Benchmark the promise and the business · 01
Freeze the rules before the race
Model + qualityPrompt shapeConcurrencyHardware + runtime
INFERENCE FOUNDER · FIELD COURSE
Benchmark the promise and the business · 02
Count work that actually succeeds
Quality floorTTFT SLOStreaming SLOSuccessful request
INFERENCE FOUNDER · FIELD COURSE
Benchmark the promise and the business · 03
Price the successful token
GPU costAchieved utilizationIdle headroomRetries + overhead
03
INFERENCE FOUNDER · FIELD COURSE
Benchmark the promise and the business · 04
The cheapest GPU can produce the costliest request
Retail GPU rateAchieved utilizationSLO failuresSuccessful-token COGS
INFERENCE FOUNDER · FIELD COURSE
Benchmark the promise and the business · 05
A speedup is bounded by quality and regime
Self-reported claimQuality disclosureReproduce the regime
INFERENCE FOUNDER · FIELD COURSE