US-East vs US-West LLM Latency & TTFT Benchmarks (2026)
Empirical latency comparison between AWS us-east-1 (N. Virginia) and us-west-2 (Oregon) across premier LLM providers. Discover how edge routing eliminates transatlantic and cross-country network jitter for interactive software.
Live Cross-Country Time-To-First-Token (TTFT)
| Model | Provider | US-East TTFT | US-West TTFT | US-East p95 | US-West p95 |
|---|---|---|---|---|---|
| OpenAI GPT-4o mini | OpenAI Direct | 210ms | 275ms | 840ms | 980ms |
| Claude 3.7 Sonnet | Anthropic AWS Bedrock | 290ms | 340ms | 1180ms | 1310ms |
| Llama 3.3 70B | Groq Cloud | 145ms | 92ms | 420ms | 340ms |
| Gemini 2.5 Flash | Google AI Studio | 195ms | 185ms | 780ms | 720ms |
| DeepSeek V3 | Together AI | 360ms | 310ms | 1450ms | 1390ms |
Key Architectural Takeaways for US Engineering Teams
- Groq Dominates Streaming Speed: Groq's custom LPU clusters in San Jose (US-West) deliver sub-100ms TTFT, making them ideal for speculative decoding and quick classification steps.
- N. Virginia (us-east-1) Network Clamping: OpenAI and Anthropic direct clusters show 40–80ms lower latency from US-East clients due to physical proximity to their primary inference clusters.
- Edge Semantic Caching Advantage: Routing through SentioGate caches repeated system prompts and test fixtures at Cloudflare edge PoPs (300+ cities), serving sub-20ms responses before reaching provider clusters.
Frequently Asked Questions: US LLM Latency & Speed
Why is US-East (N. Virginia) latency lower than US-West for OpenAI and Claude?
OpenAI and Anthropic host their primary cloud inference clusters in AWS us-east-1 and Azure East US. Requests originating from US-East clients experience 40-80ms lower network transit time compared to cross-country transit from California or Oregon.
How does Groq achieve sub-100ms TTFT in US-West?
Groq operates dedicated Language Processing Unit (LPU) data centers in San Jose, California, and Dallas, Texas. Their SRAM architecture delivers deterministic token generation speeds exceeding 300 tokens per second for Llama 3.3.
How does edge caching reduce TTFT across US data centers?
Autonomous agents frequently repeat system prompts, project schemas, and tool definitions. SentioGate's edge semantic cache serves verified responses directly from Cloudflare edge nodes in under 20ms, bypassing upstream model computation entirely.