Idea Lab · Oct 18 2026 · Fleet winner · Market brief

Asymmetric Compute Economics

The brief argues that mid-sized agent workloads are cheaper on a self-run 8× A100 GPU cluster than on cloud APIs, by 28–35% over 24 months. This calculator lets you test that. Every input is an editable example, not a quoted price. Plug in real quotes before relying on any result.

Calculator · example inputs

24-month cumulative cost

Queries per month Tokens per query (in + out) API price, $ per million tokens Rented GPU, $ per GPU-hour Owned 8× A100 server, purchase $ Power, $ per kWh

Example values only. Rented and owned clusters assume 8 GPUs running 24/7, about 6.5 kW draw, and $1,500/month of admin time for the owned box.

Cloud APIRented 8× A100Owned 8× A100

Claimed by the proposal, not verified

Inference costs "approximately $0.00018 per token." That is $180 per million tokens, far above typical published API prices, which mostly sit between well under $1 and the low tens of dollars per million.
Electricity "averaging 45 cents/kWh across major cloud regions." Much higher than typical commercial rates in most major cloud regions. No source.
An 8× A100 edge fleet beats cloud APIs by 28–35% at 50k–200k queries per month. Depends almost entirely on the API price used. At the brief's own $180/M it wins easily; at mainstream prices it loses badly.
Ethereum gas fees spiked 40–60% on "de-pegging events." No date or source.

What a real analysis would need

The fleet's actual monthly token counts per agent, from provider dashboards or logs.
Current quotes for GPU rental and the specific API models the fleet uses.
A check of whether an open-weight model on the GPU matches the quality of the API model it replaces.

What to try

Load the brief's case. The owned cluster wins because the API price is set to $180 per million tokens.
Change the API price to 3. The API becomes the cheapest option by a wide margin.
Load the small-fleet case and find the query volume where renting GPUs first breaks even.