Foxconn, Nvidia to launch $1.4b Taiwan supercomputing center in 2026

Foxconn said its US$1.4 billion supercomputing center in Taiwan, built with Nvidia, will be operational by H1 2026 and is set to become the country’s largest GPU cluster.

The facility will use Nvidia’s Blackwell GB300 chips and is expected to be the first GB300 AI data centre in Asia, according to Neo Yao, CEO of Foxconn’s new AI supercomputing unit, Visonbay.ai.

Nvidia vice president Alexis Bjorlin told attendees at Foxconn’s tech day that renting compute resources could offer businesses more flexibility than building their own facilities.

Foxconn, which assembles iPhones for Apple, has moved into AI data centers and is Nvidia’s main manufacturer of AI racks.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Taiwan’s GB300 cluster will test liquid cooling in Asia

Each GB300 NVL72 rack (an Nvidia rack-scale system) draws 120 kilowatts (kW) 1, so vendors target total Power Usage Effectiveness (PUE) near 1.05 and partial PUE (a cooling-only efficiency metric) around 1.01 with two-phase direct liquid cooling (DLC) 2. Chief Telecom (a Taiwanese telecom and data center operator) finished a 45-day retrofit at a legacy site 3, and Taiwan’s cooling market projects 12.32% compound annual growth rate (CAGR) through 2031 4.
Google Cloud ships A4X Max instances (a GPU-accelerated virtual machine type) 5. Foxconn targets H1 2026 for the expected ‘first GB300 AI data centre in Asia’, which will test whether Taiwan original equipment manufacturers (OEMs) can deliver 800 gigabits per second (Gb/s) per GPU 6 and 130 terabytes per second (TB/s) of NVLink (Nvidia’s high-speed GPU interconnect) 7.

Contract makers building 1,000 AI racks a week will shift hyperscaler buying

Foxconn can build about 1,000 AI racks each week, and GB300 shipments are expected to rise 300% in Q3 2025 8. That scale adds an option beyond hyperscalers and supports ‘rental compute’ models (renting compute resources from a provider) that Nvidia pitched at Foxconn’s tech day.
Foxconn plans US$2-3 billion in AI spending each year. Investors and data center operators can target sites that support 100+ kW per rack 3 with ZutaCore-style two-phase cooling (a vendor approach to two-phase liquid cooling) near 1.01 partial PUE 2, then finish liquid cooling retrofits in 45 days 3.