Once we start to see diminishing returns to model size and model performance begins to plateau (we are already nearing the plateau) we will have a set of models

Once we start to see diminishing returns to model size and model performance begins to plateau (we are already nearing the plateau) we will have a set of models all pretty close to one another in performance which will be made available for downstream workloads on demand. The consequences of this will be twofold. First, there will be less incentive for new entrants to train proprietary models as performance across SOTA models will be uniform and thus less possibility of differentiation and gaining market share. Second, there will be less incentive for SOTA model developers to invest heavily in GPUs because of diminishing returns. Now the beauty of LLMs is that inference requires orders of magnitude less compute than training. So once we have a stable of performance plateaued models, compute demand will drop off a cliff. Once we get over this mad rush to train train train, GPU orders are going to crater. The two critical assumptions I'm making are 1) we are rapidly approaching a performance plateau and 2) models other than LLM's are much less widely useful than LLM's. Image generation I see as a fad, I played around with it a ton last year but haven't touched it since. Other niche models will have very little general applicability compared to LLM's.