Supermicro has begun selling the Super AI Station, a workstation built around Nvidia's GB300 Grace Blackwell Ultra Desktop Superchip and aimed at teams that want data-center-class AI development hardware beside a desk instead of inside a full rack.

The configured Gold Series system, listed as ARS-511GD-NB-LCC-01-G2, appears in Supermicro's US eStore with a starting price of $92,887.47. That makes this much more than an "AI PC" headline. It is workstation-shaped infrastructure for organizations that want local control over model testing, fine-tuning, inference, data science, and agent workloads.

What Supermicro is selling

The Super AI Station is Supermicro's implementation of Nvidia's DGX Station-class GB300 platform. The system packages a 72-core Nvidia Grace CPU, an onboard Blackwell Ultra B300 GPU, and Nvidia ConnectX-8 networking into a liquid-cooled tower that can also be rack-mounted as a 5U system.

The headline numbers are large for a workstation:

  • 748GB of coherent memory, split between 496GB LPDDR5X system memory and 252GB HBM3e GPU memory
  • up to 20 PFLOPS of FP4 AI performance
  • 2x QSFP 400GbE ports through Nvidia ConnectX-8
  • Ubuntu with Nvidia AI Developer Tools preinstalled
  • closed-loop liquid cooling
  • optional RTX Pro Blackwell GPU expansion for visualization, rendering, simulation, or physical AI work

Supermicro says the system is designed for research labs, national labs, AI development, AI inference and training, data science, and agentic AI. The company is also positioning it as a local staging system for workloads that later scale into larger Nvidia-based AI factory deployments.

Why it matters

The important shift is not just performance. It is where the performance lives.

Most serious model work has moved toward rented cloud GPUs or managed infrastructure because large models quickly outgrow normal workstations. Supermicro is now selling a machine that tries to pull part of that work back on-prem, with enough memory to run or test much larger models locally and enough networking to plug into serious storage or multi-system environments.

That matters for teams with sensitive data, high experiment volume, or unpredictable cloud costs. A local workstation does not remove the need for clusters, but it can change the development loop. Engineers and researchers can test agents, fine-tune models, benchmark workflows, and validate data pipelines closer to where proprietary data already lives.

The timing also fits a broader hardware trend we have been tracking. Nvidia's smaller local-AI push shows up in creator-class systems like ASUS ProArt RTX Spark laptops, while larger infrastructure fights are moving down into chips, memory, networking, and custom accelerators, as seen in OpenAI's Broadcom inference chip project. Supermicro's GB300 workstation sits between those worlds: far above a laptop, but more approachable than a full data-center rack.

The price signals the buyer

At nearly $93,000 before any buyer-specific configuration or volume terms, the Super AI Station is not competing with normal developer desktops. It is aimed at AI labs, enterprise platform teams, higher education, deep-tech startups, and organizations where local infrastructure can be justified by speed, privacy, utilization, or avoided cloud friction.

That price also makes the buying decision clearer. A team does not choose this because it wants a faster PC. It chooses this if local access to hundreds of gigabytes of coherent memory, Blackwell Ultra acceleration, and Nvidia's enterprise AI software stack changes what it can safely build or test.

What to watch

The technical promise is straightforward, but the operational questions are still important:

  • how much useful real-world performance developers get from FP4 on their own workloads
  • whether 748GB of coherent memory is enough for the target models teams actually want to run locally
  • how noisy, serviceable, and office-friendly the liquid-cooled system is in practice
  • how quickly Windows support and Nvidia's broader agent tooling mature around the platform
  • whether buyers compare it against cloud GPU credits, a small rack, or multiple lower-cost workstations

For now, the product is a strong signal that "local AI" is splitting into two categories. One is consumer and creator hardware that can run smaller models conveniently. The other is deskside infrastructure for professional teams that want frontier-adjacent development capacity without moving every experiment into the cloud.

Our take

Supermicro's Super AI Station is expensive, specialized, and probably exactly priced for its real audience. The interesting part is that the GB300 platform is now showing up as something a team can buy as a workstation, not only as a data-center building block.

If the system delivers stable local performance for large-model development and agent testing, it could become a useful bridge between individual AI workstations and full AI factory deployments. The broader lesson is simple: AI infrastructure is moving closer to the desk again, but only for teams whose workloads are already large enough to justify data-center hardware in workstation clothing.