Rack-scale demonstration Deployment

SoftBank–Foxconn H200 rack cooling pilot

SoftBank, ZutaCore, and Foxconn demonstrated a rack-integrated two-phase direct-to-chip system using NVIDIA H200 GPUs.

NVIDIA H200SoftBankFoxconntwo phase

Direct answer

What this record says

SoftBank–Foxconn H200 rack cooling pilot is a pilot cooling deployment in Japan, operated by SoftBank. It is recorded at One rack-scale demonstration. It was announced May 21, 2025. Last reviewed Aug. 9, 2026 against 1 source.

1 sourceReviewed Aug. 9, 2026
Operator
SoftBank
Stage
Pilot
Capacity
One rack-scale demonstration
Rack density
Not disclosed

Project brief

What is publicly known

Status: pilot. SoftBank, ZutaCore, and Foxconn reported that they operated and evaluated one rack-scale system at a SoftBank data center in February 2025. Foxconn built the server around NVIDIA H200 graphics processing units, ZutaCore supplied HyperCool two-phase direct liquid cooling, and SoftBank integrated the rack.

The partners say the server passed NVIDIA's NVQual temperature test and that the rack achieved a measured partial power usage effectiveness of 1.03. That ratio means the measured cooling scope used about three units of power for every hundred units attributed to computing. It is not whole-building power usage effectiveness, and the source does not publish the test boundary or raw data.

Cooling relevance comes from integration. The dielectric fluid boils in cold plates on processors, graphics processors, and memory, then condenses in the loop. SoftBank also designed the rack around 19-inch and Open Rack version 3 equipment, rear power and cabling, and future robotic handling.

This is stronger evidence than a concept announcement because hardware ran and a result was reported. It is still supplier and partner evidence from a controlled demonstration. Unknowns include rack kilowatts, test duration, inlet and return temperatures, load profile, redundancy, failure behavior, service procedure, fluid loss, and whether the system carried production work.

The release says the parties will pursue commercialization, which confirms that the demonstrated design was not yet a generally deployed product. A named production site, customer order, commercial warranty, repeat test by an independent lab, or sustained workload data would strengthen the status. Broad rollout should require several commissioned racks and operating evidence.

A result that cannot be reproduced at full load would weaken the efficiency claim without erasing the fact that the pilot occurred.

Status is labeled from public evidence. Supplier orders, MoUs, pilots, projects under delivery, and operating sites are not counted as equivalent.

Status
Pilot
Operator
SoftBank
Location
Japan
Capacity
One rack-scale demonstration
Rack density
Not disclosed
Compute
NVIDIA H200 GPUs
Announced
May 21, 2025

Connected records

Companies involved

Connected records

Cooling architecture

Connected records

Equipment chain

Evidence ledger

Sources and evidence

The links below show where the factual claims came from. Supplier specifications remain supplier-reported unless the record names independent operating evidence.

  1. 01
Record information
Record ID
DCC / DEPL / SOFTBANK-FOX
Record reviewed
Aug. 9, 2026
Newest source published
May 21, 2025
Record first published
Aug. 9, 2026
External sources
1
Search visibility
Offered to search engines

This field was researched and is not published by the party who would know:

  • Rack density

Decide between

Comparable deployments

Records covering the same cooling architecture, most-evidenced first.

View all deployments

Add two or more, then open the comparison to see every stated claim, its duty point, and its source side by side.