GPUs
Reserve whole NVIDIA GPUs for a term, launch confidential GPU instances into the reservation, or run on spot.
GPU instances get whole NVIDIA GPUs passed through to a confidential VM. GPUs are never sold on demand: you reserve them for a term of at least seven days, or run on spot capacity that no reservation holds.
B300 GPUs are offered in the latam-central-1 location (San Salvador) on TDX, with the confos-gpu image: the general-purpose image with NVIDIA's confidential-computing driver, CUDA, NCCL and Docker. You log in over SSH as root, as with confos-cpu, and there is no Kubernetes in the guest. GET /v1/images lists the confos-gpu versions whose sizes include the GPU sizes; use the newest, 2026.11.10-3 today. B300 reservations there start on 2026-10-22 or later.
GPU sizes are listed in GET /v1/sizes before they are offered. A size whose tees list is empty is not offered in any location yet, and GET /v1/reservation-offerings returns no offerings for it.
GPU sizes
| Size | GPUs | vCPUs | Memory |
|---|---|---|---|
g1-b300.1x | 1 × B300 | 8 | 64 GiB |
g1-b300.2x | 2 × B300 | 16 | 128 GiB |
g1-b300.4x | 4 × B300 | 32 | 256 GiB |
g1-b300.8x | 8 × B300 | 64 | 512 GiB |
g1-b200.1x | 1 × B200 | 8 | 64 GiB |
g1-b200.2x | 2 × B200 | 16 | 128 GiB |
g1-b200.4x | 4 × B200 | 32 | 256 GiB |
g1-b200.8x | 8 × B200 | 64 | 512 GiB |
g1m-b300.1x | 1 × B300 | 16 | 224 GiB |
g1m-b300.2x | 2 × B300 | 32 | 448 GiB |
g1m-b300.4x | 4 × B300 | 64 | 896 GiB |
g1m-b300.8x | 8 × B300 | 128 | 1792 GiB |
g1m-b200.1x | 1 × B200 | 16 | 224 GiB |
g1m-b200.2x | 2 × B200 | 32 | 448 GiB |
g1m-b200.4x | 4 × B200 | 64 | 896 GiB |
g1m-b200.8x | 8 × B200 | 128 | 1792 GiB |
g1x-b300.1x | 1 × B300 | 24 | 346 GiB |
g1x-b300.2x | 2 × B300 | 48 | 692 GiB |
g1x-b300.4x | 4 × B300 | 96 | 1384 GiB |
g1x-b300.8x | 8 × B300 | 192 | 2768 GiB |
A family gives each GPU the same vCPUs and memory at every count: g1 8 vCPUs and 64 GiB, g1m 16 vCPUs and 224 GiB, and g1x 24 vCPUs and 346 GiB. Every vCPU is a host CPU of the instance's own. Every family runs as spot at the same price for a GPU. confos-gpu 2026.11.10-3 offers g1m at one, two and four GPUs, and at eight on B300, and g1x at one, two and four; the other g1m and g1x sizes are not yet offered: an image lists a size in its sizes once it sells it.
Each instance gets its own GPUs and only those. With 8 GPUs the guest has the whole board and can attest its NVSwitch topology. With 1, 2 or 4 it shares the board's NVSwitches with other instances on the host, which the guest cannot verify.
Reserve GPUs
1. Find an offering
curl -G https://api.cloud.confidential.ai/v1/reservation-offerings \
-H "Authorization: Bearer $CC_API_KEY" -H "Api-Version: 2026-10-15" \
--data-urlencode gpuModel=b300 \
--data-urlencode gpuCount=8 \
--data-urlencode term=P30D \
--data-urlencode start=2026-11-01T00:00:00ZA reservation is of GPUs, at one price per GPU-hour for a model. Any of the model's sizes with reservation in purchaseOptions launches against it, so you split its GPUs as you like: eight B300s can run one g1x-b300.4x beside two g1m-b300.2x, and other sizes later. Each instance needs its vCPUs and memory free on the host when it starts. A host keeps two CPUs for itself, so the family that splits every CPU of the host, g1x on a B300 and g1m on a B200, runs on at most seven of its eight GPUs at once, and every reserved GPU always starts at the family below it.
gpuCount is the GPU count of one of the model's sizes (1, 2, 4 or 8) or a multiple of the largest. term is a whole number of days from P7D to P365D. The answer is up to six priced offerings, starting at 00:00 UTC on the days nearest to start that have room, each with its location, tee, start, end and price (perGpuHour and total). A price holds until the offering's expiresAt. An empty list means nothing fits.
