Beyond the GPU-hour · checked October 2, 2026
What else GPU clouds bill
The hourly rates on this site are the GPU-hour. The bill can also carry data sent out, persistent storage, an instance kept while stopped, and time rounded up to the billing step. What each provider's own pages say about those, read by hand, each linked to the page it was read from.
- Providers read24five questions each
- Data out free165 charge per GB
- Not published7answers the pages do not give
Side by side
| Provider | In the rate | Data out | Storage | While stopped | Billing step |
|---|---|---|---|---|---|
| Amazon Web Services | Instance price covers vCPU, RAM and local NVMe instance store (p5.48xlarge: 192 vCPU, 2 TiB, 8x3.84 TB NVMe); EBS volumes billed separately our reading source | 100 GB/month free across AWS, then $0.09/GB for the first 10 TB/month out to internet, US East (N. Virginia) source | EBS gp3 $0.08/GB-month in US East (N. Virginia); gp2 $0.10 source | Stopped instance not billed for usage; attached EBS volumes (and Elastic IPs) keep billing source | Per second, one-minute minimum each time the instance starts (Linux) source |
| Civo | Included: vCPU, RAM, local NVMe and free data transfer. The H100 Small is 23 vCPU, 184 GB RAM, 200 GB NVMe. our reading source | Free and unlimited. No egress or ingress fees. source | Persistent volumes: $0.11/GB-month ($0.000149/GB-hour). source | A stopped instance bills the same as a running one, at the full hourly price, until it is deleted. source | Hourly, rounded up to the next full hour, so the effective minimum is one hour. Invoiced monthly. source |
| CoreWeave | Instance price listed with its vCPU, RAM and local NVMe (HGX H100: 128 vCPU, 2,048 GB, 61.44 TB) at $49.24/hr our reading source | Free: no charge for data transfer to the internet source | Distributed File Storage $0.070/GB-month (binary GB); billed on provisioned capacity source | Not published; nodes have no stopped state on the pricing or billing pages. DFS billed on provisioned capacity source | Increment and minimum not published; usage counted in instance-hours, attributed every 30 seconds source |
| Crusoe | Local NVMe (ephemeral) included; instance types bundle vCPU and RAM; persistent and OS disks billed separately our reading source | free: no charge for ingress or egress, in VPC or to/from internet source | Persistent disk $0.08/GiB-month (shared disks $0.07) source | Stopped on-demand VM: no compute charge; 128 GB OS disk billed at $0.08/GiB-month; reserved VMs bill regardless source | Per second while running (pricing FAQ says per-hour or per-second); no minimum commitment source |
| Cudo Compute | GPU only. vCPU ($0.0049/vCPU-hr), RAM ($0.0049/GiB-hr), storage and IPv4 are billed separately (A100 machine type in the official API). source | Not published. Checked the pricing page, the billing docs, the help FAQ and the full docs text. source | VM disk storage is $0.00014/GiB-hr, about $0.10/GiB-month on Cudo's 730-hour month. The field is named minStorage, so it may be the lowest storage class. our reading source | A stopped VM keeps billing for all its resources. You must delete it to stop charges. source | Usage is metered to the nearest second and recalculated hourly. A $10 minimum top-up is needed to deploy. source |
| Denvr Dataworks | Included: vCPU, RAM and direct-attached NVMe. On-demand local NVMe is ephemeral and erased on stop. source | Free internet ingress and egress. source | Persistent disks: $0.08 per GB-month (network shared filesystems are the same price). source | Stopping from the console ends compute billing, but the root disk and filesystems keep billing. A shutdown from inside the OS keeps the full instance billing. source | Per minute according to the pricing page; Denvr's billing docs say one-second increments. No minimum stated. source |
| Gcore | Included: each listed configuration bundles vCPU, RAM and local SSD/NVMe in one price, e.g. Virtual GPU 1×H100 with 24 vCPU, 232 GB RAM and 1250 GB. Prices exclude VAT. our reading source | Free egress from GPU clusters. Gcore's cloud docs also say Bare Metal outgoing traffic is billed separately, but no per-GB price is published. source | File Shares (the persistent storage named for GPU clusters): €0.10 per GB, billing period not stated. Block volume price is not on a static page. source | Bare Metal GPUs bill until deleted, even when not running. Shelving a Virtual GPU stops compute billing. Volumes and IPs bill until deleted. source | Per minute while in use, pay-as-you-go. A balance of at least €4/$4 is needed to start. source |
| Google Cloud | Listed A3/A2 price includes GPUs, vCPU, memory and bundled Local SSD (a3-highgpu-8g: 208 vCPU, 1,871 GB, 6,000 GiB SSD); boot disk extra source | Premium Tier (default), us-central1 to North America: first 1 GiB free, then $0.12/GiB up to 1 TiB source | Balanced Persistent Disk $0.000136986/GiB-hour in us-central1, about $0.10/GiB-month; Hyperdisk Balanced about $0.08 plus IOPS/throughput our reading source | Stopped instance: no vCPU/memory charge; disks and external IPs billed regardless of state source | Per second after a 1-minute minimum (vCPUs, GPUs, memory) source |
