CH·02CLI reference

gpu pricing

Show GPU pricing across regions and tiers

Synopsis

Shows the current price grid (USD/hour) for the GPU.ai catalog. Filter by GPU type, region, or tier. The CLI follows server pagination automatically, so the full catalog is always shown.

Each row is one offer per SKU: gpu_count is the number of GPUs in the configuration and the price covers the whole configuration (an 8-GPU node lists ~8x the single-GPU price — compare rows by count, not price alone). Offers for the same SKU — same gpu_type, gpu_count, region, tier, and boot class — are aggregated into a single row: the cheapest price wins and availability is summed, so a SKU never appears as multiple rows that differ only by price. (Fast- and slow-boot capacity stay separate rows — see boot class below.) Out-of-stock offers (available 0) are hidden by default; pass --include-unavailable to see them.

Each row also lists the hardware for the offering — the VCPU (vCPU count), RAM (GB), and DISK (GB) columns in table view (cpu_cores, ram_gb, instance_disk_gb in JSON). DISK (GB) is the filesystem a launch of that row actually receives — the platform default (100 GB) on capacity where GPU.ai sets the disk, or the machine's own disk on fixed-disk (typically full-VM) capacity. A trailing + marks rows that accept a custom size via gpu instances create --disk (disk_configurable in JSON); those rows also carry disk_price_per_gb_hour in JSON — the retail price of each GB beyond the included default, folded into the instance's hourly rate at launch (0 = larger sizes are currently free). A spec the upstream did not report renders as -. Filter to one SKU with --gpu-type/--region/--tier to see exactly what a launch will get.

Each row also carries a boot class: instant capacity is reachable within about a minute of launch, while slow-boot capacity takes ~3-5 minutes to provision (BOOT column in table view, instant_boot bool in JSON). Use --instant-boot to hide the slow rows.

Each row also carries a capacity class: secure marks data center-operated hardware, while community marks supplier/peer-hosted capacity with a softer SLA (SOURCE column in table view; capacity_class string plus its community bool twin in JSON). The class is per *offering*, not per upstream — some upstreams price both secure and community capacity for the same SKU, and those stay separate rows. Use --community to show only community rows.

Each row also carries an environment class (ENV column) and a deployment class (TYPE column):

ENV is the image class the offer can launch: certified (can run the GPU.ai Certified Image — the one normalized, digest-pinned environment), raw-vm (a VM we bootstrap for you with root access, but that cannot take the Certified Image), or provider-template (container-native or otherwise non-normalizable capacity that launches from the provider's own runtime template). Pass --certified to show only offerings that can run the GPU.ai Certified Image.
TYPE is the deployment class: vm (we bootstrap the machine ourselves) or container (the provider runs our image natively). Filter with --type vm or --type container. A pre-field catalog row shows -.

Both columns are derived from provider *attributes*, never provider identity. See the Certified Image guide for the full contract.

For every certified-capable row, the per-machine pinned-environment spec (D-63-07) is printed as a footnote block after the table under Certified environment (pinned per machine): — the image ref, the short image digest, the minimum NVIDIA driver, and the CUDA generation:

Certified environment (pinned per machine):
  rtx_pro_6000_blackwell: image=ghcr.io/gpuai-dev/base-cuda128 digest=sha256:… driver>=570.26 cuda=12.8

In JSON view the same data is the environment object on each row (image, image_digest, driver_min, cuda), present only on certified-capable rows.

gpu pricing [flags]

Examples

# All prices
gpu pricing

# Just H100 SXM
gpu pricing --gpu-type h100_sxm

# Spot prices in a region
gpu pricing --region us-east --tier spot

# Only capacity that boots in under a minute
gpu pricing --instant-boot

# Only community-tier capacity
gpu pricing --community

# Only offerings that can run the GPU.ai Certified Image
gpu pricing --certified

# Only container-native capacity
gpu pricing --type container

# Certified image on VM-class capacity only
gpu pricing --certified --type vm

# Include out-of-stock offers too
gpu pricing --include-unavailable

# JSON for jq
gpu pricing --output json | jq '.[] | select(.available)'

Options

      --certified             Only show offerings that can run the GPU.ai Certified Image (server-side certified_image filter)
      --community             show only community-tier capacity
      --gpu-type string       Filter by GPU type ID
  -h, --help                  help for pricing
      --include-unavailable   Include offers with zero current availability
      --instant-boot          Only show capacity that boots in under a minute (hides ~3-5 min boot offerings)
      --region string         Filter by region code
      --tier string           Filter by tier: on_demand | spot
      --type string           Filter by deployment type: vm | container

Options inherited from parent commands

      --api-base string   API base URL (override with GPUAI_API_BASE env) (default "https://api.gpu.ai/v1")
      --debug             Enable debug logging to stderr
  -o, --output string     Output format: table|json (default table on TTY, json otherwise)

Synopsis

Examples

Options

Options inherited from parent commands

SEE ALSO