[arathore]

If by running models you mean just the inference phase, then even today you can run large family of ML models on commodity hardware (with some elbow grease, of course). The training phase is generally the one not easily replicated by non-corporations.