AI Economics
Model capability gets the attention; unit economics decide whether a feature survives contact with production. These pages cover how inference is priced, why utilisation dominates self-hosted cost, and how to work out a break-even with your own numbers rather than someone else's.
Cost and capacity
Prices are cited to the vendor's published pricing page and dated, because they change.
Why LLM inference is a recurring per-request cost, how input and output tokens are priced separately, and where cached and batch tiers actually apply.
Hosted API against self-hosted open weights - the break-even formula, why utilisation and latency SLOs decide it, and costs that never appear per token.
Context is a per-request budget, not a capability. Attention cost scaling, long-context price tiers, prompt caching, and why chat history compounds.
Utilisation dominates self-hosted inference cost. Continuous batching, KV cache memory limits, why a model that fits can still be throughput-bound.
Retrieval adds per-request cost and infrastructure; fine-tuning adds training cost and a versioning burden. Query volume and knowledge churn decide it.
Deciding what enters the context window and what stays out: budgeting the prompt, compaction, tool-result handling, and retrieval precision over recall.
A high-cost delivery model where engineers embed with one customer. Why AI deployments generate demand for it, and when the economics do not pay back.
Every figure on these pages links to the vendor page it came from and states when it was observed. Where a number could not be sourced - several providers render pricing in the browser, so it cannot be read reliably - the economics are given as a relationship or a formula you can apply to your own measurements instead.