The briefing your
competitors already read.

The margin that justified closed-model API pricing has effectively collapsed. Qwen2.5-Max's February release — scoring 93.0 on MMLU — isn't a near-miss; it's a tie, achieved at inference costs that undercut OpenAI's o1 by a reported 97%. The practical implication for anyone building on closed APIs is not theoretical: the switching cost just became the only remaining moat, and switching costs erode.

What's less discussed is the compression dynamic underneath the headline numbers. Each open-source release is now arriving with full weight files, fine-tuning recipes, and distillation pipelines — meaning the next generation of smaller, faster, cheaper models is already baking in the oven. The capability gap isn't closing; it's inverting. The question for infrastructure teams in Q2 is not whether to evaluate open weights, but which evaluation framework to trust.

When DeepSeek-R1 landed in January 2025 at $0.55 per million output tokens, it wasn't a pricing anomaly — it was the announcement that the inference market had structurally bifurcated. By February 2026, Cerebras-hosted open models are processing at $0.07 per million tokens for GPT-4-class quality. The startup that built its margin model on a 70% gross margin at $0.50/1K tokens is now looking at a world where the commodity floor is 40× lower.

The downstream effect isn't just on AI-native startups. Enterprise software companies embedding LLMs into existing products — the "AI wrapper" cohort — are facing a strange problem: their differentiation was partly the cost of switching, and the cost of switching just fell through the floor. The companies surviving this compression are the ones who locked in on data flywheel advantages eighteen months ago. Everyone else is repricing their roadmap.