AI Inference Costs Drop 90% Since 2023: Democratization Accelerates

A new analysis by Andreessen Horowitz documents the dramatic cost collapse in AI inference: GPT-4 class intelligence that cost $60 per million tokens in 2023 now costs under $3/million tokens, with continued decline expected. The drivers: NVIDIA Blackwell hardware efficiency gains, improved model architectures (mixture-of-experts, smaller context caches), quantization improvements, and intense provider competition. The practical implication: AI features that were economically viable only for high-value use cases in 2023 are now viable for nearly any software application. This is expected to trigger a second wave of AI product development targeting cost-sensitive use cases.