Thunder Compute Pricing: GPU Instances & Sandboxes — Full 2026 Guide
VerifiedThunder Compute GPU pricing: RTX A6000 from $0.35/hr, A100 80GB from $1.09/hr, H100 SXM $3.20/hr. Sandboxes from $1.95/GPU-hr, billed per second.
Categories & Tags
Thunder Compute Review 2026
Thunder Compute is a cloud GPU rental platform that positions itself as the 'World's Cheapest GPUs,' offering one-click GPU servers at rates it claims are roughly 80% less than AWS for comparable hardware. The platform is the product of a systems lab that spent four years in stealth productionizing GPU virtualization research, with a team that includes alumni from Citadel Securities, Aquatic, and AWS. For developers, data scientists, and startups, the core value proposition is straightforward: access to high-end GPUs like the H100 and A100 on a pure usage-based, per-minute billing model without data egress fees.
Quick Navigation
Quick Summary
Overall Rating 4.3/5 Best For ML engineers and researchers needing affordable, on-demand GPU instances
What Is Thunder Compute and Why Does It Matter?
Thunder Compute positions itself as a GPU cloud built for developers, offering GPU instances and sandboxes for AI workloads such as evals, RL environments, and post-training. Its strategic role is to make GPU compute cheaper and simpler, as stated on its About page. The site emphasizes two core products: persistent Instances that users can SSH into and snapshot, and short-lived Sandboxes for parallel rollouts, both running on the same GPUs. Pricing transparency is central: Instances start at $0.35/hr and are billed per minute, while Sandboxes start at $1.95/GPU-hr and are billed per second. A comparison table shows Thunder Compute's H100 80GB at $3.20 per GPU-hour versus higher on-demand list prices from Lambda, Runpod, Paperspace, CoreWeave, AWS, Azure, Oracle Cloud, and Google Cloud, checked Oct 8, 2026. The company also offers reserved capacity for 8 to 512+ GPUs and states no egress fees. Testimonials from ML engineers and researchers highlight UX, savings, and snapshot flow.
Who Should Use Thunder Compute?
- ML engineers running post-training and fine-tuning Thunder Compute offers persistent GPU instances with RTX A6000, L40, A100 80GB, and H100 SXM options, ready to SSH into in minutes, with snapshots to keep your disk between runs. Instances are billed per minute, and the site cites a testimonial from a bootstrapped data science consultancy running 2-3 finetuning jobs around the clock.
- RL researchers and eval engineers Sandboxes are isolated, short-lived CPU or GPU environments for agents, letting you fan out hundreds of parallel rollouts from a few lines of Python. They scale up and down to 0 and are billed per second for each resource, with H100 SXM at $2.95/GPU-hr and A100 80GB at $1.95/GPU-hr.
- Cost-conscious startup teams The pricing page compares Thunder Compute's H100 80GB rate of $3.20/GPU-hr against on-demand list prices from Lambda ($3.29), Runpod ($3.49), Paperspace ($5.95), CoreWeave ($6.16), AWS ($6.88), Azure ($6.98), Oracle Cloud ($10.00), and Google Cloud ($10.98), checked Oct 8, 2026. The homepage claims switching saves up to 68% vs AWS.
- Developers who prefer CLI and IDE workflows Instances can be launched from your terminal (e.g. 'tnr create --gpu a100' then 'tnr connect 0'), from VS Code, or from your browser. The site lists a VS Code extension, CLI, MCP server, templates, and REST API under Developers, and a testimonial from a UC Berkeley ML PhD cites the VS Code extension for faster iteration.
Professional reality: Sandbox access requires contacting Thunder Compute to request access, and reserved capacity for 8 to 512+ dedicated GPUs is only available by talking to sales, so teams needing immediate self-serve sandbox or reserved-cluster provisioning may face a wait.
Thunder Compute Features That Drive Results
Persistent GPU instances you SSH into in minutes
Spin up RTX A6000, L40, A100 80GB, or H100 SXM GPUs and connect from your terminal, VS Code, or the browser. Create with a single command like tnr create --gpu a100, then tnr connect to get an SSH session. Snapshot the instance when you are done.
Hands-on development on a GPU machine you keep, ready to SSH into in minutes.
Isolated, short-lived CPU or GPU sandboxes for agents
Fan out hundreds of parallel rollouts from a few lines of Python using the thunder_sandbox library. Configure cpu, memory, storage, gpu_type, gpu_count, and timeout per sandbox, then wait_until_ready and exec commands such as nvidia-smi. Sandboxes scale up and down to 0 and are billed per second.
