Dell Technologies began selling the Dell Pro Max with GB300, a tower workstation built around the NVIDIA Grace Blackwell Ultra GB300 Superchip, in Japan on August 20. It brings a data center class configuration to the side of a desk, with 748GB of coherent memory and 20,000 TFLOPS of FP4 compute. The price on Dell's official online store is 23,531,150 yen.
Data center silicon moves to the desktop
Until now, Grace Blackwell Ultra was a component for teams building large AI infrastructure. The Dell Pro Max with GB300 packs that GB300 Superchip into a single tower chassis so research labs and enterprise AI teams can run it under their own control. Dell announced shipments of a GB300 desktop in the United States back in March, and the Japanese order window follows from that. The model number is FCT6263.
Memory capacity is the headline of the configuration. The system pairs 252GB of HBM3e on the GPU side with 496GB of LPDDR5X on the system side, for 748GB of coherent memory in total. Dell says that capacity is enough to work with models of up to 1 trillion parameters. Storage takes four M.2 NVMe 2242 Gen4/Gen5 SSDs for up to 16TB. Display output is handled by an NVIDIA RTX PRO 2000 Blackwell with 16GB of GDDR7 on a PCIe card.
Where that much memory pays off
When you work with large models, memory usually becomes the ceiling before compute does. If parameters and intermediate data will not fit in GPU memory, you end up splitting the model or spreading it across multiple machines, and every iteration gets slower. A 748GB pool removes much of that partitioning work.
Dell points to prototyping, fine-tuning, local inference, data science, and edge application development as the target workloads. The appeal is being able to train on data that is hard to send outside the building, such as financial documents or medical imaging, on a single machine at hand. For teams that want to avoid cloud metering, latency, or compliance constraints, that character alone can justify the purchase.
On the networking side, an NVIDIA ConnectX-8 SuperNIC provides up to 800 Gb/s of bandwidth. Linking two Dell Pro Max with GB300 units directly doubles the available compute and lets teams scale up to larger models in steps.
Cooling designed for continuous operation
A machine meant to sit beside a desk has to stay cool and quiet. Dell fits this system with its own MaxCool technology, which combines a smart cold plate design, dual heat exchangers, and next-generation thermal interface materials. Dell states that internal testing as of February 2026 showed more efficient heat removal than a baseline cooling solution.
The redundant pump system is easy to overlook but matters in practice. Operation continues if one loop fails, so a multi-day training job does not get rolled back by a cooling fault. Sustaining peak performance without thermal throttling also makes run times easier to predict.
Another notable feature is the MIG-based personal cloud capability. A single Dell Pro Max with GB300 can be divided among up to 7 users, each with isolated memory. The design intent is clearly a shared team resource rather than a personal machine.
Software ships as Ubuntu with a preconfigured AI stack that includes the CUDA-X libraries. The system also connects to Dell AI Factory with NVIDIA, giving a path to carry code, models, and workflows built locally into on-premises or cloud production environments. Build locally, then deploy anywhere is the intended flow.
Reading the price
At 23,531,150 yen, this is not a machine an individual buys on a whim. Set against the cost of renting comparable GPU resources in the cloud for years, or the projects an organization has abandoned because the data could not leave the premises, some buyers will find it lands within range. Dell also offers the smaller Dell Pro Max with GB10 for deskside AI development, so there is a choice by scale.
Summary
The Dell Pro Max with GB300 puts an NVIDIA GB300 Superchip and 748GB of coherent memory in a tower chassis so trillion-parameter models can be handled locally. MaxCool cooling and a redundant pump system support long training runs, and MIG allows sharing among up to 7 users. Orders opened in Japan on August 20 at 23,531,150 yen on Dell's official store. For AI teams working with data that cannot be handed to the cloud, there is now one more practical option.
