May 26th news reveals that due to the high demand for artificial intelligence (AI) chips in the market, NVIDIA's data center GPU sales have been booming, leading to tight capacity for TSMC's CoWoS packaging over the past year.
As one of the market's high-performance AI chips, NVIDIA's GPU based on the Blackwell architecture is set for a large-scale release, further exacerbating the tight packaging capacity situation.
To address this challenge, NVIDIA intends to introduce advanced FOPLP (panel-level fan-out packaging) technology on the Blackwell architecture GB200 chip and plans to start implementing it in 2025. This technology choice not only provides NVIDIA with more flexibility in the packaging field but also helps alleviate the pressure caused by insufficient packaging capacity to some extent.
According to market research, NVIDIA's GB200 supply chain has been initiated and is currently in the design fine-tuning and testing phase. It is expected to ship approximately 420,000 units this year, with a potential increase to 1.5 to 2 million units next year. This positive shipment forecast further underscores the strong demand for NVIDIA's AI chips in the market.
It is worth noting that NVIDIA originally planned to introduce FOPLP packaging technology in 2026, but given the rapidly changing market conditions, the company has decided to accelerate the timeline. Reportedly, NVIDIA's chosen FOPLP packaging technology utilizes a glass substrate that can withstand high temperatures for extended periods while maintaining optimal performance, ensuring the stability and reliability of the chips.
