llmcost.xyzTensorDock vs Vast.ai: GPU Marketplace Pricing Compared

TensorDock vs Vast.ai: GPU Marketplace Pricing Compared

Two GPU marketplaces compared on pricing, GPU availability, and billing — see which fits your workload and budget.

📅 Data verified: 2026-06-13

TensorDock and Vast.ai are both GPU marketplaces, but they package supply differently: TensorDock combines marketplace capacity with partner data centers, while Vast.ai is a decentralized bid-based marketplace. Pricing is aggressive on both sides, with H100s starting around $2.25/hr on TensorDock and about $1.87/hr on Vast.ai, while RTX 4090 listings start near $0.35/hr and $0.20/hr respectively.

FeatureTensorDockVast.ai
Pricing ModelMarketplace + partner data centers (incl. Voltage Park)Decentralized GPU marketplace (bid-based)
BillingPer-hour (per-minute on some hosts)Per-minute
H100 Pricingfrom $2.25/hr (on-demand ~$2.50, spot ~$1.60)from ~$1.87/hr (PCIe)
A100 80GB Pricingfrom ~$1.30/hrfrom ~$0.67/hr
RTX 4090 Pricingfrom ~$0.35/hrfrom ~$0.20/hr
Founded20202018

💡 Pricing reflects publicly listed on-demand rates as of 2026-06-13 and may vary by region, GPU availability, or change over time. Always confirm current pricing on the provider's site before committing.

Why Choose Each Provider

TensorDock

  • Competitive H100/A100 pricing close to marketplace lows
  • Spot pricing for further discounts on interruptible jobs
  • Simple self-service deployment, pay-as-you-go
  • Smaller company than hyperscalers — less brand track record
  • Availability and exact pricing vary by region/host

Vast.ai

  • Often the cheapest marketplace prices for any given GPU
  • Huge GPU variety from thousands of independent hosts
  • Interruptible/spot instances for extra savings
  • Reliability and network speed vary by host
  • No enterprise SLA or guaranteed uptime
  • Prices fluctuate live with marketplace supply/demand

Verdict

For self-service on-demand workloads where you want straightforward deployment and still-competitive pricing, TensorDock is the better pick because its H100 pricing starts around $2.25/hr, with on-demand closer to $2.50/hr and spot around $1.60/hr, and it also offers strong A100 80GB pricing from about $1.30/hr. For bargain-hunting solo researchers or teams running flexible, interruptible experiments, Vast.ai is the better pick because its marketplace often posts lower starting prices — about $1.87/hr for H100 PCIe, $0.67/hr for A100 80GB, and $0.20/hr for RTX 4090 — though you trade that for host-to-host variability and no enterprise SLA.

FAQ

Is TensorDock cheaper than Vast.ai for H100 rentals?

Usually no at the absolute floor price: Vast.ai H100 PCIe listings start around $1.87/hr, while TensorDock H100 starts around $2.25/hr. That said, TensorDock spot can drop to about $1.60/hr on interruptible capacity, so the cheaper option can depend on whether you are comparing on-demand versus spot and what hosts are available in your region.

Which is better for short training jobs: TensorDock or Vast.ai?

Both can work well for short jobs because billing is granular: Vast.ai bills per-minute, and TensorDock is per-hour with per-minute billing on some hosts. If your priority is the lowest possible entry price, Vast.ai is attractive with A100 80GB from about $0.67/hr and RTX 4090 from about $0.20/hr; if you want simpler self-service deployment with competitive pricing, TensorDock starts around $1.30/hr for A100 80GB and $0.35/hr for RTX 4090.

Does Vast.ai support interruptible or spot-style GPU instances like TensorDock?

Yes. Vast.ai offers interruptible/spot-style marketplace instances for extra savings, similar in spirit to TensorDock’s spot options. TensorDock explicitly lists H100 spot pricing around $1.60/hr versus roughly $2.50/hr on-demand, while Vast.ai’s live marketplace pricing fluctuates with supply and demand and can be very low, but reliability and uptime vary by host.

Check egress before committing

In 2026, the biggest surprise cost on GPU marketplaces is often network egress, not hourly GPU price. Vast and TensorDock listings can look nearly identical on $/GPU-hour, but a multi-day fine-tune or checkpoint sync can become more expensive on the host with stricter outbound bandwidth caps or paid public IPs. Before renting, estimate how many GB of datasets, model weights, checkpoints, and logs will leave the machine, then compare total run cost including storage persistence and egress fees. Cheapest hourly usually loses if you move data often.