Foxconn's Visionbay.ai brings a 5 MW Nvidia HGX B300 cluster online in Taiwan, with 90 percent utilisation at launch

On 31 July 2026, Foxconn subsidiary Visionbay.ai commercially launched a 5 MW AI cluster in Taiwan built on Nvidia HGX B300 systems, claiming it is the largest deployment of its kind on the island^[1]. The company says more than 90 percent of capacity is already in use^[1], an unusually high utilisation figure for a facility that entered commercial operation the same day it was announced.

What the cluster is

The cluster features direct liquid cooling, a 2N high-availability architecture, and Nvidia Quantum-X800 InfiniBand networking, purpose-built for large-scale AI training and inference. It has signed its first long-term enterprise customers and entered commercial operation with utilisation exceeding 90 percent. The facility operates to Tier III-class standards, according to Visionbay.ai.

The hardware choice matters for the power story. Each Nvidia B300 GPU in the HGX configuration draws up to 1,100 W TDP, roughly double the draw of an equivalent H100 system. At that density, a 5 MW envelope leaves limited headroom for cooling overhead and power distribution losses - direct liquid cooling is not optional at this power class.

Visionbay.ai was launched in November 2025 as Foxconn's dedicated business unit for supercomputing and cloud-AI operations^[1], making its first public debut at Hon Hai Tech Day that month. The 31 July commercial launch is therefore the first live revenue-generating deployment from a unit that is less than nine months old.

The certification angle

Alongside the launch, Visionbay.ai achieved Nvidia Exemplar Cloud validation for training workloads on HGX B300, making it one of the first ten organisations globally - and the second in Asia - to receive the certification. The Exemplar Cloud programme is Nvidia's production-grade validation for AI training infrastructure; qualifying for it signals that the cluster's networking, storage, and operational procedures meet Nvidia's own benchmarks for sustained large-scale training runs, not just peak throughput.

That certification is commercially significant. Enterprise buyers running foundation-model training at scale increasingly require validated infrastructure rather than raw GPU counts, a shift Visionbay.ai itself flagged: demand is moving from simply owning GPUs to having systems capable of reliably supporting commercial applications.

Taiwan's grid context

The launch lands against a tightening power backdrop. Taiwan's AI data centre push is exposing a wider problem: AI needs vast, reliable power, but grids, permits, and green-energy rules are not keeping up. Taipower has received 79 applications for AI data centre power connections in five years, and the real strain is not on total island supply but on regional grids and substation bottlenecks.

Taiwan's Ministry of Economic Affairs began enforcing a PUE limit of 1.5 for hyperscale sites in November 2025 and rolled out tiered electricity tariffs in January 2026 that charge surcharges of up to 20 percent for inefficient facilities. Visionbay.ai's use of direct liquid cooling positions it to meet those thresholds, though the company has not disclosed the facility's PUE or its power source.

What comes next

Visionbay.ai says it plans to expand AI infrastructure across the Asia-Pacific region, deploying both Nvidia Blackwell and the next-generation Vera Rubin architecture. Foxconn's vertical integration - spanning component manufacturing, server R&D, supply-chain logistics, and cooling technology - gives the unit a structural cost advantage over pure-play colocation operators trying to source the same hardware on the open market.

The 90 percent utilisation claim at day one is the number to watch. If it holds through Q3 2026, it will confirm that enterprise demand for validated, liquid-cooled Blackwell compute in Asia is running well ahead of available supply - and that Foxconn's move to monetise its hardware supply-chain position directly, rather than selling servers to others, is generating real load.