Hedra GPU Sandboxes

Product · Developer tool by Hedra

Isolated GPU environments for custom training and evaluation code.

Hedra GPU Sandboxes offer isolated, high-performance GPU environments for custom training and evaluation. They provide developers with full runtime control on frontier hardware like NVIDIA B200 and AMD MI355X with granular per-second billing.

Launched
2026-11-18

Category: AI Infrastructure

Platforms
Web, API
Editors and tools
CLI, MCP
Deployment
Cloud-based isolated GPU environments
Billing Increment
Per GPU-second
Availability Status
Research Preview
Supported Hardware
NVIDIA B300, B200, H100, RTX 4090, RTX PRO 6000, AMD MI355X
  • Isolated GPU environments with full runtime control
  • Frontier hardware including NVIDIA B300 and AMD MI355X
  • Granular billing calculated per GPU-second
  • Designed for custom training and evaluation pipelines

Overview

Hedra GPU Sandboxes provide isolated, high-performance GPU environments for businesses and developers to run custom training and evaluation code. The service allows users to control their own runtime while leveraging Hedra's infrastructure, which includes frontier hardware such as NVIDIA B300, B200, and AMD Instinct MI355X chips. Designed for research and development tasks, these sandboxes support model fine-tuning, reinforcement learning rollouts, and complex evaluation pipelines. Billing is calculated per GPU-second, providing a granular way to access enterprise-grade compute without long-term commitments. Currently available as a research preview, these environments offer a secure space for teams to develop proprietary AI models.

Compute Infrastructure

  • Isolated Runtime Control: Hedra GPU Sandboxes provide isolated GPU compute environments where users maintain full control over the runtime.
  • Custom Code Execution: The service supports running custom training pipelines and model evaluation code directly on frontier infrastructure.
  • Frontier Hardware Access: Users can leverage high-performance hardware including NVIDIA B300, B200, H100, and RTX 4090 chips.
  • AMD Accelerator Support: The platform provides access to AMD Instinct MI355X and MI455X accelerators for research and development.

Resource Management

  • Per-Second Billing: Compute usage is billed strictly per GPU-second, allowing for granular cost management.
  • Research Preview Status: GPU Sandboxes are currently offered as a research preview for teams developing proprietary models.
  • Cost Estimation Tools: The platform includes a GPU cost estimator to calculate monthly expenses based on GPU count and billable hours.
  • Developer Tooling: Users can manage their compute resources through the Hedra CLI and MCP integrations.

Use cases

  • Research teams run custom training pipelines in isolated environments to develop proprietary AI models.
  • Developers execute reinforcement learning rollouts on frontier NVIDIA and AMD hardware with full runtime control.
  • Engineers perform complex model evaluations using per-second billing to optimize compute costs.

Specifications

Platforms
Web, API
Editors and tools
CLI, MCP
Deployment
Cloud-based isolated GPU environments
Billing Increment
Per GPU-second
Availability Status
Research Preview
Supported Hardware
NVIDIA B300, B200, H100, RTX 4090, RTX PRO 6000, AMD MI355X
Environment Type
Isolated GPU compute
Control Level
User-controlled runtime

How it works

  1. Select Hardware: Select the required GPU hardware from the catalog, such as NVIDIA B200 or H100, based on your workload needs.
  2. Deploy Custom Code: Deploy your custom training, evaluation, or rollout code into an isolated environment where you control the runtime.
  3. Execute and Bill per Second: Monitor usage and pay only for the compute consumed, billed at a per-second rate for the duration of the task.

Pricing

NVIDIA B300
$0.00229167 per GPU-second ($8.25/hr equivalent)
NVIDIA B200
$0.00201667 per GPU-second ($7.26/hr equivalent)
NVIDIA H100
$0.00083333 per GPU-second ($3.00/hr equivalent)
NVIDIA RTX 4090
$0.00024722 per GPU-second ($0.89/hr equivalent)
AMD Instinct MI355X
$0.00138889 per GPU-second ($5.00/hr equivalent)

FAQ

How are Hedra GPU Sandboxes priced?

GPU compute for sandboxes is priced per GPU-second. The specific rates depend on the hardware selected, such as NVIDIA B200 or H100.

Can I buy compute without a research engagement?

Yes, users can discuss compute directly with the Hedra team. However, model training and GPU Sandboxes currently remain in research preview.

What GPU models are available in the sandboxes?

Hedra provides access to frontier hardware including NVIDIA B300, B200, RTX PRO 6000, H100, RTX 4090, and AMD Instinct MI355X.

Explore GPU Sandboxes

Added by AntSpace from public sources (hedra.com).

Related

More products by Hedra