Foxconn and Nvidia Plan $1.4 Billion Taiwan Supercomputing Center for First Half of 2026

A 27 MW project puts Taiwan at the center of a new AI push

Foxconn is building a $1.4 billion supercomputing center in Taiwan with Nvidia, targeting operations in the first half of 2026. The facility is rated at 27 megawatts (MW) and is being developed by Visonbay.ai, Foxconn's new unit for AI supercomputing and cloud services.

Once operating, the center is set to house Taiwan's largest advanced graphics processing unit (GPU) cluster and become the first installation in Asia built around Nvidia's Blackwell GB300 platform. GPUs are processors that handle many calculations at once. Connecting them into a cluster allows them to share the enormous computing workload required by modern artificial intelligence.

As of November 23, 2025, the project brings together two parts of Foxconn's expanding AI business: manufacturing the racks that hold powerful computing systems and selling access to those systems as a service. Nvidia supplies the underlying computing platform, while Foxconn brings its growing capacity to manufacture and integrate AI racks.

The center is intended for customers across the region that need computing power for training large language models, running foundation models at scale, and developing robotics and autonomous systems. Training teaches a model patterns from data. Inference is the work a trained model performs when it answers a question, makes a prediction or processes a new request.

Foxconn's plan also addresses a practical problem for customers: assembling a private AI facility takes money, specialist staff and time. Visonbay.ai is expected to offer access to GB300 capacity without customers having to arrange their own hardware, cooling and computing space.

What makes Blackwell GB300 different?

Nvidia's Blackwell generation is designed for larger models, faster training and a lower cost for each unit of computing work. The GB300 line combines fast GPU cores, high bandwidth memory and an updated NVLink fabric, the connections that allow dozens of GPUs to operate as a shared pool.

Large AI models spread their work across many processors. Those processors must repeatedly exchange model parameters, the values learned during training, and activations, the intermediate results produced as data passes through a model. Faster exchanges help keep the GPUs working instead of waiting for information.

GB300's memory and connections are therefore important alongside the speed of its processors. Moving data more efficiently between GPUs can improve both training and inference when models are too large for a single processor to handle.

Blackwell also includes security features and support for confidential computing. These capabilities matter to organizations that need to protect their data and model weights, the learned values that shape a model's behavior, while meeting regulatory obligations and internal rules during training and inference.

Inside the racks that connect the cluster

An AI rack contains servers packed with GPUs, fast network switches and storage, together with power distribution equipment and cabling. The rack turns a collection of components into a system that can be installed, connected and operated as part of a larger cluster.

Within a rack, NVLink allows GPUs to communicate with high bandwidth and low latency. Bandwidth describes how much data can move at once; latency is the delay before it arrives. Both matter when processors must coordinate their work continuously.

Between racks, fast Ethernet or InfiniBand networks connect the cluster. A job that exceeds the capacity of one rack can then be distributed across several. The Taiwan center's design combines these dense computing systems with factory built racks to shorten the time needed to bring capacity into service.

Why liquid cooling is central to the design

GB300 systems concentrate so much computing power in a small area that their heat loads exceed what traditional air cooling can comfortably handle. Vendors are shifting toward direct liquid cooling and other advanced designs that move heat away from hot components more effectively.

Nvidia's rack systems paired with GB300 GPUs reflect that change. Plumbing, manifolds that distribute liquid, and cooling controls are built into the racks at the factory. Liquid cooling makes it practical to fit more computing power into each square meter of floor space.

Cooling with liquid can also reduce electricity spent on fans and help lower power usage effectiveness (PUE). PUE compares all the electricity used by a facility with the electricity used by its computing equipment. Keeping the two amounts closer together means less energy is being spent on cooling and other supporting operations.

Foxconn's Taiwan project places particular weight on liquid cooling to maintain performance and control energy overhead. Operators must balance those gains against new maintenance requirements, including checks on coolant chemistry, leak detection and the quick disconnect fittings used when servicing equipment. The expected benefits include more predictable temperatures, manageable noise and better efficiency through the year, including Asia's hot summer months.

Renting computing power instead of building a facility

Nvidia used Foxconn's tech day to set out a choice facing companies that need AI capacity: buy and operate a cluster, or rent computing resources from a specialist operator. Rapid improvements in GPUs make it harder for buyers to plan around long equipment lifetimes unless they expect to refresh their systems frequently.

Alexis Bjorlin, a vice president at Nvidia, told attendees that renting could offer businesses greater flexibility as products and models change:

Renting compute resources may offer a far better return on investment, enabling flexibility and enabling companies to scale their compute according to both product and business cycles.

Visonbay.ai is expected to offer that rental model in Taiwan. Enterprises could obtain GB300 capacity without waiting to recruit an operating team or complete their own facility work. Public sector organizations and research institutions could also spread spending through more predictable budgets instead of making a large payment at the start.

Foxconn gains a recurring services business alongside its manufacturing operations. Nvidia gains a larger installed base for Blackwell and another route for customers to use its GPUs without managing procurement or building new data halls.

How much computing can 27 MW support?

The Taiwan center's 27 MW rating makes it a major installation for the island, with enough capacity to host thousands of advanced GPUs. Power is a useful starting point for comparing AI data centers because dense clusters require substantial electricity to operate and remain cool.

The project sits toward the high end of deployments housed in a single building in Asia. Hyperscale campuses can reach 100 to 300 MW across multiple buildings. Foxconn's plan is more focused: one dense cluster capable of handling training and inference for many customers at the same time.

Capacity at this scale can support advanced model training, research on a national scale and commercial customers that lack the time or capital to assemble facilities of their own. The proposed mix of users is intended to keep the equipment busy as demand changes across industries and business cycles.

Semiconductor companies could use the cluster to simulate chip designs and photolithography, the process used to form patterns during chip production. Biotechnology teams could run protein structure research and drug screening with generative models. Financial technology and electronic commerce businesses could adapt large language models for recommendations, fraud detection and customer support.

Robotics developers could train the rules that guide machine behavior and run pipelines that generate artificial training data at production speed. Industrial groups with factories in Taiwan and Japan could use the center for simulation and computer vision linked to manufacturing.

Why Taiwan is a useful location

Taiwan brings together several parts of the advanced computing supply chain. TSMC fabricates many of the chips, contract manufacturers assemble servers and racks, and an established network services market connects customers with cloud providers.

Locating a GB300 cluster within that ecosystem puts hardware suppliers, system integrators and early customers closer together. That proximity can speed up feedback about how equipment performs and how systems should be configured. The location also reduces shipping time and import complexity for regional buyers.

Local universities and national laboratories are likely users because the cluster would give them access to substantial computing resources within Taiwan. Such sovereign computing capacity allows national research and AI programs to use infrastructure located on the island.

Regional cloud providers and software companies could reserve capacity for training, fine tuning and inference. Fine tuning means adapting an existing model for a more specific task. Research teams could also book computing time for large training runs that would be difficult to schedule on public clouds when demand is high.

Foxconn is expanding production and annual AI investment

Foxconn has spent the past two years developing AI and cloud infrastructure as a second engine of growth. The company says it can produce roughly 1,000 AI racks each week and plans to increase that rate in 2026.

Each rack requires more than assembling servers. Workers must install liquid cooling connections, wire systems for very high bandwidth and test the finished equipment under load before shipment. Foxconn's weekly production rate shows the scale of its move beyond phone assembly into infrastructure that customers depend on for continuous operations.

That manufacturing capacity can shorten customer lead times and help equip the Taiwan center quickly once it opens. Chairman Young Liu has also set out a longer investment plan, committing $2 billion to $3 billion each year to AI projects and capacity.

Foxconn's ambitions extend beyond Taiwan. Earlier plans included work with SoftBank at a former Foxconn electric vehicle site in Ohio to build data center equipment, part of a wider effort to supply large AI programs in the United States. Manufacturing arrangements abroad give the company more logistical options as supply bottlenecks move between regions.

