From 'Model Race' to 'Settling Accounts': What Competition in the AI Industry Now Looks Like — Answers from WAIC 2026

At the WAIC 2026 conference, experts from various sectors — industry, capital, technology, and infrastructure — discussed the AI industry's transition from a race for model parameters to a pragmatic 'taking stock' phase. Key changes in 2026 include: the emergence of Claude Fable5, which sparked panic in the cybersecurity sector; a sharp rise in the cost of accelerators (rental and purchase prices increased several times over); increased complexity of B2B applications; a threefold imbalance between investments ($700+ billion) and revenues ($50 billion) in infrastructure; an 8-10x drop in the price per token alongside a surge in total token consumption; and a shift in investor attention from 'big stories' to real economic returns. Participants noted that models remain important but their role is changing: large companies must invest in models, while startups should focus on product concepts. Agent technologies (Agent + Harness) have enabled the application of probabilistic models to deterministic problems, particularly in coding, where token consumption has skyrocketed. Priority areas for the coming year include: AI infrastructure, trustworthy AI (reliability and security), world models (physical data), trusted data in narrow domains (medicine, industry), and model inference speed (up to 15,000 tokens per second on 8B models). Experts agreed that benchmarks remain useful, but closed tests are more reliable than open ones.