NVIDIA reportedly takes Foxconn’s entire CPO rack supply ahead of Computex

NVIDIA is moving fast on co packaged optics for future AI data centers, and Foxconn may already be feeling the pressure.

According to a new report, Foxconn has shipped its entire supply of co packaged optics, or CPO, racks to NVIDIA ahead of schedule. The supply is reportedly so tight that Foxconn may not even have display units left to show at Computex this year.

CPO technology is important because AI data centers need faster and more efficient ways to move data between GPUs, CPUs, switches, and racks. As AI systems scale up, networking becomes one of the biggest bottlenecks. Co packaged optics can help by bringing optical connections closer to the switching silicon, improving bandwidth and reducing power loss compared with traditional electrical interconnects.

NVIDIA is reportedly accelerating CPO development ahead of its future Feynman platform in 2028. The company is also expected to use CPO more broadly as Rubin and later AI platforms move deeper into data centers.

Foxconn is said to be NVIDIA’s key supplier for these optical switch racks. The company reportedly began mass prototype shipments in Q1 2026, with mass production and broader shipments expected in Q3 2026. The racks are being made at Foxconn’s Vietnam plant.

The scale of NVIDIA’s demand appears to have changed quickly.

Foxconn had reportedly planned to show CPO switch cabinets at Computex, but industry sources claim all prototype racks, including display cabinets, have already been sent to NVIDIA.

That is a strong signal of how urgent NVIDIA’s AI infrastructure roadmap has become. The company is not only competing on GPUs anymore. It is building entire rack scale systems where compute, networking, memory, optics, and power delivery all matter.

For Foxconn, this could become a valuable business. The report says CPO solutions may bring double digit gross margins and higher average selling prices, eventually contributing more than 15 percent of Foxconn’s revenue.

The timing also matters because NVIDIA is not alone. AMD is also moving into advanced AI rack scale infrastructure, and the broader industry is racing to solve networking limits as AI clusters grow larger.

CPO may become one of the key technologies behind the next generation of AI data centers. Blackwell and Rubin already push massive scale out requirements, and Feynman is expected to raise those demands even further.

For now, this remains a rumor based on Taiwanese industry sources. But if accurate, it shows NVIDIA locking down a critical part of its future AI infrastructure supply chain early.

The bigger message is clear: AI hardware is no longer just about the accelerator chip. The companies that win will also need the best networking, optics, racks, cooling, and supply chain control. NVIDIA taking Foxconn’s entire CPO rack supply would fit that strategy perfectly.