2030年までに、1兆個のパラメータを持つLLMの推論コストが90％以上削減される、ガートナーが予想 － Publickey

This page reports Gartner's prediction that the inference cost of large language models with one trillion parameters will be reduced by over 90% by 2030 compared to 2025. This reduction will be achieved through improvements in semiconductor efficiency, model design innovation, increased inference-focused silicon usage, and edge device deployment. However, Gartner warns that this cost reduction may not directly benefit enterprises due to increased token processing from widespread AI agent adoption. To maintain efficiency, companies should use smaller, domain-specific models for routine tasks and reserve large-scale models for complex processing.