2. Request it
curl -X POST https://api.cloud.confidential.ai/v1/reservations \
-H "Authorization: Bearer $CC_API_KEY" -H "Api-Version: 2026-10-15" \
-H "Idempotency-Key: $(uuidgen)" -H "Content-Type: application/json" \
-d '{ "offering": "ro-3hx8k2", "displayName": "November training run" }'The request returns 202 with the reservation. Its status then moves through:
| Status | Meaning | You can |
|---|---|---|
awaiting_approval | An agent asked; waiting for an owner to approve in the console (72 hours at most) | Withdraw it |
requested | Invoiced; waiting for the invoice to be paid | Pay it, or withdraw it |
confirmed | Paid for, and ours to honour | Nothing; it cannot be withdrawn or refunded |
active | The window is open: launch GPU instances | Launch and delete instances |
ended | The window closed and its instances were deleted | |
declined, withdrawn | Closed before confirmation |
Paying confirms a reservation. Once it is requested, it carries an invoice for its total:
"invoice": {
"status": "open",
"url": "https://invoice.stripe.com/…",
"dueAt": "2026-11-01T17:40:00Z"
}The invoice is emailed to your org's billing email, and you pay it at url, due 10 days after the request or at the reservation's start, whichever is first. A reservation totalling more than USD 999,999.99 is not offered; contact us for one. A reservation still unpaid at dueAt is declined and its GPUs go back. Withdraw a request before it is confirmed with DELETE /v1/reservations/{id}, which voids its invoice.
3. Launch into it
Once the reservation is active, create GPU instances against it:
{
"location": "latam-central-1",
"size": "g1-b300.8x",
"reservation": "r-5mf2k9x3qp",
"image": "confos-gpu",
"disks": [ { "name": "user-data", "type": "scratch", "sizeGiB": 200 } ],
"network": { "ingress": [ { "name": "ssh", "port": 22 } ] },
"access": { "sshKeys": [ "ssh-ed25519 AAAA… you@laptop" ] }
}The TEE comes from the reservation, and location must be the reservation's. gpusInUse on the reservation shows how many of its GPUs instances use; a create that needs more than are left returns 409 reservation_exhausted. To see what fits before you create, GET /v1/capacity?location=us-west-1&size=g1-b300.4x gives each size's maxCount with purchaseOption: reservation: how many instances one create against one of your reservations there could get now, on its unused GPUs and the host's vCPUs and memory beside them, counted as a create counts them. A size that asks for more vCPUs or memory than the host has left has maxCount 0 even when the reservation has the GPUs. Every instance in a reservation has terminationAt, the end of the window, when the platform deletes it; in its last ten minutes its reason is reservation_ending.
Approvals for agents
A reservation requested by an agent (an agent key, or an OAuth token from any client but the console) starts awaiting_approval with an approval:
"approval": {
"status": "pending",
"url": "https://console.confidential.ai/…",
"expiresAt": "2026-10-23T12:00:00Z"
}An owner of the organization opens approval.url in the console to approve or decline it. Nobody but an owner signed in to the console can approve, so no agent can spend on GPUs alone. Undecided at expiresAt, 72 hours after the request, it is declined.
