400Gbps Networking on a Desk: Who Actually Needs This Monster?

Almost nobody needs 400Gbps networking on a desk, and the people who do already know it. The desk-sized AI systems that brought these ports into offices use them for one job: joining two or more AI boxes into a cluster, or pulling large datasets off fast storage. For email, file shares and video calls, a 1 Gb or 10 Gb office network is still the right answer.

The hardware is real, though. The NVIDIA DGX Spark and ASUS Ascent GX10 carry an NVIDIA ConnectX-7 network card, and the NVIDIA DGX Station GB300 carries a ConnectX-8 SuperNIC. Add a NVIDIA QSFP112 400G DAC Cable, a MikroTik CRS804 DDQ or an NVIDIA Spectrum-2 SN3700 switch and you have data-centre networking under a desk. Here is who should care.

What speeds do desk AI systems actually have?

The "400G" in product names describes cables and switch ports. The systems themselves run at the speeds NVIDIA publishes, which are not all 400 Gb/s.

So a GB10 desk system talks at 200 Gb/s per port. NVIDIA's clustering guide says plainly that faster cables offer no benefit on those ports. The 400 Gb/s per port, and 800 Gb/s in total, belongs to the DGX Station class, such as the Supermicro Super AI Station ARS-511GD-NB-LCC, which is price on request and quoted per configuration.

Who actually needs 200G or 400G on a desk?

Three groups, in our view:

• Teams running models too big for one box. NVIDIA rates a 128 GB DGX Spark for models of up to 200 billion parameters and two linked units for up to 400 billion. That second figure only exists because of the fast link between them.
• Developers rehearsing for the data centre. Multi-node training and inference code behaves differently once work is split across machines. A pair of GB10 boxes is a cheap place to learn that before renting or buying a cluster.
• People moving very large datasets. Checkpoints, model weights and training data measured in hundreds of gigabytes move faster over a 200G link to all-flash storage than over 10 GbE, if the storage can keep up.

Our guide Do You Need an AI Supercomputer? GB10 vs PC vs Cloud covers whether you need the box at all, before the network question even arises.

Who does not need it?

Almost every other office user. A single GB10 system used for local chat, coding assistants or document search spends its time on compute and memory, not on the network. Its 10 GbE port and Wi-Fi 7 cover access from laptops. If staff complain that the office network is slow, the cause is usually an older switch, a busy firewall or Wi-Fi coverage, which we cover in Your Internet Is 10Gbps. So Why Is Your Office Network Still Slow? and 10GbE Networking in a Small Office: Brilliant or Completely Unnecessary?. Fix those first; a 200G link between two AI boxes does nothing for the rest of the floor.

Two AI boxes: is a cable enough?

Yes. NVIDIA states that two DGX Spark systems connect with a direct QSFP cable and no switch. That is the cheapest high-speed setup there is: two units and one short cable. The store's NVIDIA QSFP112 400G DAC Cable is a 40 cm passive copper cable built for that back-to-back link. One caution: NVIDIA's clustering page lists approved cables by manufacturer part number, so check the cable against NVIDIA's current list before you buy, from us or anyone else. We confirm compatibility when quoting. Our walkthrough How to Cluster Two ASUS Ascent GX10 Units covers the software side.

Three or four boxes: when do you need a switch?

NVIDIA's Cluster Assistant supports up to three DGX Spark systems wired directly to each other and up to four through a switch, and the product page says four units require an external 200 GbE switch. This is where the "monster" switches come in.

The MikroTik CRS804 DDQ is the small-cluster option. MikroTik lists four 400G QSFP56-DD ports, each built from eight lanes, which can break out to two 200G links per port, plus two 10G RJ-45 ports. It draws up to 123 W with optics fitted, runs RouterOS v7, has dual hot-swap power supplies, and MikroTik positions it for small to medium AI and GPU clusters. It is half-width, so it suits a short rack or comms cabinet rather than a desk.

The NVIDIA Spectrum-2 SN3700 is the full data-centre answer: 32 ports at up to 200GbE each and 6.4 Tb/s of throughput. NVIDIA's hardware manual lists 250 W, a 14 kg 1U chassis and an operating range of 0 to 40 °C. That belongs in a ventilated rack, ideally on a UPS, not next to someone's monitor. If your rack is already crowded, read Your Network Cabinet Is a Mess. And That Could Cost You. before adding to it.

What about the DGX Station and 800 Gb/s?

The DGX Station GB300 is a different class of machine. NVIDIA lists 748 GB of coherent memory, a ConnectX-8 SuperNIC with up to 800 Gb/s across two 400 Gb/s QSFP112 ports, and support for linking up to two systems. It also lists 1,600 W of total system power and a 20A circuit. The networking is the least of the installation questions: power, heat and noise come first. See our NVIDIA DGX Station GB300 buyer's guide and the DGX Station GB300 office installation checklist.

The case for and against

For: the network is the cheap part. A direct cable costs a small fraction of the second AI box, and once you own two boxes, leaving them unconnected wastes the ability to run larger models. A small switch such as the CRS804 DDQ costs less than one GB10 system.

Against: clustering adds work. Someone has to set up the link, keep software versions aligned across nodes and troubleshoot when a job runs slower split across machines. Many teams would be better served by one bigger system, a cloud burst for the occasional large job, or simply a smaller model.

Buy the second AI box for a reason, and the 200G cable comes with that reason. Never buy the switch first and look for a workload later.

Our verdict: 400G-class networking on a desk is worth it for teams that have outgrown one GB10 system and want to keep work on premises. For everyone else, spend the money on a sound 1G or 10G office network, a fanless switch that suits the room (see Is a Fanless Network Switch Worth the Heat Risk?) and good storage.

Buying desk AI networking in Singapore

SourceIT supplies NVIDIA, MikroTik and Supermicro with local warranty, and ASUS Ascent GX10 systems; confirm the warranty term for your SKU when requesting a quote. Prices above are as listed on 23 Oct 2026. Browse the NVIDIA DGX Spark collection for GB10 systems and cables, and the network switch and wireless AP collection for switches. For two or more systems, or a cluster with a switch and storage, ask about volume discounts and special pricing. Use Request for Quotation on the product page, call (65) 6978 3502 or email Sales@sourceit.com.sg. Government buyers can purchase through GeBIZ.

Frequently asked questions

Does the DGX Spark have 400Gbps networking? No. NVIDIA lists a ConnectX-7 NIC at 200 Gbps, with two QSFP ports of up to 200 Gb/s each and a 10 GbE port. The 400G label appears on compatible cables and switch ports.

How do I connect two DGX Spark or ASUS GX10 units? With a direct QSFP cable between their high-speed ports; NVIDIA says no switch is needed for two units. Check the cable against NVIDIA's approved list first.

When do I need a switch for GB10 systems? NVIDIA's Cluster Assistant supports up to three systems wired directly and up to four through a switch. The product page says four units require an external 200 GbE switch.

Will a faster cable make my DGX Spark link faster? No. NVIDIA's clustering guide says the ports top out at 200 Gb/s and faster cables offer no benefit.

How fast is the DGX Station GB300 network? NVIDIA lists a ConnectX-8 SuperNIC with up to 800 Gb/s, using two QSFP112 ports at 400 Gb/s each, and support for linking up to two systems.

Can a data-centre switch sit in an ordinary office? It can, but it should not sit on a desk. NVIDIA lists the SN3700 at 250 W with a 0 to 40 °C operating range in a 1U rack chassis, so plan a ventilated rack and UPS.