| Hyperstack | Included in the GPU rate: vCPU, RAM, root disk and ephemeral local NVMe. Only the GPUs are billed. source | Free: no charge for egress, ingress or bandwidth. source | Shared Storage Volumes (Cloud-SSD): $0.000096774/GB-hour, about $0.07/GB-month. Billed even when not attached to a VM. source | A stopped (SHUTOFF) VM bills in full, GPU included. A HIBERNATED VM bills only its storage and any retained public IP. source | Billed per minute (prepaid on-demand). No minimum charge published. source |
| Jarvislabs | Included: vCPU and RAM, listed with each GPU's per-GPU-hour price. Storage is billed separately at $0.10/GB-month. our reading source | Free in-region. No price is published for internet egress. source | Instance storage: $0.10 per GB-month. source | Pausing releases the GPU and stops compute billing. Stored data bills at $0.00014/GB-hr and is deleted if the balance goes negative. source | Per minute. No minimum charge stated. source |
| Lambda | Per-GPU rate covers vCPU, RAM and local SSD listed per instance (8x H100 SXM: 208 vCPU, 1800 GiB, 22 TiB SSD) our reading source | Free: no egress fees source | Filesystem rate not published; docs' $0.20/GiB-month is an example that may not be current; price shown at creation source | No stop state: instances run until terminated; OS shutdown leaves billing on. Filesystems bill while they exist source | One-minute increments, from passing health checks to termination; no minimum published source |
| Massed Compute | Included: vCPU, RAM and storage listed with each configuration's hourly price, e.g. RTX PRO 6000 ×1 with 16 vCPU, 144 GB RAM and 725 storage (unit not shown). our reading source | Free: no bandwidth charges. source | Not published. Storage comes only bundled per configuration; the pricing page and billing docs give no separate volume price. source | A stopped VM keeps billing at the full hourly rate. Only deleting it stops billing. source | Per minute: the hourly rate divided by 60, debited every minute. Auto-recharge needs a minimum balance of at least $5. source |
| Microsoft Azure | VM price includes vCPU, RAM and local temp disk only; OS and data managed disks billed separately source | First 100 GB/month free, then $0.087/GB for next 10 TB, from North America (Microsoft premium network routing) source | Premium SSD LRS P30 (1 TiB) $135.17/month in East US, about $0.132/GiB-month; Premium SSD v2 $0.00011/GiB-hour our reading source | Stopped (allocated) keeps billing compute; Stopped (deallocated) stops compute billing, disks and networking still charged source | Per full minute; partial minute not charged; no minimum published source |
| Nebius | Preset vCPU and RAM appear inside the GPU-hour for H100/H200/B200/B300/RTX PRO; L40S bills CPU and RAM separately; disks extra our reading source | free: egress and ingress traffic not charged source | Network SSD disk $0.071/GiB-month (non-replicated $0.053; shared filesystem $0.08) source | Stopped VM: GPU, vCPU, RAM not charged; disks and filesystems keep billing source | Per second (priced per hour, charged proportionally); no minimum charge published source |
| Oracle Cloud (OCI) | Bare-metal GPU shape billed per GPU-hour, no separate OCPU line; BM.GPU.H100.8 has 112 OCPU, 2048 GB RAM, 16x3.84 TB NVMe; block/boot volumes extra our reading source | First 10 TB/month free, then $0.0085/GB, originating in North America/Europe/UK source | Block Volume Balanced: $0.0255/GB-month storage + 10 VPUs ($0.017) = $0.0425/GB-month our reading source | Stopped GPU instances keep billing (except VM.GPU.A10 series); must terminate to stop billing source | Per second; bare-metal (incl. BM.GPU) minimum one hour, VMs minimum one minute source |
| OVHcloud | Included: vCPU, RAM and local NVMe (l40s-90: 15 vCore, 90 GB, 400 GB NVMe). Public IPv4 billed separately since 2026-10-01. our reading source | Outbound included (free), except in Asia-Pacific: 1 TB/month per project, then per-GB (rate not shown). source | Block Storage Classic: $0.000066/GB-hour, about $0.04818/GB-month. High Speed and High Speed Gen2: about $0.09636/GB-month. source | A stopped or paused instance bills at the same rate. Shelved: only the snapshot bills. From OVHcloud global docs; the US support article returned 403. source | Hourly plan billed per second after a 1-minute minimum, or a monthly plan. From OVHcloud global docs; the US support article returned 403. source |
| Paperspace | Machine rate covers the GPU machine; storage, public IP addresses and other add-ons are billed on top, hourly our reading source | Free: no charge for ingress or egress on Paperspace Machines source | Block storage in size tiers with an hourly price and a monthly cap: 50 GB $0.0074/hr, at most $5.00/month source | Powering off stops compute billing; storage, public IPs and add-ons keep billing hourly until destroyed source | Hourly, by machine type; no minimum published source |
