AI Drug Discovery Faces Anti-Scaling Law as Pipeline Congestion Threatens Gains from Recent Breakthroughs Recent data highlights progress but underscores a structural shift. At the 2026 ASCO meeting, daraxonrasib demonstrated a 60% reduction in mortality risk for late-stage pancreatic cancer, doubling median overall survival to 13.2 months. Concurrently, Eli Lilly VERVE-102 program achieved an 88% reduction in PCSK9 levels via single-dose in vivo DNA editing, with LDL dropping 62%. While these milestones signal therapeutic potential, Liang Chang, biotech venture capitalist and cancer biologist, warns that the industry is approaching an Anti-Scaling Law. Contrary to the exponential performance gains seen in software, AI-driven drug development risks increasing pipeline congestion without proportionally raising clinical success rates. Chang argues that daraxonrasib and VERVE-102 succeed because they target RAS and PCSK9, among the most genetically validated targets in history, not because AI generated novel insights. Technology scaling relies on replicability, whereas biology lacks these dynamics. The first successful target often captures the highest-value opportunities. Subsequent efforts face diminishing returns as remaining targets present greater complexity regarding disease mechanisms, toxicity, and patient variability. Success on one target does not translate to adjacent targets. The pool of viable targets is severely constrained. Of approximately 20,000 human protein-coding genes, only 700 have yielded approved drugs. A 2024 study indicates that top-tier targets, comprising the top 0.28% by genetic priority, boast a nine-fold higher probability of regulatory approval compared to average candidates. These targets represent a non-random sample of high-confidence biology. While AI significantly enhances engineering efficiency, accelerating molecular design, structure prediction, and workflow optimization, it has not breached the ceiling of biological validation. Clinical approval rates have stagnated, dropping from 10% in 2014 to approximately 8% in 2020, particularly at the Phase 2 proof-of-concept stage where AI provides limited leverage. As AI lowers execution costs, capital and teams will likely converge on validated targets to mitigate risk. This behavior could trigger intense competition within mature categories. The PCSK9 landscape already includes dozens of approved or clinical assets across modalities, and CD19 exceeds 300 active clinical programs. Chang predicts that without a mechanism to identify novel biology, AI adoption will homogenize development efforts rather than diversify the pipeline. First-in-class targets emerge from human genetics and physical experimentation, not algorithmic inference. Regeneron identification of GPR75 required sequencing 640,000 exomes to detect a rare protective variant linked to obesity, a process spanning years of wet-lab validation. Similarly, HSD17B13, a metabolic target derived from genetic data, remains in development years after discovery. Chang emphasizes that AI cannot shortcut the discovery of targets absent from training data. Innovation requires deep engagement with clinical cohorts and laboratory research. The analysis concludes that AI raises the floor of drug development efficiency but does not resolve the fundamental challenge of target selection. Industry stakeholders must remain cautious, recognizing that engineering gains cannot replace the rigorous pursuit of novel biological insights required to overcome current clinical attrition rates.