What this site runs on, and where to run your own AI agent
Every figure on llmlatency.dev comes from servers we rent and code we run. This page names them,
and adds a few services worth knowing about if you run LLM-backed software yourself. Some buttons are
partner links: they are marked, they go through https://llmlatency.dev/go/<name> so we can
see whether anyone uses them, and they never sit inside a ranking.
What the measurements run on
Vultr
Three of the four probe nodes behind every number on this site are Vultr cloud instances. Every five minutes each one opens a real connection to each provider’s API and records the answer.
Chicago, Tokyo and São Paulo regions; the cheapest instance size is enough for a probe.
Hetzner
The Europe probe, the measurement database and the build that produces these pages run on one Hetzner cloud server in Nuremberg.
Hetzner closed its referral programme in 2026, so this link earns nothing. It is here because the site runs on it.
Cloudflare
Cloudflare Pages serves the static site; a Cloudflare Worker serves the redirects and counts partner clicks without cookies.
No referral programme for Pages or Workers.
Where should I host an AI agent, monitor an API or rent a GPU?
Readers of the outage log and the agent hosting pages ask the same three questions: how do I watch my own API, where do I run the agent, and what if I want the model itself. One honest answer each. Where this site does not use a product, the card says so.
Hyperping
Watch your own API, webhook or status page the way this site watches providers: checks from several regions, alerts and a public status page.
This site does not use Hyperping; its probes are our own code on our own servers.
Railway
Run OpenClaw, Hermes Agent, n8n or any container on a managed platform instead of patching a VPS. A new account starts with $20 in credits.
This site does not run on Railway; the probes are plain Hetzner and Vultr VPSes.
Vast.ai
Rent a GPU on a marketplace when you want to run Llama, Qwen or DeepSeek weights yourself rather than call a hosted API.
This site measures hosted APIs only and does not use Vast.ai.
Novita AI
Novita is one of the inference APIs measured here, by the same probes and the same code as every other. Its provider page shows the figures.
The only measured provider whose programme we have joined. The link never appears in a ranking.
How partner links are kept apart from the data
Partner blocks never appear inside a ranking, table or comparison; no measured provider can buy a position, a badge or a softer number; and the full list of partners is public on this page and on the methodology page. Programmes that pay only through channels we cannot receive were left out rather than linked for show. Checked on November 07, 2026; if a programme closes, its card goes with it.
Report a mistake: corrections · about this site