Gross Profit per Token: A New Valuation Metric for AI Inference Companies
Tomasz Tunguz analyzes Meta's $2.5B acquisition of Manus through the lens of gross profit per token, arguing this metric better predicts AI company valuations than raw token volume. By comparing six AI inference companies, he demonstrates that gross profit per token correlates 0.70 with valuation, suggesting investors prioritize monetization efficiency over scale.
Metrics in this report
0.70correlation coefficient
r
Six AI inference companies
0.47correlation coefficient
r
Raw token volume (lower explanatory power)
67x
calculated
Gross profit multiple
20x
calculated
Gross profit multiple
102x
calculated
Gross profit multiple
50x
calculated
Gross profit multiple
222x
calculated
Gross profit multiple (application layer)
24x
calculated
Gross profit multiple
55%
blended
Direct sales and cloud resale
85%
claimed
AI inference via architectural efficiency
40%
estimated
LPU-based inference
50%
estimated
Agent-based SaaS model
60%
reported
Application-layer search
45%
estimated
GPU-based inference reseller
71-72%
range
Publicly traded software companies
100$M
actual
Achieved in 8 months
147T
actual
Since launch March 6, 2025
16.3T
estimated
December 2025
2.5$B
acquisition price
Meta acquisition
48.5%
R²
Variance explained