NVIDIA DGX Station vs Traditional GPU Servers: Is the Data Centre Moving to Your Desk?

Partly. The NVIDIA DGX Station GB300 moves one data-centre-class GPU to the desk: a single Blackwell Ultra GPU with 252 GB of HBM3e, on a 1,600 W office power budget, with no rack, no raised floor and no separate cooling plant. A traditional 8-GPU server is a different class of machine. NVIDIA's own DGX B200, for example, has eight Blackwell GPUs with 1,440 GB of HBM3e and draws up to about 14.3 kW from a 10U rack chassis. For a team's development, private inference and fine-tuning, a DGX Station can replace time on a shared server. For production-scale serving, large training runs and anything that needs several GPUs working as one, the server still belongs in the data centre.

How does a DGX Station compare with an 8-GPU server?

Source: NVIDIA DGX Station datasheet (Aug 2026) and NVIDIA DGX B200 product page (accessed 5 Nov 2026). DGX B200 is used as a reference design; HGX-based servers from other vendors differ. Peak figures are NVIDIA's ratings with sparsity, not measured throughput.

On paper, the server has about 5.7 times the GPU memory and 7.2 times the peak FP4 compute, for about 8.9 times the maximum power. The bigger difference is that its eight GPUs share NVLink, so one very large model or training job can span them as if they were one. Two DGX Stations connected over ConnectX-8 Ethernet give "over 1.4 TB of coherent memory" in NVIDIA's words, but there is no NVLink between units, so multi-unit work runs over the network instead. See DGX Station's networking for what that link can and cannot do.

What does each one need from the building?

This is where the deskside model saves the most. A DGX Station plugs into an office circuit, cools itself with its internal liquid loop and stays reasonably quiet (ServeTheHome measured an MSI unit at 37–49 dBA). Our power, heat and noise guide covers what to check with your electrician and facilities team.

An 8-GPU server at around 14 kW needs a rack with high-power distribution, cooling sized to remove that heat, UPS capacity, and usually a data centre or colocation contract. Electricity alone shows the scale. At SP Group's Q4 2026 tariff of 31.16 cents per kWh including GST, running flat out all year would cost about S$39,000 for a 14.3 kW server and about S$4,370 for a 1.6 kW DGX Station. Both are upper bounds, since neither runs at maximum all the time, and the server figure excludes the cooling energy a data centre adds.

The question isn't whether a desk can match a data centre. It's how much of your AI work actually needs one.

Which workloads still belong on a rack server?

• Production inference at high volume: eight GPUs serve far more concurrent users than one, and servers fit into existing data centre redundancy, monitoring and change control.
• Training or fine-tuning very large models: jobs that need more than 252 GB of fast GPU memory, or that scale across GPUs, benefit from NVLink.
• The largest models at higher precision: a model that needs 1,000 GB at 8-bit fits within a 1,440 GB server's GPU memory but not a DGX Station's 748 GB.
• Workloads needing high availability: data centre servers can be clustered with redundant power and networking. A single deskside system is a single point of failure; see Your DGX Station Could Be a Single Point of Failure.

Which workloads move comfortably to the desk?

• Model development and evaluation: data scientists testing models up to the size that fits in 252 GB of HBM3e, or larger at reduced speed using the LPDDR5X tier.
• Departmental private inference: StorageReview measured gpt-oss-120b at 9,280 output tokens per second across 128 concurrent requests on an ASUS ExpertCenter Pro ET900N G3, enough for a sizeable team.
• Fine-tuning and adaptation: LoRA-style tuning of mid-size open-weight models on company data that should not leave the building.
• Sensitive or regulated data: work that cannot go to a cloud region but doesn't justify a data centre build.

Our post Can a Mini AI Supercomputer Replace an Entire Server? looks at the same question for smaller GB10 systems, and Can Eight NVIDIA GB10 Systems Compete With Enterprise AI Infrastructure? covers clustering them.

What about renting server GPUs in the cloud instead?

For bursty or occasional work, renting can beat both. Single Blackwell-class GPUs are available by the hour: RunPod listed one B300 at US$8.99 per hour and Nebius at US$9.50 per hour (as listed on 5 Nov 2026). That flexibility costs more if you use it constantly. One B300 rented around the clock at US$8.99 comes to about US$78,750 a year, before storage and data transfer. The only like-for-like Singapore-region price we found, Azure's H200 instance, was about 30% higher than the same instance in US East. Buy DGX Station or Keep Paying for Cloud GPUs? runs the full three-year comparison.

So is the data centre moving to your desk?

The development and departmental layer is. Work that used to queue for time on a shared GPU server, or go to a cloud API, can now run on a single deskside system with no facility changes beyond a suitable power point. The production and large-scale training layer is not moving: eight NVLink-connected GPUs, data centre cooling and redundancy are still the right tools there. Many organisations will end up with both, using DGX Stations for teams and rack servers or cloud for scale. For how much token output one deskside unit delivers, see How Many AI Tokens Can DGX Station Actually Generate?

Buying in Singapore

For deskside AI, SourceIT lists five DGX Station GB300 systems, all price on request: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. For rack-mounted GPU servers, we list the Lenovo ThinkSystem SR680a V3 (B200) and HPE ProLiant DL380a Gen11, also price on request. See the AI Workstations and AI Servers collection for more. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

Is a DGX Station as powerful as a GPU server? No. It has one Blackwell Ultra GPU with 252 GB of HBM3e. NVIDIA's DGX B200 reference server has eight GPUs with 1,440 GB of HBM3e connected by NVLink, and about 7.2 times the peak FP4 compute.

How much power does each draw? A DGX Station's system power is 1,600 W. NVIDIA lists the DGX B200 at about 14.3 kW maximum. At SP Group's Q4 2026 tariff with GST, that is about S$4,370 versus S$39,000 a year if each ran at maximum around the clock.

Can two DGX Stations act like one larger server? Partly. Two units linked over ConnectX-8 give over 1.4 TB of coherent memory, but they communicate over 400 Gb/s Ethernet rather than NVLink, so multi-GPU jobs are slower than inside an 8-GPU server.

Does a DGX Station need a server room? No. It is a liquid-cooled tower designed for an office, though it needs a suitable power circuit, ventilation space and an equipment plan. Some OEMs offer a rack-mount option.

When should I choose a rack server instead? When you need production-scale inference for many users, training that spans several GPUs, models larger than a single system can hold at your preferred precision, or data centre redundancy.