Google Develops DDR4 Recycling Supply Chain to Ease AI Memory Bottlenecks

Google is developing software and hardware solutions to address memory bottlenecks, according to Nikhil Cherian, senior director of supply chain infrastructure. According to ChainCatcher, the company is also dismantling retired servers to recover DDR4 components and build an internal recycling supply chain. Cherian said the AI industry has shifted rapidly from being compute-constrained to memory-constrained, with high-performance memory accounting for about 75% of the bill of materials cost of a given AI server. Goldman Sachs said memory prices are expected to keep rising in the third quarter, with PC DRAM prices forecast to increase 18% to 23% and server DRAM prices 13% to 18%. Trendforce data showed that in August, spot prices for DDR4 8GB and DDR5 8GB rose to $142 and $133, respectively. Google has also designed special hardware adapters that connect previous-generation DDR4 to next-generation AI servers and imports retired servers to remove and recycle their DDR4 modules. The company said two TPU ASICs launched this year were optimized for memory and could reduce memory demand to one-sixth of previous levels. Its TPU8i chip uses a dedicated layered memory design, relies on a high-speed DDR5 memory architecture for host-level tasks, and includes 288GB of HBM3e high-bandwidth memory per chip.