| RunPod | vCPU and RAM listed with each GPU rate; container and volume disk billed separately per GB-month our reading source | free: no fees for data ingress or egress source | Volume disk $0.10/GB-month on a running Pod ($0.20 stopped); network volume $0.07/GB-month under 1 TB source | Stopping releases the GPU (no GPU charge); volume disk keeps billing at $0.20/GB-month, container disk not charged source | Per second; at least one hour of credits needed to deploy an on-demand Pod (not a minimum charge) source |
| Scaleway | The rate covers vCPU, RAM and scratch NVMe (H100). The system Block Storage volume (125 GB minimum for a GPU OS) and the IPv4 are billed separately. our reading source | Free: Scaleway says Instance list prices include egress. Stated on the Virtual Instances page; the GPU price page is silent. our reading source | Block Storage 5K (lowest tier): €0.000130/GB-hour, about €0.095/GB-month at 730 h. The 15K tier is €0.000177/GB-hour. source | Powered off: compute billing stops, but storage volumes and IPs keep billing. Standby is charged as a running instance. source | GPU Instances billed per minute of uptime (RENDER-S per hour). No minimum charge published. source |
| Seeweb | Included: CPU cores, RAM and disk listed in each plan under one hourly cost, e.g. MI300X with 32 cores, 256 GB RAM and 2 TB disk at €2.35/hr. our reading source | Free: network traffic is not charged for Cloud Server GPU. source | No GPU-specific volume price. Cloud ScaleOut Filer, the add-on space the GPU page names, costs €45.50/month with 100 GB, then €11.50 per extra 100 GB (about €0.115/GB-month). our reading source | Not published. Checked the GPU page, the FAQ and the control-panel docs; they cover cancelling, not billing while powered off. source | Hourly: billed for hours of use, no advance payment. No stated minimum beyond one hour. source |
| Together AI | Per-GPU hourly rate; node vCPU/RAM not priced separately; local NVMe per node (ephemeral); shared storage billed separately our reading source | free: no data transfer fees in or out of the cluster source | Shared filesystem $0.16/GiB-month; minimum volume size 1 TiB source | No pause: reserved billed for full term, on-demand stops only on termination; storage keeps billing without a cluster source | On-demand per hour of cluster runtime; reserved 1-90 days paid upfront, non-refundable; preemptible metered every 1-2 min source |
| Vast.ai | GPU rate includes a GPU-proportional share of host CPU and RAM; storage and bandwidth priced separately source | Set by each host per offer ($/TB), charged on every byte up and down, in any instance state; no platform-wide rate source | Set by each host per offer, charged on allocated disk size; no platform-wide price source | GPU rental stops; storage keeps billing until the instance is deleted (bandwidth billed in any state) source | Per second while active; credits must be prepaid; no minimum charge published source |
| Verda | CPUs and RAM listed with each GPU instance price; OS and block volumes billed separately as storage our reading source | not published (pricing page, billing docs, instance and terms pages checked) source | NVMe block storage $0.20/GiB-month (shared filesystem also $0.20) source | Shut-down instances keep billing in full; kept volumes keep billing until deleted source | Prepaid 10-minute increments; unused part refunded if terminated early source |
| Vultr | The bare-metal GPU plan bundles CPU, RAM, local NVMe and a bandwidth allowance in one hourly_cost. www.vultr.com/pricing returned 403. our reading source | Plan allowance (15,360 GB on the GPU bare-metal plans), then $0.01/GB. Inbound is not metered. source | not read | A stopped instance keeps billing in full (CPU, RAM, storage, IP) until it is destroyed. source | Hourly, minimum one hour. Monthly cap of 730 h for Cloud GPU and 672 h for bare metal (our rows are bare metal). source |
How this is read
- By hand, from the provider's own pages - pricing pages, documentation, billing FAQs - never a third party, a sales call or a review site. Every cell links the page it was read from; the last check was .
- "Not published" means the pages were read and do not say. It is not a claim the fee is zero.
- "Our reading" marks an answer the page implies without stating it in so many words (a rate card that lists vCPU and RAM beside the GPU, for example). Everything else is stated on the linked page.
- Not added up. Providers bill these in units that do not convert (a free quota, a price per GB, a bundled disk), so no "true hourly cost" is computed and nothing here changes the Price Index or any rate.
- Read again by hand, not every hour like the rates. A wrong cell is a correction to data@gpueconomy.com with the page that says otherwise.
Data updated: