Emad Mostaque says Kimi K3 inference costs could drop 10x to 50x as infra matures

rohanpaul_ai · x · 2026-07-21

Emad Mostaque argues that Kimi K3’s current inference cost is high because the infrastructure around it is still immature, not because of a permanent technical limit.

He expects the cost to fall 10x to 50x over the next few months as specialized providers optimize kernels, routing, quantization, batching, memory, and serving.
He says Kimi K3 currently uses roughly 2x the tokens of GPT-5.6 for the same task, but that gap should narrow as the ecosystem adapts.
His broader point: U.S.-based inference companies will spend heavily optimizing Chinese open-model weights once they are available, similar to how Fireworks, Modal, and Baseten have built large businesses around inference infrastructure.

Original post →

Emad Mostaque says Kimi K3 inference costs could drop 10x to 50x as infra matures

More from Venture