GOOGLE’S TURBOQUANT SHRINKS AI MEMORY BOTTLENECK 6X — RUNS BIGGER MODELS ON THE SAME HARDWARE Google Research unveiled TurboQuant at ICLR 2026 — a KV cache compression algorithm that cuts the biggest memory bottleneck in LLM inference down to 3-4 bits per element with no retraining required, delivering a 4–6x memory reduction and up to 8x faster performance on H100 GPUs. It could dramatically cut the cost of running frontier models at scale. Keywords: TurboQuant, KV cache compression, Google AI research, LLM inference efficiency