TensorDock and Runpod both rent GPUs by the hour, and there the similarity stops. TensorDock is a marketplace: independent hosts list their own hardware and set their own prices, which is why the same card shows several different rates on the same page. Runpod runs its own fleet across two tiers, Secure Cloud and Community Cloud.
That difference decides almost everything else about the two platforms, including whether a price you see today is the price you get tomorrow.
Rates below were read from each provider's own pricing page on 21 August 2026.
| Capability | TensorDock | Runpod |
|---|---|---|
| Supply model | Marketplace of independent hosts | Own fleet, Secure and Community tiers |
| GPU models | 45, across 100+ locations | 24, across 31 global regions |
| Full VM with root and Windows | Yes, KVM virtualization | Containers only |
| Serverless for your own container | Not offered | Serverless, sub-200ms cold starts via FlashBoot |
| Multi-node training | Not published | Instant Clusters, self-serve |
| Price stability | Floors set by competing hosts | One published rate per tier |
| Billing | Continuous deduction from balance | Per second, no minimum |
| When funds run out | Servers are deleted | Pods stop, data on volumes persists |
| Ingress and egress charges | None | None |
| Ownership | Acquired by Voltage Park, March 2025 | Independent |
TensorDock capabilities and rates were read from tensordock.com on 21 August 2026. Their prices are marketplace floors: the same GPU carries different prices depending on host, location and redundancy. Runpod rates pull live.
TensorDock is now part of Voltage Park
This is the first thing to know, and it does not appear anywhere on TensorDock's own site.
Voltage Park acquired TensorDock in March 2025, with terms undisclosed. TensorDock's founder became General Manager of On-Demand at Voltage Park, and a Voltage Park director took over as General Manager of TensorDock. The marketplace was stated at the time to continue operating as normal. Voltage Park still links TensorDock from the Company section of its own site footer.
Since then Voltage Park has itself merged with Lightning AI. So a team evaluating TensorDock today is evaluating a marketplace owned by a company that is midway through its own integration.
None of that necessarily makes TensorDock a worse product. It is context you should have, particularly if your reason for choosing a vendor includes wanting to know who will be operating it in a year. It is also worth knowing if you are comparing TensorDock against Voltage Park, because those are not two independent options.
Pricing, and why the answer is not a single number
TensorDock publishes floor prices. Their own FAQ explains why: they describe themselves as a marketplace of independent hosts who compete and set their own pricing. Hosts differ by location, redundancy and hardware, so the figure you see is the cheapest listing at that moment, not a rate card.
| Card | TensorDock floor | Runpod Community | Runpod Secure |
|---|---|---|---|
| H100 SXM | From $2.25/hr | {{gpu:h100-sxm:community}}/hr | {{gpu:h100-sxm}}/hr |
| A100 SXM | From $1.80/hr | {{gpu:a100-sxm:community}}/hr | {{gpu:a100-sxm}}/hr |
| RTX 4090 | From $0.35/hr | {{gpu:rtx-4090:community}}/hr | {{gpu:rtx-4090}}/hr |
| Consumer entry | From $0.12/hr | {{gpu:rtx-a5000:community}}/hr | {{gpu:rtx-a5000}}/hr |
TensorDock figures are marketplace floors read from tensordock.com on 21 August 2026, not fixed rates. Their consumer entry rate covers cards Runpod does not carry, so the final row is a floor comparison rather than a like-for-like one. Runpod rates pull live.
The result splits by card rather than favoring one platform. Anyone telling you one of these platforms is simply cheaper than the other has not checked more than one card.
What TensorDock does well
Breadth. Forty-five GPU models across more than 100 locations in over 20 countries is the widest selection in this category, and it includes hardware most providers do not carry. If you need a specific card in a specific country, TensorDock is likelier to have it.
Entry price. Consumer GPUs start at $0.12/hr and CPU-only instances at $0.012/hr. You can begin with a $5 deposit.
