TensorPool
TensorPool · Coding
TensorPool is a GPU cloud platform that rents B200, B300, H100, and H200 clusters on demand through a CLI, with git-style jobs, shared NFS storage, and per-minute billing.
About TensorPool
What Is TensorPool
TensorPool is a GPU cloud platform for renting GPU clusters by the hour. It targets ML engineers who train and fine-tune models but don't want to manage cloud accounts, quota requests, or spot interruptions by hand. You describe a job in a short config file, push it, and TensorPool finds capacity across providers and runs it.
The pitch is speed and price. TensorPool publishes transparent pricing, and its on-demand H100 SXM rate of $1.99 per hour undercuts typical hyperscaler rates by a wide margin. If you want to rent an H100 GPU or compare a B200 GPU rental against a hyperscaler, the pricing page lays the numbers out side by side. Capacity scales from a single GPU to clusters of hundreds, depending on the plan you use.
The main catch is that TensorPool is built around a developer workflow, not a click-through dashboard. You need comfort with a terminal and a container-based mindset. It also leans on spot and interruptible capacity in places, so long unattended runs carry some risk unless you use its recovery features.
Getting Started
- Sign up at tensorpool.dev with GitHub or email, then claim the $20 free credit for starring the GitHub repo.
- Install the TensorPool CLI and authenticate it with your account token.
- Write a job config that names the GPU type, the container image, and the command you want to run.
- Push the job with the CLI. TensorPool picks the cheapest available provider and starts the run.
- Pull logs and outputs back through the CLI, or stream them live while the job runs.
Product Information
A quick look at TensorPool's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- ML engineers
- Small research teams
- Startups on tight budgets
Tasks
- Fine-tuning large language models
- Running parallel experiments
- Batch inference and data processing
Scenarios
- Overnight training runs
- Running a GPU cloud for ML training without a contract
- Teams moving off hyperscalers to cut GPU spend
Key features
CLI-First GPU Jobs
TensorPool runs work through a command-line tool rather than a web console. You define a job in a config file, then push it the way you'd push a commit. It appeals to engineers who already live in a terminal and want GPU provisioning to fit existing scripts and CI pipelines.
Git-Style Workflows
Jobs behave like code changes. You submit a job, watch its state, and pull results back when it finishes. TensorPool bills only for runtime, so idle time between steps doesn't add to the bill. The model suits repeatable training pipelines more than one-off interactive sessions.
Multi-Provider Spot Recovery
Behind the CLI sits a scheduler that shops across GPU providers and picks the cheapest option that fits your job. When an interruptible instance gets preempted, smart spot recovery restarts the run on fresh hardware so a training job can continue without manual intervention.
nb.dev ML Notebook
nb.dev is TensorPool's take on Jupyter, built for ML work. It attaches a GPU to any notebook instantly, snapshots the full machine state including GPU memory and model weights, and branches that snapshot onto as many new machines as you want. That turns sequential trial-and-error into parallel experiments.
Shared NFS and Object Storage
TensorPool offers high-performance NFS shared storage plus an S3-compatible object store. The object tier is globally distributed and caches data near your GPUs, which the company says cuts cross-region latency versus Cloudflare R2. Shared storage keeps datasets and checkpoints reachable from every node in a job.
Container-Based Execution
Every job runs in a container, so you bring your own environment and dependencies. The setup avoids the version drift that plagues hand-built GPU servers, and it means a job that works on your laptop runs the same way on the cluster.
Transparent Per-Hour Pricing
The pricing page lists each GPU by hour and compares it directly against Lambda Labs and the big clouds. B200 sits at $4.99 per hour and L40S at $1.49, with CPU at $0.015. There's no hidden platform fee, and storage is billed separately per terabyte.
Elastic Scale Up to 128 GPUs
The standard plan covers up to 128 GPUs on demand, and enterprise agreements scale into the thousands with a 99.9% uptime SLA. Teams can start small on a single card and grow the same workflow into a large cluster without switching platforms.
Pros and cons
Pros
- On-demand H100 pricing around $1.99 per hour undercuts most hyperscalers for comparable hardware.
- The CLI and git-style jobs fit existing engineering workflows instead of forcing a new dashboard.
- Spot recovery restarts interrupted jobs automatically, which reduces lost work on long training runs.
- Free $20 credit and no monthly minimum make it cheap to test before committing.
- Shared NFS and S3-compatible storage keep datasets and checkpoints available across all nodes.
Cons
- There's no rich graphical console for job management, so non-technical users will struggle.
- Committed discounts and large-scale reservations require contacting sales, with no public list price.
- The platform depends on spot capacity in some regions, which adds variability to start times and reliability.
- Documentation and support live mainly in Slack and Discord, so answers aren't always searchable.
Frequently asked questions
TensorPool rents GPU clusters for ML training, fine-tuning, and inference. You submit containerized jobs through a CLI and it handles provisioning across providers, so you don't manage cloud accounts yourself.
Related content
Explore related tools, skills, and articles for TensorPool.
TensorPool Alternatives
Udon
Udon · Coding · ProductivityUdon turns an Apple Silicon Mac into a home server you run from a browser, with container controls, file shares, and an optional AI sysadmin.
offstage
offstage cloud · Coding · Productivityoffstage gives AI coding agents and CI jobs a real Mac: a fresh macOS account on Apple silicon, driven over an API and billed by the second.
BotBus
Chengdu Tuyang Tusen Technology Co., Ltd. · Coding · ProductivityBotBus lets you control AI coding agents from your phone and watch: approve commands, start tasks and read results while the code stays on your desktop.