Problems began at the end of 2024 when NVIDIA was supposed to start mass production of Blackwell servers. It was reported that the new architecture had a design defect, leading to increased thermal loads. Despite the company’s assurances of resolving the issue, a recent report from The Information indicates that the servers are still experiencing failures related to improper chip connections.
According to experts, the key problem is the CoWoS packaging technology from TSMC used to connect components. NVIDIA attempted to make changes in production by replacing the GPU mask, but this did not yield the desired result. Amid these difficulties, major customers have started switching to the previous, time-tested Hopper lineup.
