GPU Machines

Varity Team Core Contributors Updated August 2026

GPU on Varity is three separate catalogs, not one. They rent different things and use different endpoints. Containers also return a different profile schema.

Execution class	Profiles	You get	Created via	SSH
container	9	One container on one GPU	POST /api/deployments	No
virtual_machine	186	A VM with 1, 2, 4, or 8 GPUs	POST /api/machines	Yes
bare_metal	4	A physical 8-GPU host	POST /api/machines	Yes

All values on this page were read from GET /api/deployment-profiles?workload=gpu on 22 August 2026. Every profile in all three catalogs was available.

GPU prices track available supply

Unlike applications and CPU VMs, GPU rates are not a fixed price list. Varity places GPU workloads on independent capacity, so which cards exist, in which regions, at which hourly rate, follows what is actually on offer at the moment you ask. The tables below are a dated snapshot, not a rate card.

Read GET /api/deployment-profiles?workload=gpu&execution_class=<class> for the current catalog, and take the hourly figure from the quote you are about to spend against. The quote carries valid_until and a resource_fingerprint; if either the deadline passes or the configuration changes, request a fresh one.

No monthly cap on GPU

The monthly cap that CPU VMs carry does not exist on GPU. monthly_cap_microusd is absent from every GPU profile. Any monthly figure you compute from an hourly rate is your own estimate, not a ceiling Varity enforces. A forgotten 8× H100 VM at $59.562/hour accrues about $1,429 a day.

GPU Containers

Nine offers, four card models. Pick one, deploy a container image onto it. There is no host to administer, no OS image to choose, and no region in the profile.

Offer	GPU model	VRAM	Interface	GPUs	Hourly
RTX 4090 24 GB	rtx4090	24 GB	pcie	1	$1.431
A100 80 GB	a100	80 GB	sxm	1	$2.9835
A100 80 GB	a100	80 GB	sxm	1	$3.024
A100 80 GB	a100	80 GB	sxm	1	$3.1185
A100 80 GB	a100	80 GB	sxm	1	$3.294
H100 80 GB	h100	80 GB	sxm	1	$3.6585
H200 141 GB	h200	141 GB	sxm	1	$4.6305
H200 141 GB	h200	141 GB	sxm	1	$4.806
H200 141 GB	h200	141 GB	sxm	1	$5.13

Repeated labels are separate offers with their own id and their own price. There are four distinct A100 offers and three distinct H200 offers. Take the cheapest one that is available.

Container profiles differ from machine profiles field by field:

• Identifier is id with a gpo- prefix, not profile_id with mp-.
• Name is label, not display_name.
• Price is customer_price.hourly_usd, a decimal in dollars, not amount_microusd.
• hardware describes the card (vendor, model, vram_mb, interface, count), not a host. There are no cpu_units, memory_mb, or storage_mb.
• There is no region and no os_images.
• availability is an object (status, matching_capacity_count, aggregate_available_units, source_age_seconds), not a string.
• schema_version is gpu-container-profiles-v1.

Every container offer carries price_basis: "exact" and pricing_policy_version: "gpu-hourly-v2-2026-07-28".

Deploy a GPU Container

Quote it, then deploy it as an application with the GPU attached.

1. Read the container catalog

   curl "https://varity.app/api/deployment-profiles?workload=gpu&execution_class=container" \
     -H "Authorization: Bearer $VARITY_API_KEY"

2. Quote the offer

   curl -X POST https://varity.app/api/pricing/accelerator-quote \
     -H "Authorization: Bearer $VARITY_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "execution_class": "container",
       "profile_id": "gpo-8057971715b95861f8775a0f"
     }'

   Returns quote_token, hourly_usd, authorization_usd, and valid_until. Pass the token through unchanged; do not inspect or log it.

3. Create the deployment

   curl -X POST https://varity.app/api/deployments \
     -H "Authorization: Bearer $VARITY_API_KEY" \
     -H "Idempotency-Key: gpu-container-001" \
     -H "Content-Type: application/json" \
     -d '{
       "name": "my-inference",
       "image": { "ref": "ghcr.io/acme/infer:1.0", "port": 8000 },
       "accelerator": {
         "profile_id": "gpo-8057971715b95861f8775a0f",
         "count": 1,
         "alternatives": [
           { "vendor": "nvidia", "model": "h100", "vram_mb": 81920, "interface": "sxm" }
         ]
       },
       "accelerator_quote_token": "<quote_token from step 2>"
     }'

   Copy alternatives straight out of the profile. accelerator and accelerator_quote_token must be sent together. Either one alone is rejected. count is 1 to 24 and alternatives holds 1 to 8 entries. Static-hosting deployments cannot carry an accelerator at all.

