Google is now putting dusty old DDR4 RAM into brand new AI servers

Supply chains are struggling, and Google has decided that if the world runs out of shiny new memory, the graveyard of old hardware is the next best thing. Nikhil Cherian, the company’s supply chain infrastructure lead, recently admitted that memory is the final boss of AI development. Since HBM3e stacks now account for the vast majority of server costs, the company is harvesting DDR4 modules from retired servers to prop up their latest TPU8i setups.

The current TPU8i chips are absolute monsters, packing 288 GB of high-speed memory to feed the hungry AI models. To make this work alongside Google Axion host processors that were originally meant for DDR5, the engineers had to invent custom hardware adapters. These little boards act as a bridge between the ancient memory standard and modern motherboards, allowing the old sticks to function in a high-stakes AI environment.

Software trickery is doing the heavy lifting here as well. By optimizing data streaming and minimizing latency, the team is trying to ensure the slower DDR4 doesn't turn the entire cluster into a bottleneck. It is a desperate but effective way to shift the budget toward those expensive HBM3e stacks that actually matter for performance.

Sustainability is a nice word for it, but this is clearly a frantic attempt to keep margins from collapsing under the weight of AI obsession. When the trillion-dollar kings of the internet start dumpster diving for memory, it’s a sign that the current hype cycle might be running on fumes and recycled scraps.

Source: Wccftech