Meta is acquiring Manus for $2 billion.1 Alongside the announcement, Manus disclosed $100 million in ARR achieved in eight months, 147 trillion tokens processed since launch.2

Can we use those figures to explain the acquisition price?

Publicly traded software companies have gross margins of 71-72%. AI companies run lower. Gross profit per token may be a better indicator of earnings potential. Some AI companies already use gross profit as a quota metric rather than revenue.

Here’s how six AI inference companies compare.3

Company

Monthly Tokens

Valuation ($B)

Gross Margin

GP Multiple

DeepSeek

15T

3.4

85%4

20x5

Together AI

60T

3.3

45%6

24x

Manus

16.3T

2

50%7

40x

Anthropic

50T

183

55%8

67x

Groq

1T

6.99

40%10

102x

Perplexity

2.8T

20

60%11

222x

DeepSeek & Together AI trade at the lowest multiples, they resell inference. Perplexity commands the highest at 222x, it’s an application. Different layers, different monetization, different multiples. Manus sits at 40x.

On a log-log scale, gross profit per token correlates 0.71 with valuation (R² = 50.2%). Raw token volume? Just 0.47. Investors value token monetization more than overall volume.

A caveat : these figures are estimates from press reports, funding announcements, & industry analyses.

The sample is small. But the correlation suggests investors & acquirers already believe what the industry hasn’t yet named.

---

1. Meta acquires Singapore AI agent firm Manus, Investing.com, December 30, 2025 ↩︎

2. Singapore’s Manus AI hits $100m ARR in 8 months, Tech in Asia, December 2025. The 147 trillion tokens figure represents cumulative tokens since Manus launched on March 6, 2025. To estimate monthly token generation, I used a geometric growth model assuming near-zero tokens at launch with compound growth, validated against revenue trajectory ($100M ARR by month 8). This yields approximately 16.3 trillion tokens monthly by December 2025. ↩︎

3. xAI Grok excluded from this analysis. Its $230B valuation reflects a composite business including X (formerly Twitter) & the AI inference operation, making it difficult to isolate the pure-play AI economics. ↩︎

4. DeepSeek claims 85%+ gross margin due to architectural efficiency. Source: DeepSeek AI Statistics ↩︎

5. GP Multiple = Valuation ÷ Annual Gross Profit, where Annual Gross Profit = Monthly Tokens × 12 × Gross Profit per Token. This metric normalizes for both volume & margin efficiency. ↩︎

6. Together AI margin estimated based on GPU ownership trends. Source: Together AI Series B ↩︎

7. Manus margin estimated at 50%, typical for agent-based SaaS models ↩︎

8. Anthropic margin blended from 60% direct sales & cloud resale arrangements. Source: Anthropic Raises Series F ↩︎

9. Groq valuation reflects pre-Nvidia acquisition. Nvidia acquired Groq for $20B, but the inference business remains standalone. ↩︎

10. Groq margin estimated considering LPU cost advantages at current scale. Source: Groq Inference Tokenomics ↩︎

11. Perplexity 60% margin per The Information. Source: Perplexity AI Statistics ↩︎