Self-hosted AI agents: which VPS region is closest to the LLM APIs
As of November 07, 2026, the median LLM API answered its first byte in 222 ms from Europe; 178 ms from US Central; 215 ms from Tokyo; 338 ms from São Paulo. These are measured from rented VPSes of the kind you would run OpenClaw, Hermes Agent or n8n on — Hetzner in Nuremberg, Vultr in Chicago, Vultr in Tokyo, Vultr in São Paulo — against 44 APIs every five minutes. US Central is closest to the most of the 23 APIs built into these agents (11 answer fastest from there), but the API you use decides.
How far is each region from the LLM APIs?
Median time to first byte over the last 24 hours across every API this site probes, from each probe VPS.
| Probe region | VPS | Median API | Under 100 ms | Fastest API |
|---|---|---|---|---|
| Europe | Hetzner · Nuremberg, Germany | 222 ms | 7 of 44 | Fireworks 16 ms |
| US Central | Vultr · Chicago, US | 178 ms | 14 of 44 | Fireworks 29 ms |
| Tokyo | Vultr · Tokyo, Japan | 215 ms | 6 of 44 | Fireworks 19 ms |
| São Paulo | Vultr · São Paulo, Brazil | 338 ms | 1 of 44 | OpenRouter 62 ms |
Which region is closest to the API your agent uses?
LLM APIs built into OpenClaw, Hermes Agent or n8n, measured at the host the agent calls. Bold marks each API’s fastest region.
| LLM API | OpenClaw | Hermes Agent | n8n | Europe | US Central | Tokyo | São Paulo |
|---|---|---|---|---|---|---|---|
| OpenRouter | ✓ | ✓ | ✓ | 98 ms | 59 ms | 55 ms | 62 ms |
| Fireworks | ✓ | ✓ | — | 16 ms | 29 ms | 19 ms | 254 ms |
| Google Gemini | ✓ | ✓ | ✓ | 90 ms | 41 ms | 63 ms | 137 ms |
| Baseten | ✓ | — | — | 125 ms | 47 ms | 160 ms | 154 ms |
| Cerebras | ✓ | — | — | 195 ms | 74 ms | 203 ms | 190 ms |
| Anthropic | ✓ | ✓ | ✓ | 197 ms | 95 ms | 226 ms | 194 ms |
| NovitaAI | ✓ | ✓ | — | 231 ms | 95 ms | 147 ms | 247 ms |
| xAI | ✓ | ✓ | ✓ | 221 ms | 111 ms | 165 ms | 224 ms |
| Groq | ✓ | — | ✓ | 213 ms | 127 ms | 161 ms | 233 ms |
| Together | ✓ | — | — | 232 ms | 128 ms | 162 ms | 263 ms |
| OpenAI | ✓ | ✓ | ✓ | 215 ms | 139 ms | 208 ms | 245 ms |
| Mistral | ✓ | — | ✓ | 117 ms | 196 ms | 322 ms | 263 ms |
| Moonshot (Kimi) | ✓ | ✓ | ✓ | 308 ms | 266 ms | 129 ms | 342 ms |
| Z.AI (GLM) | ✓ | ✓ | — | 184 ms | 252 ms | 93 ms | 541 ms |
| Qianfan (Baidu) | ✓ | — | — | 286 ms | 272 ms | 138 ms | 412 ms |
| DeepSeek | ✓ | ✓ | ✓ | 290 ms | 318 ms | 182 ms | 432 ms |
| MiniMax | ✓ | ✓ | ✓ | 311 ms | 343 ms | 223 ms | 478 ms |
| Upstage | — | ✓ | — | 586 ms | 355 ms | 58 ms | 616 ms |
| DeepInfra | ✓ | — | — | 508 ms | 157 ms | 446 ms | 533 ms |
| Qwen Cloud | ✓ | ✓ | ✓ | 359 ms | 415 ms | 167 ms | 706 ms |
| StepFun | ✓ | ✓ | — | 429 ms | 392 ms | 325 ms | 672 ms |
| Featherless AI | ✓ | — | — | 589 ms | 284 ms | 426 ms | 685 ms |
| Volcano Engine (Doubao) | ✓ | — | — | 350 ms | 457 ms | 495 ms | 705 ms |
How much does the region add to each agent step?
On an open connection every request waits at least one round trip to the API host before the model starts; the time to first byte we measure is that round trip plus the host’s answer. Across the 23 APIs built into these agents and measured from all regions, the median gap between an API’s fastest and slowest region is 213 ms. For Upstage it is 558 ms: 58 ms from Tokyo against 616 ms from São Paulo. Over a task that makes 100 sequential model calls, the slower region costs about 21 s of pure network time for a typical API and 56 s for Upstage. That is the floor under each call, not the whole wait: where we time real model requests, the first token took from 352 ms (Groq, openai/gpt-oss-120b, from Europe) to 1895 ms (Google Gemini, gemini-flash-lite-latest, from Tokyo).
Which region should the agent’s VPS be in?
The one where the API it calls answers fastest. By the last 24 hours of probes, the APIs built into these agents are closest to:
- US Central (11): Google Gemini, Baseten, Cerebras, Anthropic, NovitaAI, xAI, Groq, Together, OpenAI, DeepInfra, Featherless AI
- Tokyo (9): OpenRouter, Moonshot (Kimi), Z.AI (GLM), Qianfan (Baidu), DeepSeek, MiniMax, Upstage, Qwen Cloud, StepFun
- Europe (3): Fireworks, Mistral, Volcano Engine (Doubao)
Where do the probes run?
On rented VPSes, the same kind of machine you would run an agent on. OpenClaw’s own FAQ suggests a VPS when the agent has to stay up around the clock (source).
| Probe region | Hosting | Location |
|---|---|---|
| Europe | Hetzner | Nuremberg, Germany |
| US Central | Vultr | Chicago, US |
| Tokyo | Vultr | Tokyo, Japan |
| São Paulo | Vultr | São Paulo, Brazil |
Your own VPS will not match these numbers exactly: another data center means another network path and sometimes another DNS answer. What carries over is the comparison between regions. Time to first byte also says nothing about model speed once generation starts — see time to first token and the methodology.
Which LLM API should the agent use?
- Best LLM API for OpenClaw — 22 built-in providers measured at the host OpenClaw calls
- Best LLM API for Hermes Agent — 14 built-in providers measured at the host Hermes Agent calls
- Best LLM API for n8n — 11 built-in providers measured at the host n8n calls
Data: rankings.json · data & licence · fastest AI API by region