GPU on / off

The GPU is stopped when nobody is using it, because it costs about ₹107 an hour. Turning it on is a button on the home page. Everything comes back exactly as it was.

Turning it on

Open latency.cam
Press Turn on
Wait about 3 minutes

Or from a script:

curl -X POST https://latency.cam/api/gpu \
  -H "Content-Type: application/json" \
  -d '{"action":"start"}'

# then poll until readyInSeconds is 0
curl https://latency.cam/api/status

The three-minute wait is real

press ON
  ~40s   instance boots, SSH available
  ~45s   services start
  1-3m   models load into GPU memory
ready

"running" does not mean readyFor a minute or two after boot, /api/status reports state: "running" while readyInSeconds is still above zero. Requests during that window fail. Always wait for readyInSeconds to reach 0, not just for the state to change.

The delay is loading model weights into GPU memory, which is empty after every stop. It cannot be avoided.

What survives a stop

Thing	Survives?
Models on disk	✓ yes
Code, config, API keys	✓ yes
`api.latency.cam` domain and IP	✓ yes — Elastic IP, never changes
TLS certificate	✓ yes
Models in GPU memory	✗ reload on start
Anything in `/tmp`	✗ cleared

So yes — off then on gives you back an identical box at the same address. You can hardcode https://api.latency.cam permanently.

Auto-stop

The box shuts itself down after 2 hours idle. This is not optional and you should not disable it. A manual switch will eventually get left on, and the arithmetic is unforgiving:

Left running	Cost
A working day, 8 hours	~₹860
Overnight, 16 hours	~₹1,700
Forgotten for a month	~₹77,000

Auto-stop turns that last row into about ₹200.

Cost while stopped

Item	Monthly
Disk (200 GB gp3)	~₹1,400
Elastic IP	~₹320
Vercel Pro (docs, always up)	~₹1,760
Baseline	~₹3,500

"Off" is not free — that is the price of everything staying exactly where it was. The GPU itself adds ₹107/hour only when running.

When starting fails

InsufficientInstanceCapacity is a real possibilityStopping an instance returns the hardware to AWS. Starting requires AWS to find you a GPU again, and GPU capacity in us-east-1 is genuinely tight — we hit zero capacity across all five availability zones while setting this up.

If you get that error, wait a few minutes and retry. If it persists, the box needs rebuilding on a fallback instance type:

g5.2xlarge    A10G 24GB   (current)
g6.2xlarge    L4   24GB
g4dn.2xlarge  T4   16GB    usually available, but weaker

If you land on a T4 you get 16 GB instead of 24 GB and things will feel slower. The status panel shows which GPU you actually got, so check it before reporting a performance problem.

Region note

The box currently runs in us-east-1 (Virginia), not Mumbai, because this AWS account has zero GPU quota in ap-south-1. For development that is fine. For production it is not — Virginia adds roughly 200 ms each way to an Indian phone call, and it breaks any claim about data staying in India. A quota request for ap-south-1 should be in flight.