GPU VMs

186 profiles across 55 named configurations, 21 card models, and 53 regions. Same card in a different region or on a different host is a different profile at a different price, so the spread within one name is wide. L40S 48 GB ranges from $1.431 to $5.805.

Prices below come from customer_price.amount_microusd ÷ 1,000,000. Memory and disk are converted from memory_mb and storage_mb. “Offers” is how many profiles carry that name.

Single GPU

Profile	GPU model	VRAM	Interface	Offers	Hourly	Cheapest offer: vCPU / RAM / disk / region
A30 24 GB	a30	24 GB	pcie	1	$0.513	16 / 45 GB / 238 GB / us-central-1
RTXA6000 48 GB	rtxa6000	48 GB	pcie	5	$0.7965 to $3.78	6 / 22 GB / 238 GB / us-central-2
RTX 4090 24 GB	rtx4090	24 GB	pcie	2	$1.134 to $1.1745	12 / 65 GB / 792 GB / osl1
RTX6000ADA 48 GB	rtx6000ada	48 GB	pcie	7	$1.161 to $3.078	12 / 67 GB / 326 GB / us-central-2
RTX5090 32 GB	rtx5090	32 GB	pcie	1	$1.269	12 / 112 GB / 838 GB / oslo-norway-3
L40 48 GB	l40	48 GB	pcie	5	$1.3905 to $1.9575	14 / 67 GB / 582 GB / us-central-3
L40S 48 GB	l40s	48 GB	pcie	13	$1.431 to $5.805	12 / 67 GB / 582 GB / us-central-2
RTX4000ADA 20 GB	rtx4000ada	20 GB	pcie	1	$1.5525	8 / 30 GB / 466 GB / toronto-canada-1
A4000 16 GB	a4000	16 GB	pcie	1	$1.566	8 / 42 GB / 233 GB / newyork-usa-1
L4 24 GB	l4	24 GB	pcie	2	$1.863	8 / 45 GB / 466 GB / warsaw-poland-1
A100 80 GB	a100	80 GB	pcie / sxm	16	$2.133 to $6.426	28 / 112 GB / 792 GB / canada-1
A10 24 GB	a10	24 GB	pcie	2	$2.5245	30 / 186 GB / 1.3 TB / sanjose-usa-2
A5000 24 GB	a5000	24 GB	pcie	1	$2.781	8 / 42 GB / 233 GB / newyork-usa-1
RTXPRO6000 96 GB	rtxpro6000	96 GB	pcie	6	$3.5505 to $4.293	16 / 134 GB / 675 GB / us-central-9
A40 48 GB	a40	48 GB	pcie	4	$3.645	24 / 112 GB / 1.3 TB / bangalore-india-1
GH200 96 GB	gh200	96 GB	pcie	1	$4.4955	64 / 402 GB / 3.7 TB / dulles-usa-3
V100 32 GB	v100	32 GB	pcie	1	$4.59	8 / 28 GB / 233 GB / newyork-usa-1
H100 80 GB	h100	80 GB	pcie / sxm	11	$6.4665 to $11.745	24 / 224 GB / 931 GB / paris-france-1
H200 141 GB	h200	141 GB	sxm	5	$8.7615 to $8.7885	24 / 224 GB / 671 GB / atlanta-usa-1
B200 192 GB	b200	192 GB	sxm	1	$13.905	20 / 209 GB / 477 GB / me-west-1

A100 and H100 appear on both PCIe and SXM hosts. The interface is a property of the individual offer, so read alternatives[0].interface on the profile you actually pick.

