NVIDIA and Foxconn build AI factory in Taiwan – 10,000 Blackwell GPUs to power research and industry

The announcement came as little surprise, but still seems like a statement: NVIDIA and Foxconn want to build a so-called “AI Factory” together with the Taiwanese government. The new facility will house up to 10,000 GPUs of the current Blackwell generation and serve as a regional backbone for AI applications. The official partner in this project is the Foxconn subsidiary Big Innovation Company, which also has the status of an official “NVIDIA Cloud Partner”.

An AI factory with government support

The central component of the project is the planned provision of computing resources for research institutes, start-ups and established companies in Taiwan. The whole thing is being coordinated by Taiwan’s National Science and Technology Council (NSTC), which wants to guarantee access to the new infrastructure. Apparently, local companies and institutions in particular are set to benefit – a clear signal in the direction of technological independence and in-house development in the AI sector. Proven technology will be used to build and operate the infrastructure: NVIDIA supplies the systems based on the GB300 platform, specifically the NVIDIA GB300 NVL72 rack system is used, which relies on NVLink communication, Quantum InfiniBand and Spectrum-X Ethernet network architecture. This ensures scalability across multiple racks – a not insignificant aspect for future expansions.

TSMC wants to get involved – and accelerate

Semiconductor giant TSMC is also involved in the project. According to the announcement, its research units are planning to use the new computing power for their own development projects. There is talk of an “order of magnitude” increase in performance compared to existing systems – a statement that is not very specific, but at least indicates a high demand. Whether TSMC intends to use this to accelerate the development of new production processes or even the co-design of future chips remains an open question for the time being. One thing is clear: computing power is becoming a bottleneck factor – not only for AI start-ups, but also for established heavyweights in the industry.

NVIDIA DGX Cloud Lepton & Inception: programs for the next generation

In addition to the pure infrastructure, NVIDIA is also providing additional programs. The newly announced DGX Cloud Lepton marketplace will provide companies and research institutions with easy access to GPU resources – in the usual cloud form. At the same time, the NVIDIA Inception program remains active, which specifically promotes start-ups and provides their projects with computing time and technical support. The fact that NVIDIA also mentioned its Deep Learning Institute offerings in this context underlines its claim to not only supply hardware, but also to act as an educational platform. However, it remains to be seen whether these programs will prove to be structurally relevant beyond lip service.

Foxconn is thinking about smart cities and robot factories

For Foxconn, the focus is less on its role as a GPU provider and more on its own industrial application of the new systems. The AI Factory is intended to serve as a catalyst for the automation of Foxconn’s own business areas: Smart cities, electromobility and automated production are the keywords here. Specifically, the company plans to use the new infrastructure for connected car systems, ADAS applications (Advanced Driver Assistance Systems) and smart traffic control in urban areas. Traditional production topics such as digital twins and AI-supported quality assurance will also be addressed. It sounds a bit like “buzzword bingo”, but Foxconn at least has the industrial foundation to put these topics into practice.

Between euphoria and reality

However, the project is not entirely without dissonance. There are reports behind the scenes that sales of the new Blackwell Ultra Server (GB300) are apparently proving more difficult than expected. According to industry sources, potential customers are reluctant to buy, which is attributed to both the high investment costs and the lack of use cases. In practice, it is once again clear that even the most modern hardware is not an end in itself – without clear application scenarios, demand remains limited. The question also arises as to how many of the announced 10,000 GPUs will actually be installed initially and how busy the AI Factory will be. It is not uncommon for such large-scale projects to show discrepancies between announcement and implementation. In this respect, it remains to be seen whether Taiwan will actually become the AI hub of Asia – or whether it is just a PR footnote in NVIDIA’s expansion strategy. NVIDIA and Foxconn are building one of the country’s largest AI infrastructures to date in Taiwan – with up to 10,000 Blackwell GPUs, a clear focus on research and industry as well as close ties with government institutions. While the technical equipment appears ambitious, the economic implementation and actual benefits remain open for the time being.

Source: NVIDIA