Nvidia faces competition from startups developing specialised chips for AI inference as demand shifts from training large language models to running them efficiently.