Because of LLMs we are back to memory being king (KV cache).
But there was another thing - clusters of thousands/millions of machine instead of one big iron with all the memory.