Use ThinkStation PGX as a finished internal AI service? WZ-IT combines GB10 hardware with Open WebUI, an agreed model, hardening, networking, and operations. Explore the AI Cube
Lenovo ThinkStation PGX is Lenovo's compact AI workstation based on NVIDIA's GB10 Grace Blackwell Superchip. It offers 128 GB of unified memory, DGX OS, and ConnectX-7. Lenovo primarily targets AI development and local model work. An internal platform for employees additionally requires a usable interface, controlled access, and reliable operations.
Technical context
Lenovo documents up to 4 TB of self-encrypting NVMe storage, 10 GbE, Wi-Fi 7, and a 200 Gbps ConnectX-7 link. A 20-core Arm processor and Blackwell GPU share 128 GB of coherent memory.
PGX is not a conventional x86 Windows workstation. Containers and dependencies must support ARM64, so existing internal software should be checked before migration.
Select a model, not a specification limit
Lenovo states models up to 200 billion parameters on one system and 405 billion across two connected PGX units. This does not describe chat latency or concurrent user capacity.
Selection depends on model architecture, quantisation, task quality, context length, concurrent requests, target response speed, and headroom for KV cache and platform services.
Configure it as a business platform
Production setup covers system updates, named administrators, network and firewall rules, local model runtime, Open WebUI, persistent data, backup, and monitoring. Model access and knowledge spaces are separated by role.
Employees then work through a browser rather than developer tooling. They can use approved models and maintain personal or shared knowledge. Automated Nextcloud, SharePoint, or DMS sources are implemented as a separate integration scope.
When ThinkStation PGX fits
PGX is a good fit where Lenovo already supplies workstations and procurement wants to place a compact Linux AI platform in that vendor context. Technically it is close to other GB10 systems; practical differences mainly concern SSD configuration, availability, warranty, and supply chain.
The decision should define whether the device is a developer workstation or a central service. For the latter, fixed network paths, independent backup, and clear operational ownership matter more than local display connections. Existing x86 applications should be checked for ARM64 availability before deployment.
WZ-IT AI Cube and clustering
The AI Cube combines hardware, Open WebUI, runtime, an agreed model, hardening, functional testing, initial setup, and support. The AI Cube costs from €6,490 excl. VAT one-time; patching, backup, monitoring and support are covered by AI Cube Care from €349.90 excl. VAT per month; the first twelve months run with the device, after that Care is cancellable monthly.
A second device can serve more independent requests or distribute a larger model. The approaches need different configurations. See connecting AI Cubes with ConnectX-7.