Full machine control. KVM virtualization gives root access to a real VM, with your own drivers and Windows support. If your workload needs an operating system rather than a container, that is a genuine advantage over container-based platforms, and it is the clearest reason to pick them over Runpod.
No transfer fees. TensorDock charges no ingress or egress, the same as Runpod.
What the marketplace model costs you
None of the following is a criticism of TensorDock's execution. They are properties of running a marketplace.
Variable hardware. TensorDock describes its supply as a mix of Tier 3 and Tier 4 data centers and, in its own words, converted mining rigs for maximum price-to-performance. Both are on the same platform. Which one you get depends on the listing you pick.
Variable pricing. A floor price is not a budget. If the cheapest host for your card is unavailable, the next listing costs more.
Reliability is a host standard, not a customer guarantee. TensorDock holds hosts to a 99.99% uptime standard and removes those who fall short. That is a supplier requirement they enforce, which is not the same thing as an availability commitment to you.
Balance-based billing has a hard edge. Funds are deducted continuously, and their own FAQ states that when your balance reaches $0, your servers are automatically deleted. Not stopped. Deleted. That is worth knowing before you leave a long training run unattended.
Where Runpod is different
The clearest gap is scope. TensorDock rents virtual machines. Runpod covers three workload shapes on one account: Pods for development and long-running jobs, Serverless for inference that scales to zero, and Instant Clusters for multi-node training, all sharing the same container images.
If you are serving a model rather than running a machine, that difference matters more than any hourly rate. A VM you rent by the hour bills whether or not a request arrives. A Serverless worker that scales to zero does not.
Two smaller differences follow from running an owned fleet. Billing is per second with no minimum, and stopping a pod stops the charge rather than deleting the server. And the platform is SOC 2 Type II across 31 global regions, independently verified for HIPAA and GDPR, which is a single answer rather than one that varies by host.
Which one should you choose: TensorDock or Runpod?
Choose TensorDock if you need a specific GPU in a specific country, want a full VM with root and Windows rather than a container, or are optimizing hard on entry price and can manage variance between hosts.
Choose Runpod if your work involves serving inference, if you need per-second billing on bursty jobs, if you want the same hardware and compliance answer every time you deploy, or if your next step after renting a machine is training across several of them.
The honest split: TensorDock is the better marketplace. Runpod is the better platform. If what you want is a machine, price it on both. If what you want is somewhere to build, serve and train without changing vendors, that is the case for Runpod, and it is not a price argument.
Frequently asked questions
Who owns TensorDock?
Voltage Park, which acquired TensorDock in March 2025 with terms undisclosed. TensorDock's founder became General Manager of On-Demand at Voltage Park, and the marketplace was stated to continue operating as normal. Voltage Park has since merged with Lightning AI. TensorDock's own site does not mention any of this.
Is TensorDock cheaper than Runpod?
On some cards. TensorDock's H100 SXM floor of $2.25/hr is below Runpod's {{gpu:h100-sxm:community}}/hr Community rate. Its A100 SXM floor of $1.80/hr sits above Runpod's {{gpu:a100-sxm:community}}/hr Community rate. Because TensorDock is a marketplace, its figures are floors rather than fixed rates, so compare the specific card and configuration you need on the day you need it.
Does TensorDock offer serverless GPU inference?
No. TensorDock rents virtual machines, including an Instant VMs product for faster provisioning. There is no scale-to-zero serverless tier. Runpod Serverless scales to zero between requests and starts workers in under 200ms via FlashBoot.
What happens if my TensorDock balance runs out?
Their own FAQ is explicit: servers are automatically deleted when the balance reaches $0. Plan your deposits around long-running jobs accordingly. On Runpod, stopping a pod stops billing for compute while data on a network volume persists.
Which has more GPU models: TensorDock or Runpod?
TensorDock, with 45 models against Runpod's 24. The trade is consistency: TensorDock's selection comes from independent hosts with varying hardware and pricing, while Runpod's is one fleet with one rate card per tier.
Do either charge egress fees?
Neither does. TensorDock and Runpod both state no ingress or egress charges.