RL environments and evals at scale, paying exactly when workloads are running.
Per-minute instances and per-second sandboxes
Instances start from $0.35/hr for RTX A6000, $0.79/hr for L40, $1.09/hr for A100 80GB, and $3.20/hr for H100 SXM, billed by the minute. Sandboxes are billed per second: H100 SXM at $2.95/GPU-hr and A100 80GB at $1.95/GPU-hr, plus vCPU, memory, and storage rates. No egress fees.
Switch now and save up to 68% vs AWS on comparable GPU hours.
Snapshots to keep your disk between runs
The first 100 GB of instance disk per GPU is included and billed only while running. Snapshots save an instance disk and are kept until you delete them at $0.05/GB/mo. Data egress is $0.00/GB.
Pause experiments without cloud housekeeping or losing your disk state.
Reserve clusters of 8 to 512+ dedicated GPUs
Reserve dedicated GPUs for a fixed term at a lower rate than on-demand, with contract and volume pricing for teams. Reserved rates include H100 SXM at $3.20/GPU-hr, A100 80GB at $1.49/GPU-hr, and L40 at $0.99/GPU-hr.
Predictable capacity and lower rates for teams running sustained workloads.
CLI, VS Code extension, MCP server, and REST API
Start from your terminal or from your agent. Thunder Compute ships a CLI, a VS Code extension, an MCP server, templates, and a REST API, alongside docs, guides, and quickstarts.
Fit GPU compute into the workflow your team already uses.
Thunder Compute Pricing in 2026
Thunder Compute offers GPU instances billed per minute and sandboxes billed per second. Instance pricing includes RTX A6000 at $0.35/hr, L40 at $0.79/hr, A100 80GB at $1.09/hr, and H100 SXM at $3.20/hr. Extra vCPUs cost $0.04/vCPU-hr. Instance disk is $0.219/GB/mo, snapshots $0.05/GB/mo, and data egress is free. Sandboxes cost $2.95/GPU-hr for H100 SXM, $1.95/GPU-hr for A100 80GB, $0.051/vCPU-hr, $0.016/GiB-hr for memory, and $0.00012/GiB-hr for storage. Reserved capacity offers H100 SXM at $3.20/GPU-hr, A100 80GB at $1.49/GPU-hr, and L40 at $0.99/GPU-hr for 8–512+ GPUs.
| Plan | Price | What You Get |
|---|---|---|
| RTX A6000 Instance | $0.35/hr | 48 GB VRAM, 6 included vCPUs (8 max), 48–64 GB RAM, 100–500 GB disk |
| L40 Instance | $0.79/hr | 48 GB VRAM, 6 included vCPUs (12 max), 48–96 GB RAM, 100–500 GB disk |
| A100 80GB Instance | $1.09/hr | 80 GB VRAM, 8 included vCPUs (16 max), 64–128 GB RAM, 100–500 GB disk |
| H100 SXM Instance | $3.20/hr | 80 GB VRAM, 4 included vCPUs (16 max), 32–128 GB RAM, 100–500 GB disk |
| H100 SXM Sandbox | $2.95/GPU-hr | 80 GB VRAM, one GPU per sandbox, billed per second |
| A100 80GB Sandbox | $1.95/GPU-hr | 80 GB VRAM, one GPU per sandbox, billed per second |
| Reserved H100 SXM | $3.20/GPU-hr | 8–512+ GPUs, reserved capacity |
| Reserved A100 80GB | $1.49/GPU-hr | 8–512+ GPUs, reserved capacity |
| Reserved L40 | $0.99/GPU-hr | 8–512+ GPUs, reserved capacity |
Visit the official Thunder Compute website to check the latest pricing and plans.
Where Thunder Compute Is Strong / Where It Needs Care
- Transparent, low on-demand GPU pricingThunder Compute publishes per-GPU-hour rates for its instances: RTX A6000 at $0.35/hr, L40 at $0.79/hr, A100 80GB at $1.09/hr, and H100 SXM at $3.20/hr. Instances are billed by the minute, and the site shows a comparison table of H100 80GB on-demand list prices where Thunder Compute is listed at $3.20/GPU-hr versus Lambda $3.29, Runpod $3.49, Paperspace $5.95, CoreWeave $6.16, AWS $6.88, Azure $6.98, Oracle Cloud $10.00, and Google Cloud $10.98 (rates checked Sep 30, 2026; on-demand list prices checked Oct 8, 2026).
