Foxconn launches AI factory powered by Nvidia supercomputing cluster

Foxconn, the company best known for assembling other people’s gadgets, has launched an AI factory built around an Nvidia DGX supercomputing cluster.

The world’s most famous contract manufacturer is no longer just building the machines that power AI. It is putting them to work on its own models and its own production lines.

What Foxconn actually built

The cluster is an Nvidia DGX SuperPOD running on the chipmaker’s Blackwell architecture. Think of a SuperPOD as a prepackaged data center: a large set of Nvidia servers networked so they behave like one very large computer.

Specifically, Foxconn is using DGX GB200 systems. GB200 is Nvidia’s Blackwell-generation platform, the successor to the H100 chips that powered much of the first wave of the generative AI boom.

The performance gap between the two generations is the headline number. Foxconn’s GB200 setup trained 120 billion parameter models 2.48x faster than its earlier H100-based systems did.

The cluster also chewed through 1 trillion tokens in just 12 days. At its peak, the system hit 4,900 TFLOPS of throughput.

From chatbot to factory floor

The raw compute feeds into FoxBrain, Foxconn’s in-house large language model.

Foxconn reports that production scheduling time has dropped 25% with the new system in place. Engineering workloads tied to injection molding fell by 50%. Injection molding is the process of forcing molten plastic into a mold to create parts, like phone casings or connectors, and setting it up correctly has traditionally required a lot of skilled trial and error.

Foxconn also reports a correction rate above 90% for defects on its SMT production lines. SMT, or surface-mount technology, is the method used to place tiny electronic components directly onto circuit boards.

A long relationship with Nvidia

This launch is the latest chapter in a partnership between Foxconn and Nvidia that stretches back several years. The two companies have outlined plans for large-scale Blackwell-based supercomputers in Taiwan.

Foxconn also plays a direct role in Nvidia’s hardware business. The company manufactures the AI server racks that hyperscalers and enterprises buy to fill their data centers.

Foxconn’s annual investment in AI infrastructure is estimated at between $2 billion and $3 billion.

What this means

The most important takeaway is not the 2.48x figure, impressive as it is. It is that a manufacturer is reporting concrete operational gains, like 25% faster scheduling and a 50% cut in molding engineering work, rather than vague promises about transformation.

For Nvidia, Foxconn functions as a high-profile proof of concept. A Blackwell cluster delivering more than double the training speed of H100 hardware in a real production environment is exactly the kind of evidence that helps justify upgrade cycles for other buyers.

Risks remain, of course. Spending in the range of $2 billion to $3 billion a year needs to keep producing returns, and efficiency gains measured in one set of processes do not automatically carry over to every product line or factory.