• The 5MW cluster uses Nvidia HGX B300 systems, direct liquid cooling and Quantum-X800 networking
  • Visionbay.ai says long-term customers are signed and utilisation exceeds 90%, but has not published supporting workload data

The fact

Foxconn subsidiary Visionbay.ai has launched a 5MW AI computing cluster in Taiwan using Nvidia HGX B300 systems. The company says the infrastructure is commercially available for large-scale AI training and inference and that more than 90% of its capacity is already in use. It has not disclosed the cluster’s location.

Visionbay.ai was introduced in 2025 as Foxconn’s business for supercomputing and cloud AI services. Foxconn says the new cluster uses direct liquid cooling, a 2N infrastructure design and Nvidia Quantum-X800 networking. Nvidia lists Visionbay.ai as a provider in its Exemplar Cloud programme, which assesses participating infrastructure against workload performance, security and reliability criteria.

The assessment

Visionbay.ai is moving Foxconn from supplying AI systems into operating the computing service built around them. The company must now allocate GPU capacity, keep customer jobs running under sustained load and maintain reliable access to the cluster. The main uncertainty is what the reported utilisation rate represents. It may refer to reserved capacity, assigned systems or GPUs actively processing paid workloads.

Those measures are not interchangeable, and Visionbay.ai has not published enough detail to show which one it is using or how the cluster performs when demand is high. For BTW readers, the 90% claim becomes useful only when it is linked to operating evidence. Service availability, queue times, and completed workloads would show whether Foxconn has built a dependable AI cloud service rather than only a well-specified cluster.

What to watch

Visionbay.ai has not disclosed its installed GPU count, customer mix or method for calculating utilisation. Service availability, queue times, completed workloads and customer renewals would give a clearer picture of how the cluster performs. Any expansion beyond 5MW should be assessed against those operating results.