Sample results dashboard
A sample view of benchmark evidence: what the AI traffic cost, how often it was right, where each request went, which service targets held, and how the system behaved during failure drills and load tests.
Where requests went
What a correct answer costs
The mix sits between the two extremes
Hosted only and own GPU only are the same task run entirely on one option.
Service targets by traffic class
| Service level | Correct | Target | First response, slowest 5% | Target | Errors | Status |
|---|
Scaling under load
Failure drill: GPU server killed
The model server was deleted mid-traffic with no warning. Requests that had to stay private failed during the outage instead of being sent to the hosted API. That is the designed trade-off: an outage, not a data leak.
Failure drill: hosted provider failing (simulated)
For the whole run the hosted route was pointed at a fault simulator that failed every third call with a rate limit, a server error or a 35 second timeout. Its other answers were canned, so this run says nothing about the hosted model's accuracy. The router sent each failed call to the GPU instead, and no user got an error. The cost showed up as waiting: calls that hit the timeout took about 30 seconds before switching over.
Want this for your own system? The five-day assessment gives you a cost model, measured comparisons and a written recommendation.