- Sandboxes for agents, RL environments, and evalsThunder Compute offers isolated, short-lived CPU or GPU sandboxes designed for RL environments and evals, with the ability to fan out hundreds of parallel rollouts from a few lines of Python. Sandboxes scale up and down to 0 and are billed per second for each resource. Published sandbox rates include H100 SXM at $2.95/GPU-hr ($0.000819/GPU-sec), A100 80GB at $1.95/GPU-hr ($0.000542/GPU-sec), vCPU at $0.051/vCPU-hr, memory at $0.016/GiB-hr, and storage at $0.00012/GiB-hr.
- No egress fees and included instance diskThe pricing page states data egress has no fees at $0.00/GB. For instances, the first 100 GB of disk per GPU is included and billed only while running; additional instance disk is $0.219/GB/mo, and snapshots are $0.05/GB/mo. Snapshots let users keep their disk between runs.
- Reserved capacity for larger clustersThunder Compute offers reserved capacity for clusters of 8 to 512+ dedicated GPUs at a lower rate than on-demand, with contract and volume pricing. Listed reserved rates include H100 SXM at $3.20/GPU-hr, A100 80GB at $1.49/GPU-hr, and L40 at $0.99/GPU-hr; RTX A6000 is noted as not offered for reservation (1× only).
- Pricing comparison is based on third-party list pricesThe H100 80GB comparison table is explicitly labeled as on-demand list prices checked Oct 8, 2026, with rates checked Sep 30, 2026. These are third-party list prices shown for comparison and may not reflect negotiated, spot, reserved, or current pricing from those providers.
- Sandbox access requires contacting Thunder ComputeThe pricing page states that CPU-only sandboxes pay just the vCPU, memory, and storage rates, and that users should contact Thunder Compute for access. Sandbox access is not shown as instantly self-serve on the pricing page.
- Reserved capacity requires sales contactReserved capacity is described as clusters of 8 to 512+ dedicated GPUs for a fixed term at a lower rate than on-demand, with contract and volume pricing for teams. Each reserved GPU row lists 'Contact sales' rather than an instant checkout, and RTX A6000 is not offered for reservation.
- Instance and sandbox billing meters differInstances are billed per minute, while sandboxes are billed per second for each resource. Instance disk is billed only while running, and sandbox storage is billed only while running and wiped when the sandbox ends. Snapshots are billed separately at $0.05/GB/mo.
Real-World Use Cases
Post-Training and Fine-Tuning Jobs
Thunder Compute positions its GPU instances for post-training workloads. Spin up an RTX A6000, L40, A100 80GB, or H100 SXM instance, SSH in within minutes, and keep it running for the duration of the job. One customer quoted on the site describes running 2-3 finetuning jobs around the clock as a bootstrapped consultancy. Instances are billed per minute, and snapshots let you pause experiments without losing your disk. If your pipeline needs to move training data around first, tools like Airbyte or Fivetran handle the ingestion layer separately.
RL Environments and Parallel Rollouts
Thunder Compute's Sandboxes are built for RL environments and evals. From a few lines of Python you can fan out hundreds of parallel rollouts, each in an isolated, short-lived environment with or without a GPU. Sandboxes scale up and down to zero and are billed per second for each resource, so you pay only while workloads run. A100 80GB sandboxes run $1.95/GPU-hr and H100 SXM sandboxes run $2.95/GPU-hr. Teams running large sweeps can pair this with experiment tracking and dataset tooling such as Hugging Face Datasets.
Agent Evaluation and Benchmarking
The site describes Thunder Compute as 'GPUs for agents,' with sandboxes designed for evals at scale. Each sandbox is an isolated, short-lived CPU or GPU environment you can create programmatically, wait until ready, and exec commands against. CPU-only sandboxes pay just the vCPU, memory, and storage rates, so lightweight agent evals do not need a GPU at all. For teams building evaluation harnesses that need structured data in and out, complementary tools like Databricks or Snowflake can sit alongside the compute layer.
Cost-Conscious GPU Development and Sweeps
Thunder Compute publishes on-demand list prices checked against Lambda, Runpod, Paperspace, CoreWeave, AWS, Azure, Oracle Cloud, and Google Cloud, claiming savings up to 68% vs AWS. Instances start at $0.35/hr for an RTX A6000 and $1.09/hr for an A100 80GB, billed per minute, with no egress fees and the first 100 GB of instance disk per GPU included. The site's own estimate tool shows a single A100 80GB running 176 hours a month at $235.56/mo versus $603.84 for the same GPU hours on AWS. Reserved capacity for 8 to 512+ dedicated GPUs is available for teams that need committed clusters.
