A GPU host's worst moments happen while nobody is looking. A rental ends at 2 AM and the machine sits idle until morning. A price update fails and you find out days later. Notifications flip that around: the events that cost you money arrive in Discord or Telegram the moment they happen, in apps you already have open.
Why polling your dashboard does not scale
With one machine you can check the Vast.ai host view a few times a day and mostly get away with it. With three or four machines the math breaks. Rentals start and end at all hours, prices move every few minutes, and the gap between something happening and you noticing it is pure lost information. Sometimes it is lost revenue too: a machine that dropped offline on Friday evening earns nothing all weekend if you only look on Monday.
The fix is not more discipline. It is routing the handful of events that actually matter to a channel that interrupts you, and letting everything else stay in the log for later reading.
The events GPU Autopilot can send you
GPU Autopilot fires alerts for four families of events, and you choose which ones you want per event type:
| Event | What it tells you | Why you care |
|---|---|---|
| Rental started / ended | A renter took or released one of your machines | Confirms the machine is earning, and flags idle time the moment it begins |
| Price change | The agent moved your listing price | You see the market shifting in real time without opening the dashboard |
| Error | Something failed, for example a price write that did not go through | Failures caught in minutes instead of days |
| Scheduler event | A day-of-week floor override switched on or off | Verifies your weekend pricing actually engaged |
Per-event preferences matter more than they sound. Price changes on a busy machine can arrive many times a day, so plenty of hosts route only rentals and errors to their phone and leave price changes for the activity feed. Others want every move visible in a dedicated Discord channel. Both work; the point is that you decide.
Set up Discord alerts
Discord delivery uses a standard webhook, which means alerts land in any channel of any server you control. Setup takes about two minutes:
- 1In Discord, create a channel for your fleet (something like
#gpu-alerts) so alerts do not drown in chat. - 2Open that channel's settings, go to Integrations, and create a webhook. Copy the webhook URL.
- 3Paste the webhook URL into your GPU Autopilot notification settings. Alerts start flowing immediately, no restart or confirmation step needed.
One channel per concern
Set up Telegram alerts
Telegram alerts arrive as direct messages, which makes them the better choice if Discord is not part of your daily routine or you want fleet events on your phone's lock screen. You connect your Telegram account from the same notification settings, and the same per-event preferences apply: pick rentals, price changes, errors, scheduler events, or any combination.
There is no wrong answer between the two channels, and you do not have to choose. Discord suits a team or a shared operations channel; Telegram suits a single operator who wants a quiet phone that only buzzes when money starts or stops.
A night in the life of an alerted host
Here is the difference in practice. Without alerts:
- 11:40 PM: a renter's job finishes and your RTX 4090 machine goes idle. You are asleep.
- 7:30 AM: you check the dashboard, see the idle machine, and realize you missed roughly eight hours of potential rental time.
With alerts:
- 11:40 PM: a "rental ended" message lands in Telegram. You are still asleep, and that is fine.
- 11:43 PM: the agent has already repositioned the machine against the current market, since it reprices every few minutes whether you are awake or not. A "price change" alert documents the move.
- 2:10 AM: a "rental started" alert confirms the machine is earning again. In the morning you read the story in three messages instead of discovering a mystery.
The alerts did not make the repricing happen; the agent did. What the alerts buy you is verification without vigilance: you know the system is working because it tells you, and silence becomes information too.
Alerts tell you when, the log tells you why
A notification is deliberately short. When you want the full context behind an event, the market prices and the reasoning at the moment of the decision, that lives in the activity feed. Read the activity log guide to see how the two work together.
And when an error alert does fire, you want a fixed diagnostic routine instead of improvising at midnight. Keep the 10-minute offline checklist bookmarked for exactly that moment.