AI factories and electric vehicles widen the strategy

Foxconn and Nvidia are also working to bring AI into production lines. Their active collaboration covers vision systems, digital twins and robotics, with the aim of reducing defects, improving yields and shortening the time needed to introduce new products.

A digital twin is a virtual representation of a physical operation that can be used for simulation. Alongside vision systems and robotics, it forms part of the companies' effort to make manufacturing more responsive and efficient. Their work is expected to inform the design and operation of Foxconn's own factories.

Spencer Huang, an Nvidia product leader focused on robotics and the son of Nvidia founder Jensen Huang, described the collaboration:

Nvidia is working with Foxconn to bring AI to factories and manufacturing lines.

Foxconn also showed progress in its electric vehicle program at the event. Its Model A, designed by Japanese engineers, is aimed at Japan's market. Foxconn plans to manufacture the vehicle in Japan once sales volumes can support a local plant, as part of its strategy to serve automakers seeking to outsource more assembly.

Electricity, water and operating reliability matter

A 27 MW cluster concentrates electricity use and heat in one location. That makes grid connections, sources of cleaner power and water management important parts of operating the facility, alongside the performance of the GPUs themselves.

Data center operators in Taiwan have been expanding their use of power purchase agreements for wind and solar electricity. They are also pursuing heat reuse and chiller upgrades to reduce waste. Direct liquid cooling can reduce fan energy and capture heat more effectively while supporting dense GPU installations.

Liquid cooling requires careful design and quality controls to avoid downtime. Service teams need training in handling fittings, monitoring coolant and detecting leaks. Reliable operation depends on the cooling infrastructure keeping pace with the computing equipment during demanding workloads.

The Taiwan center's eventual efficiency and uptime will provide a benchmark for future Blackwell deployments in Asia. Other sites in the region are planning rooms cooled with liquid and modular designs suited to GPU clusters, making the experience of a large operating installation valuable.

The schedule runs into late 2026

Foxconn's target is to have the Taiwan site ready during the first half of 2026. The company also has board approval to procure equipment for AI computing clusters and a supercomputing center through late 2026, pointing to a staged build and room to expand if demand develops quickly.

The project and related investment plans follow this timetable:

• November 23, 2025: Foxconn's plans describe a $1.4 billion center rated at 27 MW, developed by Visonbay.ai with Nvidia's GB300 platform.
• First half of 2026: Foxconn targets readiness and the start of operations at the Taiwan facility.
• During 2026: Foxconn plans to raise AI rack production above its current rate of roughly 1,000 per week.
• Through late 2026: Board approval covers equipment procurement for AI computing clusters and a supercomputing center.

A precise opening date, customer prices, the full Visonbay.ai service catalog and regional connectivity options have yet to be detailed. Customers will also be watching the availability of GB300 racks as they plan their own workloads.

For enterprises replacing older accelerators, the planned service offers access to GB300 systems with higher memory bandwidth and better performance per watt. For Nvidia and Foxconn, it combines a new destination for Blackwell hardware with a business that earns revenue from operating the equipment. The $27 MW project and Foxconn's annual AI investment plan place manufacturing capacity and computing services within the same expansion strategy.

Key Points

• Foxconn and Nvidia plan a $1.4 billion supercomputing center in Taiwan, targeting operations in the first half of 2026.
• The 27 MW facility is being developed by Visonbay.ai, Foxconn's AI supercomputing and cloud services unit.
• The center is set to become Taiwan's largest advanced GPU cluster and Asia's first installation built around Nvidia's Blackwell GB300 platform.
• Visonbay.ai is expected to rent computing capacity to enterprises, researchers and public sector users.
• Foxconn says it produces roughly 1,000 AI racks weekly, plans higher output in 2026 and intends to invest $2 billion to $3 billion annually in AI.
• Liquid cooling, electricity supply and reliable operation are central to the facility's design and performance.
• Foxconn and Nvidia are also collaborating on AI for factories, while Foxconn continues its electric vehicle plans for Japan.
• Equipment procurement approval extends through late 2026; a precise opening date and detailed service terms have yet to be given.