Muse Spark Open: when does GPU self-hosting beat the API?

Put in your actual token volume and a quote for the entire serving configuration, not a single GPU. The default $40/hour is a hypothetical assumption, not a verified rental price.

Monthly uncached input (M tokens) Monthly provider-cached input (M tokens) Monthly output (M tokens) Full GPU configuration ($/hour) Reserved hours/month Ops/storage/other ($/month) Measured sustainable total tokens/sec

Hosted API estimate—

GPU + overhead estimate—

Volume required to break even—

Required sustained tokens/sec—

Illustrative cost model, not a GPU benchmark or provider quote. Default API rates of $1.25/M uncached input, $0.15/M cached input and $4.25/M output are public October 17, 2026 OpenRouter examples: provider price listing. Providers differ in cache eligibility, latency, throughput and reliability. Excludes bandwidth, taxes, redundancy and hardware performance constraints not entered above. At zero traffic, a break-even cannot be established.

Real serving requirements involve a roughly 530 GB NVFP4 checkpoint. Check Meta's deployment recipes before treating the result as deployable. UsageBox's task-cost tool can compare the cost per successful task after accounting for retries.