Citrini analyst Zephyr challenged Sandisk's investor day comparison between HBF and HBM, arguing the demo setup may have made high-bandwidth memory look weaker than it would under higher-spec configurations. He said the company set total bandwidth for both HBF and HBM at 12.8TB/s, or 1.6TB/s per stack, while running the Qwen3-480B-A35B model in bfloat16. Zephyr said current model inference more commonly uses FP4 or FP8, which would place the model's capacity needs at roughly 240GB to 480GB depending on quantization. He also said Sandisk fixed HBM capacity at 192GB per GPU, even though a 16-layer HBM4E configuration with eight stacks could reach 512GB and deliver about 32TB/s of bandwidth. In his view, those choices mean Sandisk's result showing HBF needing only one to four GPUs while HBM required eight GPUs does not fully represent performance under higher-spec HBM deployments.