How to Get Started With Thunder Compute
Create an account on the Thunder Compute website and review the available GPU options and pricing on the /pricing page.
Use the web console or CLI tool to launch a new instance, selecting your desired GPU type (e.g., A100 or H100) and base configuration.
Connect to your instance using the VS Code extension or SSH via the CLI to begin setting up your environment and running workloads.
For persistent work, create a snapshot of your instance before pausing it. You can restore this exact environment later, paying only for snapshot storage in the interim.
Is Thunder Compute Worth It in 2026?
For ML engineers, researchers, and startups where compute cost is a primary constraint, Thunder Compute is a compelling option in 2026. Its value is highest for non-mission-critical workloads like experimentation, fine-tuning, and prototyping, where the significant cost savings directly translate into more compute for the same budget. The main limitation is the lack of a flat-fee plan, which makes budgeting less predictable, and the relative newness of the company compared to hyperscalers. If you can manage variable costs and your workloads are not business-critical, it is worth serious consideration.
Thunder Compute vs the Competition
| Decision Area | Thunder Compute | When Another Option Wins |
|---|---|---|
| H100 80GB price per GPU-hour | $3.20 per GPU-hour on-demand, billed by the minute for instances and per second for sandboxes | Lambda lists $3.29, Runpod $3.49, Paperspace $5.95, CoreWeave $6.16, AWS $6.88, Azure $6.98, Oracle Cloud $10.00 and Google Cloud $10.98 per GPU-hour — all higher on-demand list prices checked Oct 8, 2026 |
| Entry-level GPU instance price | RTX A6000 48 GB from $0.35/hr, L40 48 GB from $0.79/hr, A100 80GB from $1.09/hr, H100 SXM 80 GB from $3.20/hr | Thunder Compute's published comparison table only covers H100 80GB pricing, so no like-for-like entry-level rates for the other providers are shown on this site |
| Billing granularity | Instances billed per minute; sandboxes billed per second for each resource and scale down to 0 | The site does not publish the billing granularity of Lambda, Runpod, Paperspace, CoreWeave, AWS, Azure, Oracle Cloud or Google Cloud |
| Data egress | No egress fees — $0.00/GB for instances and sandboxes | Egress pricing for the compared providers is not shown on this site |
| Reserved capacity | Reserve 8 to 512+ dedicated GPUs for a fixed term at a lower rate: H100 SXM $3.20/GPU-hr, A100 80GB $1.49/GPU-hr, L40 $0.99/GPU-hr; RTX A6000 not offered (1x only) | Reserved or contract pricing for the compared providers is not published on this site |
Thunder Compute vs Lambda
Thunder Compute's own pricing table lists Lambda at $3.29 per H100 80GB GPU-hour versus Thunder Compute at $3.20, with both figures described as on-demand list prices checked Oct 8, 2026. Thunder Compute also publishes instance rates from $0.35/hr for RTX A6000, $0.79/hr for L40 and $1.09/hr for A100 80GB, plus per-second billed sandboxes from $1.95/GPU-hr and no egress fees.
Choose Thunder Compute if: You want a lower published H100 80GB on-demand rate, per-minute instance billing, per-second sandbox billing that scales to 0, snapshots to keep your disk between runs, and no egress fees. Choose Lambda if: You are comparing on-demand H100 80GB list pricing and prefer Lambda's published rate of $3.29 per GPU-hour.
Thunder Compute vs Runpod
Runpod appears in Thunder Compute's H100 80GB price comparison at $3.49 per GPU-hour against Thunder Compute's $3.20, based on on-demand list prices checked Oct 8, 2026. Thunder Compute positions its instances for hands-on development you SSH into and keep, and its sandboxes for RL environments and evals with hundreds of parallel rollouts from a few lines of Python.
Choose Thunder Compute if: You want the lower published H100 80GB rate plus sandboxes for RL environments and evals, with per-second billing and no egress fees. Choose Runpod if: You are evaluating H100 80GB on-demand list pricing and Runpod's published $3.49 per GPU-hour fits your comparison.
Frequently Asked Questions
What GPUs can I run on Thunder Compute?
