TensorDock

What TensorDock actually is

TensorDock is a GPU compute marketplace rather than a conventional cloud. Independent hosts list GPU machines — some colocated in tier 3/4 data centers, some converted mining rigs in mining farms, in TensorDock's own description — and TensorDock handles the billing, vetting, and customer-facing platform on top. The company was founded in 2021 by Jonathan Lei, is headquartered in Boston, Massachusetts, and was acquired by GPU-as-a-service provider Voltage Park in March 2025; the marketplace continues to operate under its own brand, with a new general manager from Voltage Park's team. That last fact matters for anyone signing a long-term arrangement: the vendor you evaluate today has a parent with deeper infrastructure pockets than a typical startup marketplace.

The scale claims are marketplace-shaped: up to 30,000 GPUs through partners, 45 GPU models, and 100-plus locations in more than 20 countries, spanning consumer cards (RTX 3090, RTX 4090), professional (RTX A6000, RTX 6000 Ada, L40), and datacenter accelerators (V100, A100 SXM and PCIe, H100 SXM5). Everything is KVM virtualisation with root access, and there is an API for programmatic provisioning. TensorDock says every host is vetted before listing, that SSH access from hosts to customer machines is revoked, and that hosts are held to a 99.99% uptime standard with maintenance scheduled at least two weeks ahead.

The model has one important consequence that the pricing page states plainly: the same GPU can have multiple prices because independent hosts set their own rates and differ in location and redundancy. In our view that is the right frame for everything else in this review. You are shopping a market, not buying a standardised product, and the variability — good and bad — flows from that.

Pricing: what you really pay

TensorDock publishes a rate card, which is more than several GPU competitors do. The on-demand GPU table (the page carries a last-updated stamp of 24 July 2024 and a disclaimer that TensorDock guarantees nothing on it) lists typical hourly prices, resources not included: H100 SXM5 80 GB at $2.85, A100 SXM4 80 GB at $1.80, A100 PCIe 80 GB at $1.50, L40 48 GB at $0.95, RTX 6000 Ada 48 GB at $0.75, RTX A6000 48 GB at $0.45, RTX 4090 24 GB at $0.35, RTX 3090 24 GB at $0.20, V100 SXM2 16 GB at $0.17, and RTX A4000 16 GB at $0.10. The headline hero prices on the same page are H100 SXM5 from $2.80 per hour, A100 SXM4 from $1.80, and RTX 4090 from $0.35.

Resources are billed à la carte per second: vCPUs at $0.003 per hour each (two-vCPU minimum at $0.006), RAM at $0.002 per GB per hour (4 GB minimum), and block NVMe storage at $0.00005 per GB per hour (20 GB minimum at $0.0010). A 1 Gbps network port is included. Billing is pay-as-you-go against a prepaid balance: you deposit funds, the balance is deducted continuously while a server runs, and — read this carefully — when the balance reaches zero your servers are automatically deleted. Reserved pricing for long-term rentals is negotiated rather than published.

Two pricing observations. First, the à-la-carte model is genuinely cheap at the low end and unusually legible: a minimal RTX 4090 box works out to well under a dollar an hour all-in, which is where TensorDock earns its reputation. Second, the balance-deletion policy is a footgun for anyone used to hyperscaler invoicing. Use alerts, keep a buffer, and checkpoint training jobs; a card expiring should not mean a 20-hour run disappears. TensorDock's own comparison table claims an 8x H100 server at $29.31 per hour against AWS at $98.32 and Google Cloud at $88.49 — vendor-published numbers we would treat as directional rather than verified.

Performance & infrastructure

The infrastructure story is a spectrum, and TensorDock is unusually candid about it. The company's about page describes both ends: hardware colocated in tier 3/4 data centers, and converted mining rigs hosted by mining farms, with suppliers competing on price. Third-party reviews add the operational detail — location tiers where a lower tier can mean a facility without redundant power — which matches TensorDock's own explanation that hosts differ in redundancy and therefore cost. For production inference serving, that variance is the whole risk: pick hosts and locations deliberately, favour the datacenter-tier listings, and treat the cheapest option as a dev or batch tier.

