the ThinkStation PGX , a compact “personal AI developer workstation” designed to function as an AI mini supercomputer right out of the box. Powered by the Nvidia GB10 Grace Blackwell Superchip and equipped with 128GB of coherent unified system memory , the PGX delivers up to 1 PetaFLOP (or 1000 TOPS) of computing power — enough to handle AI models with up to 200 billion parameters .
For users needing even more performance, Lenovo allows linking two PGX units together , boosting capacity to support models with up to 405 billion parameters .
The system runs on the Nvidia DGX operating system and comes preloaded with the full Nvidia AI software stack , along with popular tools like PyTorch and Jupyter . This setup enables developers to build and train large AI models locally, without depending on cloud services or large-scale data centers.
Lenovo plans to launch the ThinkStation PGX in Q3 2025 , though specific pricing details have not been released yet.
The ThinkStation PGX appears to be Lenovo’s version of Nvidia’s Project Digits , a compact AI workstation unveiled at CES 2025 with an estimated price tag of around $3,000 — a figure we expect the PGX to closely follow.
It’s not the first third-party system based on this concept. ASUS introduced its Ascent GX10 at GTC 2025 in March 2025, and MSI has teased its own model set to debut at Computex 2025 . While some enthusiasts have welcomed these compact AI workstations, others remain skeptical. In fact, Tiny Corp , the company behind the TinyBox AI accelerator , has openly criticized the idea, suggesting that users would be better off simply buying a high-end gaming PC rather than spending thousands on a dedicated AI desktop.
One major concern is performance precision. The PGX and other Project Digits systems are rated at 1 PetaFLOP (PFLOP) , but that number is based on FP4 precision , which is largely impractical for real-world AI workloads. In more usable FP8 precision , these systems deliver only about 500 TeraFLOPS .
In contrast, the TinyBox Green , now powered by four RTX 5090 GPUs , offers 1,492 FP16 TeraFLOPS , which translates to 2,992 TeraFLOPS at FP8 — roughly 3 PetaFLOPS , or six times the performance of Project Digits systems. However, this extreme level of power comes at a much higher cost: the TinyBox Green starts at $29,000 , significantly more than the ~$3,000 price point of the compact AI workstations.
This comparison highlights a growing debate in the AI community: whether compact, consumer-focused AI PCs offer enough value compared to traditional high-performance setups.
