Memory drives most of your inference bill, a cost-optimisation layer starts forming around developer tools, and agentic models reach consumer silicon.