Four GB10 Supercomputers on One Desk: Brilliant AI Cluster or Expensive Experiment?

Both, depending on why you build it. Four NVIDIA GB10 systems joined through a 200 GbE switch give 512 GB of pooled memory, enough for models NVIDIA rates up to 700 billion parameters, for well under the cost of a DGX Station GB300. For a research group that needs frontier-class open models on premises, or a team learning multi-node AI before it rents data-centre time, that is a brilliant tool. For a business that simply wants a fast assistant for its staff, it is an expensive experiment: four separate boxes on one desk, a switch, version-matching across all of them, and speeds that suit a few developers rather than a company. Four ASUS Ascent GX10 1 TB units cost S$32,400 at SourceIT before the switch and cables (listed 28 Oct 2026).

What does a four-unit cluster cost and need?

Requirements from NVIDIA's Cluster Assistant documentation and DGX Spark playbooks. Power figures from ServeTheHome and Tom's Hardware reviews of DGX Spark. Prices listed 28 Oct 2026.

What does it actually deliver?

NVIDIA and third-party results, different models and settings. We have not run a four-unit cluster ourselves.

Two lessons stand out. First, four units do make one user faster on models split with tensor parallelism, about three times on NVIDIA's 27B example. Second, they serve a crowd worse than four independent boxes, because every token now crosses the network. Build the cluster for the models that need 512 GB, not for headcount.

Why is it "brilliant"?

• Capacity on premises. 512 GB holds 400B to 700B-class open models at 4-bit, which no single desktop GB10 can.
• The cheapest multi-node lab. Teams preparing to train or serve across data-centre GPUs can learn NCCL, tensor parallelism and cluster debugging on hardware that fits a desk.
• Flexibility. On Monday it is a cluster; on Tuesday it can be four separate developer boxes. A single big system cannot be split up that way, short of using MIG on a GB300.
• Scaling continues. NVIDIA says up to eight units are possible in manual configuration; see Can Eight NVIDIA GB10 Systems Compete With Enterprise AI Infrastructure?

A four-box cluster is a laboratory. Run it like one, and it is excellent value. Expect it to behave like a server, and it will disappoint.

Why is it an "expensive experiment"?

• Four machines to manage. Four operating systems to update in step, four sets of logs, no BMC for remote management.
• Networking you must get right. NVIDIA names no specific switch; community users report needing breakout cables and fixed 200G port speeds on the MikroTik CRS804, and say it has been backordered.
• Fragility. Community build reports include a ConnectX-7 link stuck at a low speed after a cable swap until reboot, and units freezing under sustained load. These are single reports, but they show the kind of issues a cluster adds.
• Speeds suit a team, not a company. 55.8 tokens per second for one user on a frontier-class model is impressive on a desk; 126 in total across five users is not a company-wide service.
• Not one coherent memory. Four 273 GB/s pools joined by 200 Gb/s links are not the same as one large memory space.

Four GB10s or one GB300?

A DGX Station GB300 has 748 GB of coherent memory, 252 GB of it HBM3e at 7.1 TB/s, up to 20 PFLOPS of FP4 compute, a BMC and MIG for up to seven users. It is one managed system rather than four, and it serves many users far better, but it is price on request, needs a 1,600 W circuit and costs several times as much as four GB10s. Choose four GB10s for a research group that wants flexible, per-person hardware with occasional pooled runs; choose a GB300 for a shared production endpoint. See GB10 vs GB300: When Is It Time to Move?

How to build one sensibly

1. Start with two units linked directly; add the switch only when a model needs more than 256 GB. See Two NVIDIA GB10 Systems, One AI Model.
2. Buy the units together so their software matches.
3. Place them side by side with clear airflow, not stacked.
4. Choose the switch and cables together; confirm each port negotiates 200 Gb/s. Our post 400Gbps Networking for AI: Do You Really Need Such an Expensive Switch? compares options.
5. Put the switch and units on a UPS.
6. Use NVIDIA Sync's Cluster Assistant, then run NVIDIA's link speed test before loading a model.

Buying in Singapore

SourceIT supplies GB10 systems from ASUS, NVIDIA, HP, Dell, Lenovo, GIGABYTE and MSI, the MikroTik CRS804 DDQ and NVIDIA SN3700 switches, and cluster commissioning, with local warranty and a GST invoice. For a four-unit build, ask about volume pricing and we will specify matching cables. Browse the NVIDIA DGX Spark collection. 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. Government and university buyers can purchase through GeBIZ.

Frequently asked questions

Can four DGX Sparks be clustered? Yes. NVIDIA's Cluster Assistant supports up to four GB10 units, which must connect through a switch with a 200 Gb/s Ethernet port for each unit.

What can a four-unit GB10 cluster run? NVIDIA rates four units, with 512 GB of pooled memory, for models up to 700B parameters at reduced precision.

Is a four-unit cluster four times faster? Not usually. NVIDIA measured about 3x one unit's single-request speed on Qwen3.8-27B with four units and a switch, and for many users four separate units serve more requests than one cluster.

Which switch should I use? NVIDIA names no specific model. Options include the MikroTik CRS804 DDQ (S$2,255 at SourceIT) for small clusters and the NVIDIA Spectrum-2 SN3700 (S$28,523.25) for a full data-centre switch.

Is four GB10s better than one DGX Station GB300? For a flexible research lab, often yes on cost. For a shared production endpoint with many users, a GB300's coherent 748 GB memory, higher compute and management features are better.