GPU compute
Call AI models, rent GPU servers by the hour, run your own containers or reserve a dedicated cluster. All on one global GPU network.
Choose your level of control
OpenGPU is the network. Relay is the easiest way to use it, for AI models, studio and GPU instances. The Client App runs your own containers, and Enterprise reserves dedicated GPU capacity.
CreateAI video, image and voice with no code.
Per credit
Call80+ models, one OpenAI-compatible endpoint.
Per token
RunYour containers as routed tasks.
Per task
RentFull machines with SSH access.
Per hour
ReserveReserved pools and clusters with SLAs.
ContractFully managedFull control
One HTTPS endpoint, OpenAI-compatible, with card or invoice billing. Nothing new to set up.
Native · Client App
Publish your own container as a source, define routing rules and submit tasks directly to the network. Even lower pricing than Relay.
| Relay API | Client App | Relay Instances | Dedicated | |
|---|---|---|---|---|
| Interface | OpenAI-compatible API | Client App, SDK, CLI | SSH | Any of these |
| What you run | Hosted models | Your own containers | Anything on a full machine | Production workloads |
| Billing | Per token | Per task | Per hour | Contract & invoice |
| Commitment | None | None | None (1-hour minimum) | Reserved, with SLAs |
| Best for | AI products and agents | Training, rendering, batch | Dev boxes, notebooks, services | Predictable, large-scale demand |
Studio
Relay Studio turns a prompt into finished AI videos, images and voice-overs, then edits them into one video. Powered by the same models and network as the Relay API.
Shot planning, filming, voice, music and the final edit, with the plan and cost shown first.
Video clips, talking presenters and motion from a single photo.
Image generation, voice-overs and audio in the same workspace.
Product showcases, before/after and more, in 9:16, 1:1 or 16:9, up to 3 minutes.
Instances
When you need a full server, not a single task: pick hardware from hosts across the network, connect over SSH and run anything. Training runs, notebooks, long-running services or your own inference stack.
Each offer shows the GPU, specs and host country, so you know exactly what you get.
SSH access, your OS image, drivers and network settings.
Attached storage and networking features with every instance.
Start and stop when you want. One-hour minimum, no commitments.
Capacity
GPUs, CPUs and specialized AI accelerators. The router sees all of it as one pool and places each job on the hardware that fits it best.
Datacenter
80 GB HBM
Large-model training and high-throughput inference.
Datacenter
40 / 80 GB HBM
Fine-tuning, distributed jobs and batch inference.
Prosumer
24 GB GDDR6X
Image and video generation, rendering, mid-size LLMs.
Edge
Unified memory
Lighter inference and on-device-class workloads.
Workloads
Each workload has a different resource profile. The router knows the difference.
Prefill is compute-bound and sets time to first token; decode is memory-bound and sets time per token. OpenGPU routes each request to GPUs that match its memory and speed needs, and scales out during traffic spikes.
Full pre-training, fine-tuning and continual training need large VRAM, steady I/O and resilience. Long jobs are matched to stable providers, GPUs are grouped for distributed runs, and checkpoints keep progress if a node drops.
Image, video, audio and 3D generation range from bursty interactive requests to heavy batch runs. Latency-sensitive jobs land on fast nodes; batch jobs spread horizontally. No-code teams can create directly in Relay Studio.
Scenes and videos are split into frames or segments, rendered in parallel across the network and collected back. Ray tracing, lighting passes, FX sims, transcoding and AI upscaling.
Reinforcement learning, scientific simulation and evaluation sweeps run as many small jobs instead of one giant cluster. Scale up or down on demand and recycle idle GPUs into new tasks.
Embeddings, indexing, retrieval and agent loops run as batch or real-time jobs. Relay keeps agent systems running even when individual providers churn.
Pricing model
Prices are set by the market. High-demand tasks attract more providers, prices adjust to network conditions, and you get the best performance for your budget.
Per token, per task or per hour, depending on how you use the network.
Capacity comes from a live pool, so you never pay for an idle cluster.
No reserved instances, minimum spend or multi-year commitments. Dedicated capacity is reserved under an agreement.
Providers compete for tasks. The router finds the best price-performance match.
On the roadmap
Agents will automate decisions for both sides of the marketplace, and an open Agent Marketplace will let developers publish and earn from their own.
Automated task submission, deadline-aware scheduling, budget optimization across tasks and result validation.
Automatic source registration, profitability-based task selection, hardware allocation and performance tuning.
Real-time pricing, supply and demand trends and dynamic fee adjustment.
An open market where independent developers publish agents and users pick the best performers.
Get $10 of free credit on Relay, or open the Client App and submit your first task.
