Sugon scaleX40 & scaleX640 Superpods — Sourcing

Sugon scaleX superpods — specified, not guessed

China-built AI superpods with Hygon DCU accelerators and the scaleFabric interconnect. We work out the full configuration — including the figures Sugon does not publish — and source it for the markets where it deploys.

Maintained by Haink's AI infrastructure team · Updated September 2026

The systems, specified

scaleX40-3G40 DeepComputing 3 accelerators · 5.62 TB HBM · 28 PFLOPS FP8 · 448 GB/s peer-to-peer HSL · 45 kW in 16U — roughly 1.8× a DGX H200 on aggregate memory. Full specs →

scaleX640640 accelerators in a single cabinet under phase-change immersion, PUE 1.04 — one scale-up domain, or partitioned to your workload. Full specs →

scaleFabricThe interconnect underneath — 400G adapters, 800G switch ports, 64 Tb/s aggregate, 0.9 µs end to end, scaling toward a 114,000-card ceiling. Full specs →

The rest of the stack:ParaStor storage and Sugon liquid cooling are specified alongside the compute. Sugon guide →

How we quote it

Specified firstSugon publishes no list part numbers and no prices. We build the bill of materials from the actual configuration, so a quote is comparable — not a placeholder.

Sourced for the marketAvailability, lead time and terms are confirmed per configuration and destination — not promised from a shelf.

Honest on the unknownsWhere a figure is unpublished or a claim is the vendor's own, we say so — the three specs Sugon never states are called out up front.

Weighing it against NVIDIA? See scaleX40 vs DGX H200 vs B300 and the Chinese superpod comparison.

Need a scaleX quote?

Tell us the target — accelerator count, workload or scaleX model — and we come back with a specified quote within one business day.