Buy useful output, not just GPU hours
Engineering field guide · Compute economics · Both days
What does one acceptable result really cost?
Start with Revolut’s serving results. Compare operator performance and understand how compute contracts affect access and cost.
All topic guides · Practical takeaways
Where the approaches meet
Section titled “Where the approaches meet”The useful purchase is completed work at an acceptable quality and latency. The chip model or token price alone leaves out important operating constraints.
How the approaches differ
Section titled “How the approaches differ”| Session | Approach |
|---|---|
| Revolut · Day 1 | Reports an internal self-hosted deployment and serving changes made with a fixed GPU fleet. |
| SemiAnalysis · Day 2 | Compares output from different operators using the same hardware. It separates GPU-hour inputs from useful token output. |
| SF Compute · Day 2 | Examines access to capacity through contracts and credit. Procurement can limit a workload before hardware performance does. |
Revolut · Day 1 · SemiAnalysis · Day 2 · SF Compute · Day 2
Where the evidence stops
Section titled “Where the evidence stops”Revolut’s blended token cost depends on its utilization and deployment. Its reported lead-time improvement is an internal comparison, not an isolated model experiment. Operator benchmarks and financing terms also describe particular setups.
Try this on your workload
Section titled “Try this on your workload”- Replay one representative workload against two serving options at the same concurrency.
- Set a quality rubric and latency target before comparing costs.
- Include retries, idle capacity, engineering time, and minimum commitments in the cost calculation.
Measure the result
Section titled “Measure the result”Compare cost per accepted task, latency at the 95th percentile, throughput at the target quality, and committed spending left unused.
Read the source sessions
Section titled “Read the source sessions”- Revolut · Day 1. Our session notes · Official reading edition · Source video
- SemiAnalysis · Day 2. Our session notes · Source video
- SF Compute · Day 2. Our session notes · Source video
The comparison and experiment are our editorial synthesis. They are not a reported benchmark or the speakers’ prescribed procedure.