Skip to main content

Regions

Every offer names where its machine sits. You see it on the offer card as a flag and a region code, and it follows the instance through deploy and onto the instance page.

What a region is​

A region is the machine's location. It is a fact about the hardware, not a service tier: two offers with the same GPU model give you the same GPU wherever they are.

Region codes are usually written <country>-<site>: the first two letters are the ISO 3166 country or region code — CN for China — and after the hyphen comes the site the machine is in. Each offer also carries a separate country code, taken from the first two letters of the region code; the flag and the marketplace's Region filter use that country code.

Where regions come from on Superheat CN​

FactWhat it means for you
The catalog holds only machines in Superheat CN's own fleetEvery offer on the marketplace comes from Superheat CN's fleet; it shares no inventory with any other Superheat catalog
The location is what the operator entered for the hostThe region code is exactly what the operator entered. Superheat CN does not infer a location from a machine's IP address
A machine with no location entered shows as UNKNOWNIts offers have the region code UNKNOWN and the country code XX

So go by the region code on the offer, and do not guess where a machine is from its hostname or IP address.

What a region tells you​

What it changes:

ConcernHow region affects it
Interactive latencyThe round trip between you and the machine. A distant region makes a shell feel sluggish and a JupyterLab cell slower to acknowledge, even though the work itself runs at the same speed.
Data transferMoving a dataset up or checkpoints down crosses that distance. The further away, the longer the copy.
Where your data sitsWhile the instance exists, your disk lives on that machine, in that region.

What it does not change: how fast the GPU executes your job. Once the data is on the machine, an hour of training is an hour of training wherever it runs.

The practical rule: pick a nearby region when you will be typing into a terminal or shuttling large files, and stop worrying about it for a job you launch and leave alone overnight.

Data location in plain terms​

Everything you write to the instance's disk — datasets, checkpoints, credentials you paste into a shell — is stored on that physical machine in that region until the instance is destroyed. Destroying deletes the disk.

If you have a rule about where data may be processed, the region field is the control you have over it. Read it as the operator's statement of where the hardware is, not as a compliance certification.

Filtering by region​

The marketplace's Region filter narrows by country or region, and its options show the country or region names. The list is built from the offers the catalog lists at that moment (sold-out shapes included), so it shows only the places that are on offer now rather than every place the fleet has ever covered. To narrow to one site, type the region code into the search box, which matches the GPU model, the hostname and the region.

Through the API, the region parameter on GET /v1/offers accepts either a region code or a country code, as an exact, case-sensitive match:

# every offer whose machine has country code CN
curl -s "$SUPERHEAT_API/v1/offers?region=CN&sort=price_asc" \
  -H "Authorization: Bearer $SUPERHEAT_KEY"

To see one site only, pass the full region code shown on the offer as region.

A filter is not a guarantee​

Choosing a region narrows the offers you can see. It does not reserve anything, and it does not make capacity appear.

  • If nothing in that region matches your other filters, the page is empty. The fix is to widen the filter, not to wait — there is no queue and no capacity request.
  • Availability moves. A region with four free 4× offers this morning can have none this afternoon, because other people rented them.
  • The choice applies to one deployment. Destroy that instance and deploy again, and you land wherever you pick at that moment. Instances do not migrate, and there is no setting that pins future deployments to a region.
  • Stopping an instance holds the exact machine you were on, so its region is held too. Destroying releases it.

If a specific region matters to your work, check the marketplace before you plan around it, and keep the instance stopped rather than destroyed when you intend to come back to that same machine.