Cloud inference API for physical AI that runs frontier robot models with under 100ms latency.
Homepage hero - real-time cloud inference for physical AI
1 / 2Overview
What is Dreamscale Labs?
Run the best robot AI models over the cloud
Dreamscale Labs provides real-time cloud inference infrastructure for physical AI. Instead of running large robot models on local hardware, robotics teams can run frontier models such as NVIDIA DreamZero and MolmoAct2 VLA/WAM models over the cloud, with most models responding in under 100ms end to end. The platform optimizes the full stack — model, GPU kernels and network layer — and its Real-Time Chunking (RTC) technique cuts tail latency 40-50% compared to existing cloud providers. You can start from a frontier model or connect your own fine-tuned checkpoint for your robot. Models are currently in research preview; request access through the website.
Platforms and languages
Dreamscale Labs Availability
Platforms
Languages
Capabilities
Dreamscale Labs Key Features
Real-time cloud inference for physical AI
Runs frontier robot models such as NVIDIA DreamZero and MolmoAct2 VLA/WAM models in the cloud, with most models responding in under 100ms end to end.
Real-Time Chunking latency optimization
Full-stack optimization across model, GPU kernels and network layer; Real-Time Chunking (RTC) cuts tail latency 40-50% compared to existing cloud providers.
Bring your own model
Start from a frontier model or connect your own fine-tuned checkpoint for your robot. Custom models are in research preview; contact the team for access.
High-concurrency execution
Serves a 14B model at 320ms inference and runs 8+ robots concurrently on a single H100, with on-demand compute that fits experiment-iterate-deploy robotics workflows.
Best for
Who uses Dreamscale Labs?
Robotics teams deploying or evaluating robot AI models
Run frontier robot models in the cloud with real-time latency instead of on local hardware
Common questions
Dreamscale Labs FAQs
What is Dreamscale Labs?
Dreamscale Labs is a cloud inference API for physical AI: it runs frontier robot AI models in the cloud so your robot does not need heavy local compute, with most models responding in under 100ms end to end.
How do I get started?
Pick a model such as DreamZero or MolmoAct2, connect your robot, send it instructions and run. The platform handles compute across the model, GPU kernels and network layers.
Can I use my own model?
Yes. You can start from a frontier model or connect your own fine-tuned checkpoint. Custom models are in research preview, so contact the team to request access.
How fast is it?
Most models run in under 100ms end to end. Real-Time Chunking (RTC) cuts tail latency 40-50% versus other cloud providers, and a 14B model can serve 8+ robots concurrently on a single H100.
Is it publicly available?
Models are currently in research preview. Click Get started / Get in touch on the website to request access.
Reviews
See what the community thinks and share your experience.
—
Based on 0 ratings
Rating distribution
Community reviews
No reviews
No written reviews yet.