On raw capability, the marketplace is broad. H100 SXM5 80 GB and A100 SXM4 80 GB cover serious training and large-model inference; the consumer cards are where the price-performance is extreme — TensorDock argues RTX 4090 and 3090 deliver up to 5x inference value, and the published rates ($0.35 and $0.20 per hour) support the arithmetic for workloads that fit in consumer VRAM. Because configuration is à la carte, you can size CPU, RAM, and storage around the accelerator instead of buying a fixed instance shape, and the API exposes metadata and availability for automation.

We have not benchmarked TensorDock hosts ourselves, and given the marketplace model a single number would not generalise anyway. Independent testers report H100 training throughput close to theoretical on datacenter-tier hosts with weaker figures on community hosts — directionally consistent with what the model predicts. The honest summary: the silicon is the same silicon you get anywhere; what varies is the network, the redundancy, and what happens when something breaks at 3 a.m.

Where it falls short

Support is the clearest weakness. Third-party testing reports email-only support with multi-hour response windows, and there is no phone line or chat. For a platform where hosts are independent and incidents can span two organisations, that is thin. TensorDock's answer is host vetting and the 99.99% uptime standard with removal from the platform as the enforcement mechanism — a marketplace discipline, not a contractual SLA with credits behind it. Procurement teams should price that gap honestly rather than comparing the 99.99% figure to hyperscaler SLAs.

Second, operational polish. The dashboard and API are functional but leaner than RunPod's or Lambda's; there are fewer prebuilt ML environment templates; spot instances are relatively new; and the prepaid-balance model with automatic server deletion is a world away from AWS invoicing. The published rate card itself carries a disclaimer that it may be stale — the numbers above were last stamped July 2024, so verify in the dashboard before budgeting.

Third, consistency of host quality is structural, not incidental. Vetted does not mean uniform. Redundancy, network quality, and maintenance practice vary by host, and the cheapest listings are often the least redundant. Teams running production services need to treat host and location selection as configuration work — pin known-good hosts, keep images and checkpoints external, and assume any single host can vanish.

Who should pick this

Our pick for cost-conscious ML engineers, researchers, and studios: training jobs with checkpointing, batch inference, rendering, and experimentation where the budget question is GPU-hours per dollar. The consumer-card pricing is the standout — if your models fit in 24-48 GB of VRAM, an RTX 4090 at $0.35 per hour plus pocket-change resources is hard to argue with, and the à-la-carte resource billing means you are not paying for a fixed instance shape you do not need. The breadth (45 models, 100-plus locations) also helps when you need a specific card in a specific region on short notice.

We would not pick TensorDock for latency-critical production serving where you need a contractual SLA, named support, and uniform hardware — that is what CoreWeave, Lambda, or a hyperscaler sells, at a higher price. If you are comparing the GPU cloud market broadly, our GPU cloud category lines up the alternatives, and our benchmarks page explains how we think about measuring this class of product.

Key takeaways

• GPU marketplace model: 45 GPU models across 100-plus locations in 20-plus countries, mixing tier 3/4 datacenter hardware with cheaper community hosts.
• Published rate card (July 2024): H100 SXM5 $2.85/hr typical, A100 SXM4 $1.80/hr, RTX 4090 $0.35/hr, RTX A4000 $0.10/hr — resources billed separately per second.
• À-la-carte resources: vCPU $0.003/hr, RAM $0.002/GB/hr, NVMe $0.00005/GB/hr, 1 Gbps included; prepaid balance, and servers are deleted when the balance hits zero.
• Hosts are vetted and held to a 99.99% uptime standard, but that is marketplace discipline — no contractual SLA with credits.
• Support is email-only; host quality and redundancy vary by listing, so production use requires deliberate host selection.
• Founded 2021 in Boston by Jonathan Lei; acquired by Voltage Park in March 2025, marketplace still operating independently.

Sources

• https://tensordock.com/cloud-gpus
• https://tensordock.com/about
• https://cbinsights.com/investor/tensordock
• https://gxcom.net/tensordock-review
• https://pricegpu.com/provider/tensordock
• https://gpucloudlist.com/en/blog/tensordock-review-cheapest-h100
• https://vpsrated.com/provider/tensordock