NVIDIA details agentic inference economics: Claude Code hits 85-97% cache, 11.7x read/write ratio

_ScottCondron · x · 2026-08-23

NVIDIA's technical blog post "Full-Stack Optimizations for Agentic Inference with Dynamo" explains how to bring managed-API-grade agent inference optimization to self-hosted open-source models.

Why it matters: Coding agents now write production code at scale — Stripe's agents generate 1,300+ PRs per week, Ramp attributes 30% of merged PRs to agents, Spotify reports 650+/month. Claude Code makes hundreds of API calls per session, each carrying the full history, putting heavy pressure on KV cache.

Key numbers: After the first call writes the conversation prefix, subsequent calls hit 85-97% cache; a 4-Opus agent swarm reaches 97.2% aggregate hit rate with an 11.7x read/write ratio — a write-once-read-many pattern. Maximizing cross-worker cache reuse and keeping KV blocks warm and routable is the central optimization target.

Dynamo's approach: Making Dynamo agent-native at three layers — frontend API, router, and KV cache management — so self-run GPU clusters get prefix matching, cache placement, and eviction that managed providers control out of the box.

Related event: NVIDIA details agent-aware inference caching with up to 97% hit rate(2 posts)→

Original post →

More from coding & agent

coding & agent channel →