Multi GPU

Profile	GPUs	VRAM each	Interface	Offers	Hourly	Cheapest offer: vCPU / RAM / region
2x RTXA6000 48 GB	2	48 GB	pcie	3	$1.6065 to $7.56	14 / 45 GB / us-central-2
2x A16 16 GB	2	16 GB	pcie	2	$1.998	12 / 119 GB / singapore-singapore-1
2x RTX 4090 24 GB	2	24 GB	pcie	1	$2.2815	24 / 130 GB / osl1
2x RTX6000ADA 48 GB	2	48 GB	pcie	4	$2.322 to $3.807	26 / 134 GB / us-central-3
2x L40 48 GB	2	48 GB	pcie	5	$2.781 to $3.915	26 / 134 GB / us-central-3
2x L40S 48 GB	2	48 GB	pcie	8	$2.8485 to $11.4885	24 / 134 GB / us-central-2
2x A4000 16 GB	2	16 GB	pcie	1	$3.132	16 / 84 GB / newyork-usa-1
2x L4 24 GB	2	24 GB	pcie	2	$3.726	16 / 89 GB / warsaw-poland-1
2x A100 80 GB	2	80 GB	pcie / sxm	6	$4.239 to $6.8715	60 / 224 GB / canada-1
2x A5000 24 GB	2	24 GB	pcie	1	$5.562	16 / 84 GB / newyork-usa-1
2x RTXPRO6000 96 GB	2	96 GB	pcie	4	$7.0875 to $8.586	30 / 268 GB / us-east-1
2x V100 32 GB	2	32 GB	pcie	1	$9.18	16 / 56 GB / newyork-usa-1
2x H100 80 GB	2	80 GB	pcie / sxm	5	$11.826 to $16.4295	80 / 317 GB / finland-2
4x RTXA6000 48 GB	4	48 GB	pcie	2	$3.213 to $15.1335	30 / 89 GB / us-central-2
4x A16 16 GB	4	16 GB	pcie	1	$4.023	24 / 238 GB / bangalore-india-1
4x RTX6000ADA 48 GB	4	48 GB	pcie	2	$4.6305 to $7.6005	52 / 268 GB / us-central-2
4x RTX 4090 24 GB	4	24 GB	pcie	1	$4.698	60 / 328 GB / oslo-norway-3
4x L40 48 GB	4	48 GB	pcie	5	$5.562 to $11.6775	50 / 268 GB / us-central-3
4x L40S 48 GB	4	48 GB	pcie	6	$5.697 to $22.842	46 / 268 GB / us-central-2
4x A4000 16 GB	4	16 GB	pcie	1	$6.2775	32 / 168 GB / newyork-usa-1
4x L4 24 GB	4	24 GB	pcie	1	$7.452	32 / 179 GB / warsaw-poland-1
4x A100 80 GB	4	80 GB	pcie	2	$8.478 to $10.584	124 / 447 GB / canada-1
4x A5000 24 GB	4	24 GB	pcie	1	$11.1375	32 / 168 GB / newyork-usa-1
4x RTXPRO6000 96 GB	4	96 GB	pcie	5	$13.716 to $17.172	120 / 335 GB / finland-1
4x V100 32 GB	4	32 GB	pcie	1	$18.3465	32 / 112 GB / newyork-usa-1
4x H100 80 GB	4	80 GB	sxm	2	$23.409 to $23.652	176 / 633 GB / finland-3
4x H200 141 GB	4	141 GB	sxm	2	$28.755 to $28.998	176 / 633 GB / finland-3
8x RTX 4090 24 GB	8	24 GB	pcie	1	$6.2775	88 / 212 GB / casper-usa-2
8x RTX5090 32 GB	8	32 GB	pcie	1	$10.9755	120 / 212 GB / casper-usa-2
8x L40 48 GB	8	48 GB	pcie	3	$12.555 to $15.687	252 / 432 GB / canada-1
8x L4 24 GB	8	24 GB	pcie	1	$14.904	64 / 358 GB / warsaw-poland-1
8x A100 80 GB	8	80 GB	pcie / sxm	12	$17.577 to $47.0475	252 / 1.7 TB / canada-1
8x RTXPRO6000 96 GB	8	96 GB	pcie	1	$28.3635	120 / 1.4 TB / us-central-9
8x H100 80 GB	8	80 GB	sxm	3	$40.149 to $59.562	192 / 1.6 TB / canada-1
8x H200 141 GB	8	141 GB	sxm	3	$50.1795 to $57.51	208 / 1.7 TB / tokyo-japan-5

Total VRAM and price move independently. 192 GB of VRAM costs $3.213/hour as 4x RTXA6000 48 GB in us-central-2, or $6.2775/hour as 8x RTX 4090 24 GB in casper-usa-2. Multiply accelerator.count by vram_mb before comparing prices.

GPU VM Regions

53 regions:

amsterdam-netherlands-2, atlanta-usa-1, austin-usa-1, bangalore-india-1, beltsville-usa-1, calgary-canada-1, canada-1, casper-usa-2, chicago-usa-2, chicago-usa-3, culpeper-usa-1, dallas-usa-3, desmoines-usa-1, dulles-usa-1, dulles-usa-3, eu-north-1, eu-west-1, finland-1, finland-2, finland-3, frankfurt-germany-7, houston-usa-1, houston-usa-2, jerusalem-israel-1, kansascity-usa-1, kansascity-usa-6, london-uk-1, me-west-1, mon1, montreal-canada-2, mumbai-india-1, newyork-usa-1, newyork-usa-2, osl1, oslo-norway-3, paris-france-1, paris-france-5, phoenix-usa-2, saltlakecity-usa-1, sanjose-usa-2, singapore-singapore-1, sydney-australia-1, tokyo-japan-1, tokyo-japan-5, toronto-canada-1, us-1, us-central-1, us-central-2, us-central-3, us-central-9, us-east-1, us-southeast-1, warsaw-poland-1

Region is a property of the profile, not a create-time argument. Filter the catalog by region to pick where a machine runs.

