Engagement
Before work starts, we agree one metric, measure the baseline, and set the target and the quality checks.
Who we work with
We work with teams that serve their own models on their own or cloud GPUs, where GPU spend is part of cost of goods sold. We support H100, B200, RTX PRO 6000 and MI300X to MI355X.
Metric
We agree one metric with you before work starts, such as time to first token, p90 latency, requests per GPU or cost per token.
Baseline
We measure your current stack on that metric, on your hardware and your traffic, before we change anything. That measurement is the baseline.
Target
We agree a target for the metric, stated against the baseline, for example 2× the baseline throughput. The target counts as reached when the optimized deployment meets it on the same hardware and traffic as the baseline, with every quality check passing.
Quality checks
We agree the checks your outputs must pass at the same time as the target. A change that fails any of them does not ship.
Payment
If the deployment does not reach the agreed target, you do not pay.
What we need from you
We need access to the deployment or a replica on the same hardware, and a representative sample of its traffic. We train drafters on that sample.
Bring your model, hardware, traffic and target metric to a call. We will tell you whether the target is in reach.