A new NVIDIA technical blog post discusses AI model co-design, focusing on how model-design choices shape throughput and interactivity without sacrificing accuracy. AI performance is defined by three dimensions: accuracy, throughput (tokens per second), and interactivity (dominated by latency). Holding accuracy fixed, the problem becomes a two-dimensional Pareto frontier where improving one usually costs the other. The trade-off involves the system deployer, who prioritizes fleet throughput, and the user, who values lower first-token and inter-token latency. Deployment varies along two axes: workloads range from short to long context, and service goals range from throughput-oriented to latency-oriented. Each quadrant needs different optimizations; for example, long-context, throughput-oriented serving spends most of its time in attention, while latency-oriented serving adds model parallelism. Amdahl’s law applies: if attention is 77% of runtime, tuning feed-forward layers yields only marginal gains. The post provides rules of thumb for hardware-aware design, with subsequent chapters targeting different dimensions to help avoid compute bottlenecks.