Datadog's free guide shows how to connect AI spend, infrastructure, and model performance into a single view, correlating cost increases to architecture changes before they hit your cloud bill. You can break down AI costs by token, model, provider, and team, get alerts when inference volume spikes or API spend exceeds budget, and perform root-cause analysis in minutes.

A 70 billion parameter model takes up to 140 GB of space, while even a very good graphics card has only 48 GB. Models have grown 100-fold in recent years while consumer graphics memory has only doubled. The simplest way to run such models is purchasing capable hardware, but this is costly and doesn't work well with consumer hardware. Shrinking the model is the other option, but it shouldn't compromise intelligence.

Three techniques help shrink models without quality loss: packing fewer details, trimming unused pathways, and mimicking behavior. The model's weights—70 billion numbers stored as 16-bit values—form the bulk of a language model's intelligence, storing grammar, facts, and reasoning shortcuts. These weights are arranged into matrices across layers, and their relationships determine the model's capability, not individual values.