Thunder Compute offers RTX A6000 (48 GB), L40 (48 GB), A100 80GB, and H100 SXM (80 GB) GPUs. Instances are ready to SSH into in minutes, and you can launch them from your terminal, VS Code, the CLI, or your browser.
How much does an H100 cost on Thunder Compute compared to other clouds?
Thunder Compute lists H100 80GB at $3.20 per GPU-hour. For comparison, the same page shows Lambda at $3.29, Runpod at $3.49, Paperspace at $5.95, CoreWeave at $6.16, AWS at $6.88, Azure at $6.98, Oracle Cloud at $10.00, and Google Cloud at $10.98 per GPU-hour (on-demand list prices, checked Oct 8, 2026).
What are Sandboxes and how are they billed?
Sandboxes are isolated, short-lived CPU or GPU environments for agents, RL environments, and evals. They scale up and down to 0, are billed per second for each resource, and storage is wiped when the sandbox ends. H100 SXM sandboxes cost $2.95/GPU-hr and A100 80GB sandboxes cost $1.95/GPU-hr.
How does Thunder Compute handle storage and data egress?
Instance disks include the first 100 GB per GPU and are billed only while running at $0.219/GB/mo. Snapshots cost $0.05/GB/mo and are kept until you delete them. Data egress has no fees at $0.00/GB.
Can I reserve GPU capacity for a fixed term?
Yes. Reserved capacity lets you reserve clusters of 8 to 512+ dedicated GPUs for a fixed term at a lower rate than on-demand, with contract and volume pricing for teams. Reserved rates include H100 SXM at $3.20/GPU-hr, A100 80GB at $1.49/GPU-hr, and L40 at $0.99/GPU-hr.
Key Takeaways
- Thunder Compute is best for ML engineers and researchers who need affordable, on-demand GPU instances without hyperscaler pricing.
- Pricing is pure usage-based, starting at $0.35/GPU/hr, with no free plan or subscription options.
- Biggest strength is its aggressive cost leadership; main limitation is the lack of a flat-fee plan and its newer market presence.
Best Thunder Compute Alternatives
- RunPod — Choose RunPod if you need a more mature GPU cloud platform with a broader ecosystem and a longer track record of reliability.
- Vast.ai — Choose Vast.ai if your primary concern is finding the absolute lowest price on a marketplace, even if it means variable performance.
- Lambda — Choose Lambda if you prefer a more established AI cloud provider with a strong reputation for large-scale deployments.
Bottom Line: For teams where GPU cost is the primary bottleneck, Thunder Compute is a compelling and strategically sound choice in 2026, offering unmatched value for flexible, non-mission-critical workloads.
Reviewed by theaitoolsbox.com editorial team
Quick Answers
What does Thunder Compute cost?
Instances from $0.35/hr; Sandboxes from $1.95/GPU-hr; Reserved from $0.99/GPU-hr
Who is Thunder Compute best for?
Thunder Compute is best for ML engineers and researchers needing affordable, on-demand GPU instances.
How do you get started with Thunder Compute?
You create a Thunder Compute account, then use the web console or CLI to launch an instance by selecting a GPU type such as an A100 or H100. You connect via the VS Code extension or SSH to set up your environment and run workloads.
Thunder Compute
AI Data Processing Tools
Pricing Plans
Paid
Check website for details
RTX A6000 Instance
$0.35/hr48 GB VRAM, 6 included vCPUs (8 max), 48–64 GB RAM, 100–500 GB disk
L40 Instance
$0.79/hr48 GB VRAM, 6 included vCPUs (12 max), 48–96 GB RAM, 100–500 GB disk
A100 80GB Instance
$1.09/hr80 GB VRAM, 8 included vCPUs (16 max), 64–128 GB RAM, 100–500 GB disk
H100 SXM Instance
$3.20/hr80 GB VRAM, 4 included vCPUs (16 max), 32–128 GB RAM, 100–500 GB disk
H100 SXM Sandbox
$2.95/GPU-hr80 GB VRAM, one GPU per sandbox, billed per second
A100 80GB Sandbox
$1.95/GPU-hr80 GB VRAM, one GPU per sandbox, billed per second
Reserved H100 SXM
$3.20/GPU-hr8–512+ GPUs, reserved capacity
Reserved A100 80GB
$1.49/GPU-hr8–512+ GPUs, reserved capacity
Reserved L40
$0.99/GPU-hr8–512+ GPUs, reserved capacity