TensorDock review
Updated September 28, 2026 · DeployCue Editorial Team · Facts checked September 26, 2026
We earn a commission when you sign up through links on this site. Partner status can affect where a brand appears in our rankings.
Our verdict
TensorDock suits developers who need a specific GPU on demand without long commitments, and who would rather not pay to move data in or out. It scores 9.2 overall. Per-second billing and No ingress/egress fees keep short experiments and bursty inference cheap.
Key facts
- Best for
- Short, per-second GPU bursts with no egress bill at the end of the month.
- GPU models
- H100 SXM5 80GB, A100 SXM4 80GB, A100 PCIe 80GB, V100 SXM2 16GB, L40 48GB, RTX 4090 24GB, RTX 3090 24GB, RTX 6000 Ada 48GB, RTX A6000 48GB, RTX A4000 16GB (Source, checked September 26, 2026)
- GPU hour from
- $0.10 (Source, checked September 26, 2026)
- Billing granularity
- Per second (Source, checked September 26, 2026)
- Egress fee
- No ingress/egress fees (Source, checked September 26, 2026)
- Locations
- from 100 (Source, checked September 26, 2026)
- Interconnect
- 1 Gbps minimum port included (Source, checked September 26, 2026)
- Compliance certifications
- SOC 2, PCI DSS, ISO 27001, HIPAA (Source, checked September 26, 2026)
- Legal entity
- TensorDock.com, Inc (Source, checked September 26, 2026)
- Headquarters
- United States (Source, checked September 26, 2026)
- Support channels
- Email, Help-centre, Community (Source, checked September 26, 2026)
- Pricing model
- Hourly billing
- Official website
- tensordock.com
Apps and profiles
Pros and cons
Pros
- per second billing means an idle or unfinished test costs only what it used
- No ingress/egress fees, so pulling checkpoints and datasets out does not add a surprise line
- GPU hours start at $0.10, below the market median of €0.63
- SOC 2, PCI DSS and ISO 27001 cover regulated workloads without a separate audit
- from 100 with a 1 Gbps minimum port included attached
Cons
- Per-second billing suits bursts, but a long steady training run bills on the same meter with no committed-use rate listed under hourly
- Support runs on email, help centre and community only, with no live phone channel for incidents
- from 100 is broad but plan names are not given per region, so pinning a workload to a single market needs care
Who it suits
- Developers running inference or fine-tuning experiments on rented GPUs
- Teams that move large checkpoints and datasets in and out of their cluster
- Small teams wanting datacentre-grade assurances without a hyperscaler contract
Who should look elsewhere
- Workloads that need Blackwell-class silicon, where ComputePrices.com lists a wider set at H100 SXM, A100 SXM and H200
- Anyone who needs phone or chat support, since email, help centre and community covers only email, help centre and community
- Regional sovereign requirements, because SberCloud publishes its own card list at B300, H100 and A100
Where and what you run
TensorDock spreads its machines across from 100, which puts a node close to more users than the category median of 4 would suggest. The card list is H100 SXM5 80GB, A100 SXM4 80GB and A100 PCIe 80GB, so a workload can be matched to a memory size without waiting for a new generation. Networking is stated as 1 Gbps minimum port included, enough to stream weights and batches without special-casing the host. Compare that with SberCloud, whose published list is B300, H100 and A100, a shorter set focused on datacentre parts rather than the desktop and workstation boards TensorDock carries.
Plans, egress and billing
GPU hours start at $0.10 for the entry plan, below the category median of €0.63, and hourly is the only model listed under hourly. Metering is per second, so a job that runs for a short while is charged for what it actually used rather than a full hour. Data movement is No ingress/egress fees, which removes the line that usually surprises teams shipping checkpoints home. Against the wider market, a median entry price of €5.83 shows the category is uneven, and TensorDock sits with the cheaper end.
Core features
The core of the service is H100 SXM5 80GB, A100 SXM4 80GB and A100 PCIe 80GB, spanning datacentre SXM cards and workstation boards. That range is the feature: one provider can host both a training job and a lighter inference endpoint. Distribution comes from from 100 of locations, well above the 4 typical count, and every node is quoted with 1 Gbps minimum port included included. ComputePrices.com publishes a much longer catalogue at H100 SXM, A100 SXM and H200, so breadth alone is not the reason to pick TensorDock; the per-second meter and the absent egress bill are.
