AI Compute Shifts from Pre-Training to Post-Training Finds SemiAnalysis

18 Oct 2026, 3:21 pm · 1h ago · 1 min read · OfficeChaiNew

Research firm SemiAnalysis reports a fundamental shift in frontier AI compute allocation over the last two years. While inference compute share has remained relatively steady, resources previously dedicated to massive pre-training runs are increasingly moving toward post-training techniques, such as reinforcement learning and test-time compute. This shift reflects the industry's evolving focus on model alignment, reasoning capabilities, and efficiency rather than solely scaling raw parameter sizes. As frontier AI labs encounter diminishing returns on brute-force pre-training, optimizing post-training pipelines is emerging as the primary frontier for driving real-world intelligence improvements in modern large language models.