Skip to lesson
Exit
LLM Inference Fundamentals for the Routing Engineer1 / 3

2 min lesson

GPU economics and provider model

Use "GPU cost is mostly about keeping expensive silicon busy" to describe the practical rule.

Step 1 of 3

Everything in the previous four sections eventually becomes a number on a bill or a capacity plan. The routing engineer who can do this math is the one who earns the gray-area decisions.

GPU cost is mostly about keeping expensive silicon busy. Utilization is driven by batch size and how well the KV cache fits in memory: idle GPUs and tiny batches burn rent for nothing, while a well-packed batch spreads the fixed cost of loading weights across many requests. The mechanics from earlier sections - paged KV, continuous batching, prefix-cache hits - are the same levers that move utilization.

Learn more

Full explanation

From Silicon to the Dollar

FROM SILICON TO THE DOLLAR

Interactive diagram. Tab through its regions; each focused region shows its detail in the panel below.

diagram: stack

Each layer rests on the one below: cost is set at the bottom, then earned back by how well the layers above pack it.

The cost vocabulary you'll be tested on
Per-token pricing
Separate input (prefill) and output (decode) rates; output is usually pricier
Rate limits
RPM (requests/min) and TPM (tokens/min) caps per key/account
Reserved vs on-demand
Committed/provisioned capacity (cheaper, fixed) vs on-demand (flexible, pricier, throttled)
GPU utilization
Driven by batch size and KV-cache fit - the core efficiency metric for self-hosting
Cache hit rate
Fraction of prefix work served from cache - a first-class cost metric, not a footnote