Every reservation says who asked for it in requestedBy: the API key by id and name, or the userId of the person whose sign-in made the request, with agent: true when the credential was an agent's:
"requestedBy": { "apiKey": { "id": "k-7k2m9q4x1bfa", "name": "training-agent" }, "agent": true }Spot GPUs
A GPU size whose purchaseOptions include spot can run without a reservation, on GPUs no active reservation holds, at the size's spotPrice:
{
"size": "g1-b300.1x",
"purchaseOption": "spot",
"location": "latam-central-1",
"tee": "tdx",
"image": "confos-gpu",
"disks": [ { "name": "user-data", "type": "scratch", "sizeGiB": 200 } ],
"network": { "ingress": [ { "name": "ssh", "port": 22 } ] },
"access": { "sshKeys": [ "ssh-ed25519 AAAA… you@laptop" ] }
}The size's locations in GET /v1/sizes name the regions that have its GPUs, and each carries spotAvailable: whether one spot instance of the size had room there as the API answered, its GPUs free together with the host's vCPUs and memory beside them, counted as a create counts them. The size's own spotAvailable is true when any region's is. It is a look, not a hold: a create can still be refused, and false turns true as instances stop and reservations end. GET /v1/capacity says how many, as maxCount with purchaseOption: spot, with the same caveat.
When a reservation's window opens on those GPUs, or on other GPUs of the host whose vCPUs and memory the instance is using, the platform preempts the instance: its reason becomes preempting, with preemptAt 60 seconds ahead, and the guest gets an ACPI power-button press so it can shut down cleanly. At preemptAt it becomes stopped with reason preempted. Its disks, addresses and endpoints stay, but whatever was in GPU memory is lost, and every GPU is reset before another instance uses it.
Start it again on the same host once GPUs are free:
curl -X POST https://api.cloud.confidential.ai/v1/instances/$ID/start \
-H "Authorization: Bearer $CC_API_KEY" -H "Api-Version: 2026-10-15" \
-H "Idempotency-Key: $(uuidgen)"If the host has no GPUs free, the start returns 409 insufficient_capacity and the instance stays stopped.
Use the GPUs
confos-gpu needs a user-data disk, as confos-cpu does: Docker, model weights and caches belong there, not in the guest's memory. The disk is not encrypted by the guest; put secrets on the optional data disk, which the guest encrypts and which reads empty after every boot. See Disks.
Verify the instance before you log in, as in Attestation, with one addition: ask the attest endpoint for the GPUs' evidence too, and the guest binds every GPU's NVIDIA evidence to the same report data as its quote:
curl -s -X POST "http://<attest host>:<attest port>/attest" \
-H "Content-Type: application/json" \
-d "{\"nonce\": \"$nonce\", \"nvidia_gpu\": true}" -o answer.jsonThe endpoint answers 502 until the GPUs' evidence is ready, which can take several minutes after boot; retry with a new nonce. Then add the GPU checks to the verify command of step 5:
attestation-cli verify -e evidence.json \
--expected-report-data "$report_data" \
--expected-mrtd "$mrtd" --expected-rtmr1 "$rtmr1" --expected-rtmr2 "$rtmr2" \
--expected-rtmr3 "$rtmr3" \
--nvidia-gpu-user-nonce "$report_data" --nvidia-gpu-required \
--nvidia-gpu-expected-archs BLACKWELLAccept the instance only if, besides the TDX checks, claims.nvidia_gpu.overall_ok and claims.nvidia_gpu.nonce_binding_ok are true and claims.nvidia_gpu.devices lists exactly your size's GPU count, each with arch BLACKWELL and its own ueid.
Once logged in, the image's own check confirms the GPU count, that confidential computing is on and ready and, with more than one GPU, that NVLink traffic between them is encrypted:
CONFOS_EXPECTED_GPUS=8 general-purpose-gpu-check
nvidia-smiGPU metrics
Your GPUs' utilization, memory, temperature, power, clocks and errors can appear in your organization's metrics and Instances dashboard. That is off by default: turn it on for an instance with "gpuMetrics": true at create, or later with PATCH /v1/instances/{id}. Only an image that ships NVIDIA's GPU metrics exporter serves the figures, confos-gpu from its next version, on port 9400 inside the instance, so your workload can't listen on 9400 there. Software inside the instance can always read that port itself. See Metrics and Grafana for the series.