AI cannot replace SaaS economics under current cost structures
Abstract
Artificial intelligence is widely assumed to scale like software. This assumption underpins current valuations, capital allocation, and product strategy across the technology sector. This paper challenges that assumption using observed financial data.
AI introduces a usage-based cost structure at the unit level, fundamentally different from the fixed-cost model of traditional SaaS. Analysis of public cloud companies shows cost of revenue growing faster than revenue following AI deployment, resulting in measurable margin compression.
This paper defines and applies the AI Cost–Revenue Divergence (ACRD) framework to identify early signals of this structural shift. The findings suggest that, under current cost conditions, AI workloads do not exhibit SaaS-like operating leverage and may instead degrade it.
1. The Claim
AI does not currently scale like software.
Traditional SaaS economics depend on a simple principle:
build once, sell many times, and margins expand as revenue grows.
AI breaks this assumption.
Each unit of AI output requires incremental compute (inference), introducing a variable cost that scales with usage, not deployment.
The constraint is mechanical:
If cost per inference does not decline faster than revenue per inference increases, margins compress rather than expand.
2. Framework: ACRD (AI Cost–Revenue Divergence)
ACRD measures the relationship between revenue growth and cost of revenue growth.
Definition:
ACRD Spread (bps) = Cost Growth – Revenue Growth
Interpretation:
- Positive spread → cost growing faster than revenue (margin compression signal)
- Negative spread → revenue scaling faster than cost (operating leverage)
Classification:
- Emerging → early divergence
- Active → clear margin pressure
- Structural → persistent divergence across periods
This framework is descriptive, not predictive. It identifies signals already present in reported financials.
3. Evidence from Public Cloud Companies
Cloudflare (Form 10-Q)
- Revenue growth: ~32%
- Cost of revenue growth: materially higher
- Spread: +2,179 bps
- Signal: Active divergence
Interpretation:
AI-related workloads are increasing cost of revenue faster than revenue, indicating early margin pressure within a usage-based delivery model.
C3.ai (Form 10-K)
- Revenue growth: modest
- Cost of revenue growth: elevated
- Spread: positive (divergent)
- Signal: Active divergence
Interpretation:
Despite positioning as an AI-native company, cost structure expansion is outpacing revenue scaling, consistent with inference-driven cost pressure.
Pattern Across Cohort
Across a broader cohort of AI-exposed public companies:
- Cost of revenue is accelerating post-AI rollout
- Revenue expansion lags or scales linearly
- Gross margin expansion stalls or reverses
This is not isolated. It is repeatable.
An extended cohort of 72 AI-exposed companies has been analyzed using the ACRD framework (see Appendix B).
4. Structural Implication
SaaS and AI operate under fundamentally different cost mechanics:
| Model | Cost Structure | Scaling Behavior |
|---|---|---|
| SaaS | Fixed (build once) | Margins expand with scale |
| AI | Variable (per inference) | Margins compress unless cost declines |
AI does not eliminate cost — it reintroduces it at the unit level.
This creates a new operating constraint:
- More usage → higher cost
- More intelligence → higher compute intensity
- More scale → not necessarily more margin
5. Conclusion
AI is not software in the economic sense that SaaS is software.
Under current cost structures:
- AI workloads introduce variable, usage-based costs
- These costs are already observable in financial statements
- Margin compression is not theoretical — it is present
The implication is direct:
AI cannot replace SaaS economics under current cost structures.
The companies that succeed in AI will not be those with the best models, but those that control the cost of delivering intelligence at scale.