What you share on a machine
A slice is not always a whole machine. When you rent a 1× offer on a two-GPU host, the other card can be rented by someone else at the same time — so it is worth knowing exactly where the line falls.
Yours alone
| Resource | How it is guaranteed |
|---|---|
| GPUs | A GPU is assigned to one instance at a time and is never split between renters. The container is given its specific cards by hardware ID, not by index, so it cannot see or address the others. |
| CPU | Your instance is capped at the cores its offer includes, enforced by the kernel rather than by convention. |
| RAM | Capped the same way. A neighbour that exhausts its own memory is killed by the kernel; your instance is unaffected. |
| Your filesystem | Another renter cannot read or write your instance's disk. |
| Your processes | Separate containers with no shared process namespace. Nothing you run is visible to a co-tenant, and nothing they run is visible to you. |
| Your private network | Your instance gets its own container network. A co-tenant on the same machine has no route to it — they cannot reach a port you did not publish, and a port you did publish is still reached the same way you reach it: through the machine's address or the connection URL, with whatever authentication the service asks for. |
Shared with the machine
| Resource | What that means for you |
|---|---|
| Disk capacity | The Storage figure on an offer card is the machine's total disk, which is the ceiling on what you can attach — not a private allocation. You pick your own size at deploy time and pay for it per GB per hour. The size you pick is committed to you: we will not sell a machine more disk than it has, so a deploy can be refused with MACHINE_DISK_EXHAUSTED when the machine is already spoken for. |
| Network bandwidth | The host's uplink is shared. The figures on the card are the machine's link speed, not a reserved share of it. This is throughput only — sharing the uplink does not put anyone on your network. |
| The host's uptime | If the machine goes offline, every instance on it is affected. This is what the host's reliability percentage measures, and why it is on the card. |
Your instance's disk is private — nobody else can read or write it — and the capacity is accounted for: the machine cannot be sold more disk than it physically has, so a co-tenant cannot be sold space that was already promised to you. That accounting happens at purchase.
Whether it is also enforced while running depends on the machine. An offer card says which:
- · guaranteed — the host's filesystem supports per-instance quotas, so a co-tenant writing without limit hits their own ceiling and stops. Your space is genuinely yours.
- · shared — the accounting still holds, but there is no hard cap inside the running machine, so a co-tenant writing far past what they bought can still fill the volume and leave you unable to write.
Most machines are still shared while hosts migrate their filesystems. Either way,
checkpoint to somewhere durable rather than relying on the instance's disk being there.
How to tell whether you are alone on a box
The Available figure on an offer card tells you the shape of the machine, not who is on
it. A card reading 1 of 2 means the machine can serve two of that shape and one is already
taken — so if you rent it, you will have a co-tenant.
A card reading 2 of 2 means the machine is idle right now. It does not mean it will stay
that way: the second card can be rented by someone else a minute later.
If you need a machine to yourself, rent the largest shape it publishes. Taking the 2× on a two-GPU host leaves nothing else to sell, so nobody else can land on it.