Renting a GPU

Pick a box by VRAM, price, and region from one cheapest-first feed, then rent it from the console or one API call.

Every GPU on IPX lives in a single price-sorted list, no matter which cloud it came from. You filter that list down to the hardware you need, rent a listing, and we provision the box, inject your SSH key at boot, and open a reverse tunnel back to our gateway. A rental moves through a short lifecycle and, once it reaches active_connected, hands you a connection string. This page covers choosing a box, renting it both ways, polling the lifecycle, setting a spend cap, and tearing down.

Choosing a box

There is exactly one feed and it is always sorted cheapest first. You do not choose a provider. You choose hardware, and we route you to the lowest-priced box that matches.

From the API, the feed is GET /api/v1/listings. It returns boxes ordered cheapest first, excludes interruptible (spot) hardware by default, and accepts min_vram_gb, interruptible, and limit. Full details live in the listings reference.

# one price-sorted feed across every cloud, cheapest first
$ curl -s "https://gpu.ipxtechholding.com/api/v1/listings?min_vram_gb=40" \
    -H "Authorization: Bearer $IPX_TOKEN" | jq '.listings[0]'

{
  "id": 8821,
  "gpu_name": "RTX 4090",
  "vram_gb": 48,
  "num_gpus": 2,
  "region": "us-east",
  "interruptible": false,
  "price_micro_usd_per_hour": 680000,
  "price_usd_per_hour": 0.68
}

Renting from the console

The console is the fastest path. No token, no JSON, just two clicks.

1. Open Browse GPUs

   Go to Browse. You land on the same cheapest-first feed, with filters for VRAM, price, and region across the top.

2. Filter to your hardware

   Set a VRAM floor, a price ceiling, and a region if you care about one. The list re-sorts and keeps the cheapest matching boxes at the top.

3. Pick a key and a cap, then Rent

   Choose which SSH key to inject and, optionally, a hard spend cap for the box. Click Rent. Provisioning starts immediately.

4. Wait for the connection string

   The box appears on your dashboard in a provisioning state. When it reaches active_connected, the connection string is shown. Copy it and SSH in.

Renting via the API

One POST creates a rental and kicks off provisioning. Authenticate with a bearer token from Settings → API access, sent as Authorization: Bearer <token> and shown below as $IPX_TOKEN.

$ curl -s -X POST "https://gpu.ipxtechholding.com/api/v1/rentals" \
    -H "Authorization: Bearer $IPX_TOKEN" \
    -H "Content-Type: application/json" \
    -d '{
      "gpu_listing_id": 8821,
      "customer_ssh_key_id": 42,
      "max_cost_micro_usd": 5000000
    }'

On success the endpoint returns 201 Created with the new rental. It has just been created, so its state is provisioning_requested and connection_string is null until the tunnel comes up.

{
  "id": 5140,
  "state": "provisioning_requested",
  "gpu_name": "RTX 4090",
  "vram_gb": 48,
  "price_micro_usd_per_hour": 680000,
  "accumulated_cost_micro_usd": 0,
  "max_cost_micro_usd": 5000000,
  "connection_string": null,
  "created_at": "2026-06-24T17:02:11Z"
}

Response fields

The lifecycle and polling

Provisioning is asynchronous. After the create call returns, fetch the rental until it connects. There is one feed of truth per rental at GET /api/v1/rentals/:id.

A rental moves through these states:

1. provisioning_requested · we accepted the request and reserved a gateway port.
2. booting_instance · the box is starting on the underlying cloud with your key.
3. awaiting_tunnel · the box is up and dialing home to open its reverse SSH tunnel.
4. active_connected · the tunnel is up, connection_string is populated, and the per-second meter is running. This is where you SSH in.
5. terminated · the box is gone and the meter has stopped, whether you tore it down or the cap was hit.

Poll GET /api/v1/rentals/:id until state is active_connected, then read connection_string. A short fixed interval is plenty. Boxes usually connect within a minute or two.

RENTAL_ID=5140

# poll every 5s until the tunnel is up
while :; do
  RESP=$(curl -s "https://gpu.ipxtechholding.com/api/v1/rentals/$RENTAL_ID" \
    -H "Authorization: Bearer $IPX_TOKEN")
  STATE=$(echo "$RESP" | jq -r '.state')
  echo "state: $STATE"
  [ "$STATE" = "active_connected" ] && break
  sleep 5
done

# grab the connection string and SSH in
echo "$RESP" | jq -r '.connection_string'

Setting a spend cap

A spend cap is a hard ceiling per box, in micro-USD. When accumulated_cost_micro_usd reaches the cap, we reclaim the box automatically. Set it with max_cost_micro_usd at rent time, in the console or the API.

• Set it. Pass max_cost_micro_usd. A $5 ceiling is 5000000; a $20 ceiling is 20000000.
• Omit it. The cap defaults to your full prepaid balance, so the box can run until your credit runs out.
• Floor. Whatever you pass must cover at least one hour at the box's rate, and your balance must be able to cover the cap. Otherwise the rental is rejected.

Tearing down

Tear down to stop the meter. You can do it from the console or the API, and the cap will also tear a box down on its own if it is hit.

From the console

Open the box on your dashboard and click Terminate. The box moves to terminated and billing stops.

From the API

Send a DELETE to the rental. We terminate the box and return its final state.

$ curl -s -X DELETE "https://gpu.ipxtechholding.com/api/v1/rentals/5140" \
    -H "Authorization: Bearer $IPX_TOKEN" | jq '.state'
# "terminated"

The response is the same rental object, now with state set to terminated and a final accumulated_cost_micro_usd. Teardown is idempotent in effect: once a box is gone, it stays gone.

Next steps