Foxconn just landed orders to build compute trays for Google's TPU racks, and apparently, they are shipping them out at a 1:1 ratio with the TPU racks themselves. Google's 7th-gen TPUs come in these massive Superpod setups that pack over 9,000 chips per pod, and Foxconn is handling the computing tray side of things. The whole TPU platform is supposedly getting traction with companies like Meta since everyone is obsessed with inference performance right at this moment.
The supply chain is apparently going wild over Google's TPU stuff because inference workloads are becoming way more important than training, and companies want the best performance without burning cash. This does not mean NVIDIA is getting replaced anytime soon, but the interest in custom silicon from Big Tech is definitely picking up steam.
The supply chain is apparently going wild over Google's TPU stuff because inference workloads are becoming way more important than training, and companies want the best performance without burning cash. This does not mean NVIDIA is getting replaced anytime soon, but the interest in custom silicon from Big Tech is definitely picking up steam.
