TensorDock GPU Marketplace: Pricing, GPU Types, and Provider Aggregation in 2026

TensorDock operates as a GPU marketplace aggregator rather than a direct infrastructure provider. The platform connects tenants to GPU compute from 12+ backend providers including CoreWeave, Vast.ai, RunPod, DataCrunch, and several smaller European and Asian GPU operators. TensorDock adds a unified API, standardized pricing, automated provider switching, and a reputation layer on top of the heterogeneous provider landscape. The platform supports H100, H200, A100, L40S, A10G, RTX 4090, RTX 6000 Ada, and several AMD GPU types.

TensorDock's key differentiator is automated multi-provider failover. If a provider's GPU becomes unavailable or performance degrades below a threshold, TensorDock automatically migrates the workload to another provider with available capacity. This failover works for stateless inference workloads but requires manual checkpointing for training jobs. The unified pricing model shows a single price per GPU type, with TensorDock taking a 10-20 percent margin on the backend provider's rate.

TensorDock's pricing sits in a distinct position: above the raw provider floor but below the cost of contracting with multiple providers individually. For H100, TensorDock's lower bound ($1.45/hr) is 2.6x Vast.ai's floor ($0.55/hr) but below AWS spot ($1.05-1.35/hr) and well below Lambda on-demand ($2.50/hr). The value proposition is avoiding the overhead of managing accounts across 8+ providers: unified billing, single API, and automated failover. For teams that would otherwise spend 5-15 hours/month managing multi-provider GPU infrastructure, the 15-20 percent margin is cost-effective.

For single-GPU workloads with low reliability requirements, going directly to Vast.ai's top providers offers better pricing ($0.55-0.68/hr for H100 versus $1.45 on TensorDock). For teams needing consistent access across multiple GPU types and automated provider failover, TensorDock's simplicity premium is justified. The pricing is competitive for workloads that require 100+ GPU-hours per week, where the management overhead savings offset the per-hour premium.

TensorDock provides a unified REST API and CLI that abstracts across all backend providers. The API accepts standard GPU configuration parameters (GPU type, count, region preference, max price, Docker image) and returns a provisioned instance URL within 30-180 seconds. The API supports persistent storage volumes, private networking between TensorDock instances, and automated health check configuration. A single API call can request 50 H100 GPUs and TensorDock distributes them across available backend providers, returning a list of provisioned instances with connection details.

The TensorDock CLI supports JSON output for programmatic consumption and integrates with common workflow tools (Slurm integration via plug-in, Kubernetes via a custom GPU device plugin, and Airflow/Prefect operators). Documentation covers migration from AWS, GCP, and RunPod with configuration templates. The developer experience bridges the gap between Vast.ai's raw marketplace interface and hyperscaler's managed services, offering intermediate-level abstraction that reduces multi-provider operational overhead.

TensorDock's automated failover provides measurable reliability improvements. With failover configured, the platform achieves 99.2 percent uptime for H100 instances over 30-day windows, versus 98.1 percent for individual provider reliability when averaged across all TensorDock's backends. The failover triggers on network connectivity loss, GPU temperature anomalies, and provider API unavailability. Average failover time is 45-120 seconds, sufficient to maintain stateless inference endpoints but causing training job interruption requiring checkpoint restart.

Regional availability spans 12 global locations through TensorDock's backend providers: US West, US East, Canada, UK, Germany, Netherlands, France, Poland, Singapore, Japan, South Korea, and Australia. This exceeds the geographic reach of any single backend provider and approaches AWS's regional breadth for GPU instances. The most constrained GPU type is H100 SXM, with 85 percent of available capacity in US and European data centers. Asia-Pacific H100 availability is limited to Singapore and Tokyo with 1-2 week wait times for 8+ GPU allocations.

TensorDock occupies a unique niche: an aggregation layer that reduces the overhead of multi-provider GPU procurement. It is not the cheapest option (Vast.ai direct is 30-50 percent cheaper for the lowest-tier providers), nor the most reliable (Lambda/CoreWeave offer 99.95%+ uptime on dedicated nodes), nor the most feature-rich (AWS offers 100+ managed services). It is the most flexible option for teams that value access to multiple GPU types and providers through a single interface.

The ideal TensorDock user is a team running 5-50 GPU instances across multiple providers, valuing the ability to switch GPU types and providers without renegotiating contracts or learning separate APIs. Teams running fewer than 5 GPUs can use direct provider platforms at lower cost. Teams running more than 50 GPUs should negotiate dedicated pricing with a single provider (Lambda, CoreWeave, or hyperscaler reservations).