The company behind it
TensorDock is published as TensorDock.com, Inc and is headquartered in the US, so contracts and invoices sit with a named US entity rather than an anonymous reseller. On paper the security posture is SOC 2, PCI DSS and ISO 27001, a broad set that covers most enterprise procurement checklists. Support is listed as email, help centre and community, a self-service and asynchronous mix: quick for routine questions, thinner for a production incident at an unusual hour. Buyers who need a counterparty they can contract with directly have one.
How TensorDock compares
TensorDock vs BytePlus9.7 vs 8.7
TensorDock vs RunPod9.7 vs 9.5
TensorDock vs ScalaHosting9.7 vs 8.3
TensorDock vs Contabo9.7 vs 9.2
TensorDock vs Cloudways9.7 vs 9.2
TensorDock vs PlanetHoster9.7 vs 8.7
TensorDock vs Acronis9.7 vs 9.9
TensorDock vs Decodo9.7 vs 9.1
TensorDock vs Hostinger9.7 vs 9.1- TensorDock vs MarsProxies9.7 vs 9.0
TensorDock vs Proton Drive9.7 vs 8.3
TensorDock vs AIS Cloud9.7 vs 8.1
Frequently asked questions
What does an hour of GPU cost on TensorDock?+
GPU hours start at $0.10 on the entry plan, which sits below the category median of €0.63. Hourly is the only model listed under hourly, and each GPU is quoted separately rather than as a bundled node. Compute the cost against the work you intend to run, since the entry rate attaches to the smaller boards in the catalogue.
Is there an egress or data transfer fee?+
Data movement is No ingress/egress fees, so pulling a trained model or a dataset off the platform does not add a bill. That is the sharpest contrast with providers charging per gigabyte on top of the compute charge. It also means the GPU hour sits closer to the real total than it does elsewhere in the category.
How precisely am I billed for a GPU?+
Metering is per second, so a job that starts and stops pays only for the slice it used. This suits tests, spot checks and bursty inference, where an hour minimum is dead money. A long, steady training run bills the same way, so the advantage matters most when jobs are short or fragile.
How many locations does TensorDock run in?+
The stated footprint is from 100, comfortably above the category median of 4. That spread helps with latency and data residency, since a job can sit nearer its users or its inputs. Check the location list before committing, as capacity varies by model and region.
Which GPUs can I rent?+
The published list is H100 SXM5 80GB, A100 SXM4 80GB and A100 PCIe 80GB, running from datacentre SXM parts down to smaller RTX boards, so both training and lighter inference are covered. Blackwell parts do not appear in that list, which is the main gap against ComputePrices.com at H100 SXM, A100 SXM and H200. If you need the newest silicon, check that catalogue first.
Is TensorDock suitable for regulated workloads?+
The brand publishes SOC 2, PCI DSS and ISO 27001, a set that normally satisfies a corporate security review. The contracting entity is TensorDock.com, Inc, headquartered in the US, which helps when procurement needs a named counterparty. Teams in payments or healthcare can point to a named certification on the list rather than negotiating one from scratch.
How do I get help when a job fails?+
Support channels are email, help centre and community, meaning email, a help centre and community forums. That is quick enough for billing or provisioning questions, and thinner for a production incident that needs someone watching a screen. Teams needing a live phone line should weigh a provider whose support listing includes one.
What network speed comes with a node?+
Each machine is quoted with 1 Gbps minimum port included, which is the stated minimum rather than a peak figure. That is ample for streaming weights and pushing batches, and it removes the need to plan around a slow host link. It matters most for pipelines that move large artefacts in and out of the node between runs.
Sources
- docs.tensordock.com/legal-information/legal-information.md checked September 26, 2026
Backs: Headquarters, Legal entity
“TensorDock.com, Inc 16192 Coastal Highway Lewes, DE 19958”
- tensordock.com checked September 26, 2026
Backs: Egress fee, Locations, Pricing model, Support channels
“99.99% uptime standard and no ingress/egress fees”
- tensordock.com/cloud-gpus.html checked September 26, 2026
Backs: Billing granularity, GPU hour from, GPU models, Interconnect
“Price Per Hour, Billed Per-Second”
- tensordock.com/security.html checked September 26, 2026
Backs: Compliance certifications
“TensorDock aggregates GPUs from Tier 3/4 data centers with certifications ranging from SOC 2, PCI DSS, ISO 27001, and HIPAA”
Compiled from 4 source pages, checked September 26, 2026. We did not open an account.
Related by the facts
TensorDock
Short, per-second GPU bursts with no egress bill at the end of the month.