NVIDIA has increased chip fabrication orders at TSMC as demand for GPU-based computational accelerators for AI applications grows. NVIDIA appears so bullish on demand for its CoWoS (chip on wafer on wafer) packaged GPUs that the company has placed additional orders for the full year.
TSMC has reportedly pledged to release an additional 10,000 wafers packaged with NVIDIA CoWoS during 2023 to meet growing demand for widely used chips. It is estimated that this will mean additional production of 1,000 to 2,000 wafers per month for the rest of the year. The article does not reveal which GPUs NVIDIA plans to increase production. TSMC’s monthly CoWoS packaging capacity is between 8,000 and 9,000 wafers. So if NVIDIA makes another 1,000 to 2,000 wafers available per month, utilization of its advanced packaging chip manufacturing capacity will increase significantly. This can leave other companies in the industry suffering from a lack of CoWoS services.
With the widespread adoption of the NVIDIA CUDA platform for AI and other high-performance workloads, dozens of major customers rely on the company’s hardware to run their AI applications. Just yesterday, Google announced its new A3 supercomputer with 26,000 NVIDIA H100 accelerators and 26 exaflops of AI power. Meanwhile, Microsoft, Oracle, and even Elon Musk’s upcoming AI project campaign have bought tens of thousands of NVIDIA AI GPUs over the past few quarters.
It is notable that NVIDIA is increasing orders for GPUs even after the company was unable to ship its top-performing accelerators to Chinese organizations without US government approval. It is likely that Chinese companies will be willing to buy less powerful A800 and H800 accelerators, or demand from American, European and Japanese companies will offset the drop in shipments to China.