While the consensus narrative for the AI boom focuses on centralized hyperscalers (AWS, Azure, GCP), I argue that Cloudflare (NET) has constructed a strategically superior substrate for AI application inference and governance. With a globally distributed footprint reaching 95% of the Internet population within 50ms, Cloudflare is transitioning from a “CDN” to a mandatory AI application runtime. This shift is driven by three architectural moats: distributed edge inference, cross-provider AI governance via AI Gateway, and the structural destruction of the “egress tax” through R2 storage.
1. The Infrastructure Thesis: Decentralizing Inference
The AI deployment cycle is shifting from training (centralized) to inference (distributed). Unlike the massive, region-centric “power plants” of the hyperscalers, Cloudflare’s infrastructure behaves like an interactive grid.
Proximity as a Performance Primitive
The current AI stack often suffers from the “Distance to Inference” bottleneck. High-throughput agents and interactive voice LLMs cannot tolerate the round-trip latency of reaching a central USD-East-1 data center from a global user base.
Cloudflare’s Workers AI allows for “one API call” inference across 200+ cities globally. By embedding GPUs at the edge, Cloudflare reduces perceived latency for interactive agents, turning geographic proximity into a non-obvious performance moat.
2. Product Deep Dive: The AI Control Plane
Cloudflare’s AI stack is not just a collection of features; it is an inline control plane that secures and optimizes model traffic.
AI Gateway: The Abstraction Layer
As enterprises adopt multi-model strategies (OpenAI, Anthropic, Meta), they face provider lock-in and observability gaps. Cloudflare’s AI Gateway acts as a neutral proxy, adding:
- Unified Interface: An OpenAI-compatible endpoint that can route across providers.
- Observability: Centralized logging, caching, and rate limiting.
- Resilience: Intelligent retries and model fallback mechanisms.
R2 & The Egress Arbitrage
Data gravity is the primary barrier to AI switching. Hyperscalers utilize egress fees—effectively a “data exit tax”—to prevent workloads from leaving their ecosystem. Cloudflare’s R2 object storage destroys this model with $0 egress fees. This creates a strategic wedge: once an enterprise stores its AI embeddings in R2, it can route that data to whichever model provider offers the best price-performance, effectively turning Cloudflare into the “Switzerland” of the AI stack.
3. Quantitative Analysis: Growth and Unit Economics
Cloudflare’s fundamentals reflect a high-velocity growth engine fueled by AI-adjacent demand.
Financial Performance Indicators
| Metric (USD) | FY2024 | FY2025 | YoY Change |
|---|---|---|---|
| Revenue | $1.670B | $2.168B | +30% |
| Gross Margin (GAAP) | 77.3% | 74.5% | -280bps |
| Total RPO | — | — | +48% |
Note: The decline in gross margin is consistent with the capital-intensive deployment of GPUs across the global edge, which Cloudflare expects to monetize through higher-value AI compute services.
Developer Velocity
As of late 2025, Cloudflare disclosed over 4.5 million active developers. In the AI era, software is increasingly built using serverless primitives and “fast-start” templates. Cloudflare’s developer retention is a leading indicator of future enterprise capture, as teams move from experimentation to production on the runtime they used for prototyping.
4. Risks & Competitive Headwinds
Despite structural advantages, Cloudflare faces critical execution risks:
- Valuation Gravity: Trading at a high P/S multiple, any deceleration in Net Retention Rate (NRR) or a miss in RPO targets could trigger significant multiple compression.
- Specialized Competition: While Cloudflare excels at general inference, specialized AI infrastructure providers may outcompete on raw GPU throughput for massive non-interactive batches.
- Regulatory Complexity: The “Data Localization Suite” is a powerful tool, but managing residency across 120+ jurisdictions remains a persistent operational burden.
5. Conclusion
Cloudflare is a structural “AI Winner” not because it builds the models, but because it owns the network through which they are called. For the institutional investor, the thesis is clear: the hyperscalers own the cloud’s foundation, but Cloudflare is building the cloud’s connectivity layer—a position that is increasingly indispensable in an agentic, data-sensitive world.
Technical Appendix: Reference Diagram
graph TD
User[Global Interface / Agents] --> Network[Cloudflare Edge Network]
Network --> AI_Gateway[AI Gateway: Governance & Routing]
AI_Gateway --> OpenAI[OpenAI]
AI_Gateway --> Anthropic[Anthropic]
AI_Gateway --> Local[Workers AI: Local Edge Inference]
Local --> R2[R2 Storage: Zero Egress Data]