GPU VM OS Images

Image choice is per profile and varies a lot: 80 of 186 profiles offer exactly one image, and the largest offer 18. Across the catalog there are 49 distinct labels.

149 of 186 profiles include at least one image whose label names a CUDA version, for example Ubuntu 24.04 + CUDA 12.8 Open + Docker, ubuntu22.04_cuda12.2_shade_os, or Ubuntu Server 22.04 LTS R570 CUDA 12.8. The remaining 37 do not, and some catalog images are plain OS builds (Debian 12 Plain, AlmaLinux 9 Plain, ubuntu20.04).

Do not assume a driver or CUDA toolchain is present. Read os_images on the specific profile and pick the label you need.

GPU Bare Metal

Four profiles. One configuration, in four regions, at one price. No hypervisor.

Region	Profile	GPUs	VRAM each	vCPU	RAM	Disk	Hourly
ams	8x RTXPRO6000 96 GB	8	96 GB	192	1.4 TB	1.2 TB	$53.4735
chicago-usa-4	8x RTXPRO6000 96 GB	8	96 GB	192	1.4 TB	1.2 TB	$53.4735
syd2	8x RTXPRO6000 96 GB	8	96 GB	192	1.4 TB	1.2 TB	$53.4735
tyo4	8x RTXPRO6000 96 GB	8	96 GB	192	1.4 TB	1.2 TB	$53.4735

All four offer exactly one OS image: ubuntu24.04_cuda12.4_shade_os. There is nothing to choose.

The GPU VM catalog also carries an 8x RTXPRO6000 96 GB at $28.3635/hour in us-central-9. Bare metal costs more because you get the physical host, not a VM on it.

Create a GPU VM or Bare-Metal Machine

Identical to the CPU VM flow, with execution_class set to virtual_machine or bare_metal.

1. Quote it

   curl -X POST https://varity.app/api/pricing/machine-quote \
     -H "Authorization: Bearer $VARITY_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "profile_id": "mp-0157cd6e4cf82ff83291d645",
       "execution_class": "virtual_machine",
       "os_image": "os-ubuntu24-04-a2801be96ab2"
     }'

   additional_storage_gb is not accepted here. It is valid only for cpu_virtual_machine. What the profile lists as storage_mb is what you get.

2. Create it before valid_until

   curl -X POST https://varity.app/api/machines \
     -H "Authorization: Bearer $VARITY_API_KEY" \
     -H "Idempotency-Key: gpu-vm-create-001" \
     -H "Content-Type: application/json" \
     -d '{
       "name": "trainer-01",
       "profile_id": "mp-0157cd6e4cf82ff83291d645",
       "execution_class": "virtual_machine",
       "os_image": "os-ubuntu24-04-a2801be96ab2",
       "ssh_public_key": "ssh-ed25519 AAAA...",
       "accelerator_quote_token": "<quote_token from step 1>"
     }'

   The quote carries valid_until and a resource_fingerprint that binds it to that exact configuration. If either the deadline or the configuration changes, request a fresh quote rather than reusing the token.

3. Poll for access

   curl https://varity.app/api/machines/$MACHINE_ID \
     -H "Authorization: Bearer $VARITY_API_KEY"

   access is null until ready, then carries protocol: "ssh", host, port, and username.

ssh -p <port> <username>@<host>

Billing

• Every GPU profile quotes billing_period: "hour" in USD. The binding rate is the hourly_usd in the quote you accepted, not the catalog list price.
• There is no monthly_cap_microusd on any GPU profile.
• month_to_date reporting on the machine record is a CPU VM field. A GPU machine does not report its own running total there. Reconcile GPU spend through billing.
• Extra storage cannot be attached to a GPU machine.

Restart and Delete

curl -X POST https://varity.app/api/machines/$MACHINE_ID/actions \
  -H "Authorization: Bearer $VARITY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"action": "restart"}'

restart and soft_reboot act in place. The machine keeps its identity, its access, and its billing.

Deleting is the only way to stop GPU billing

Shutting the guest OS down from inside stops your workload, not your reservation. The GPUs stay allocated and keep billing, with no cap to stop them.

curl -X DELETE https://varity.app/api/machines/$MACHINE_ID \
  -H "Authorization: Bearer $VARITY_API_KEY"

Deletion is asynchronous and acceptance is not proof of closure. Read the machine back and check lifecycle.cleanup_state for complete. Deletion destroys the disk. Copy anything you need off the machine first.

Related

Machine Catalogs All four catalogs, with counts and price ranges.

CPU Virtual Machines All 13 CPU VM sizes, prices, and the monthly cap.

API Reference Authentication and the full endpoint surface.