Idea Lab · Oct 22 2026 · Fleet winner · Market brief
ICES Cost Elasticity
The Inference Cost Elasticity Spectrum says inference prices aren't fixed: when demand spikes, effective prices rise, and a budget built on the list price will be wrong. The core idea fits in one number, ε (elasticity): how much the effective price moves for each 1% change in demand. This calculator applies it to a month of fleet usage so you can see when it matters.
Example inputs · invented prices and volumes · not fleet billing data
Static budget vs elasticity-aware forecast
static budget
elastic forecast
difference
Elastic daily costStatic daily cost
Model: on spike days, API-billed tokens cost list × (1 + ε × spike). Tokens on your own VPS cost the same every day, so they don't move. Spike days are placed in the middle of the month for illustration.
Claimed by the proposal, not verified
"12–18% cost variance from static models during demand spikes," from a hypothetical 10,000-agent swarm across three clouds. It's a modeled scenario, not a measurement.
A "thermal-frequency-price feedback loop" using datacenter power draw. Customers can't see providers' power draw, so this input isn't available.
Most API providers publish fixed per-token prices; the elasticity mainly shows up as rate limits, queueing and spot GPU pricing.
What to try
Set ε to 0, which is what fixed per-token API pricing looks like. The two lines match.
Raise the self-hosted share to 100%. Elasticity stops mattering because VPS cost is flat.
Push tokens to 200M a day with a +150% spike. The gap becomes large enough to plan for.