Your fleet,
fully utilized.
One command runs any model across heterogeneous AI compute, from a single edge box to large GPU clusters.
AI compute is mainly one type of cluster and nobody serves the rest.
Neoclouds and hyperscalers, with a layer of NV-certified software vendors make those clusters robust at scale. But everything outside of this big-iron goes unserved. The heterogeneous fleet most organizations actually have is more available, more accessible and far larger than the clusters everyone designs for.
These run the models people pay for, and no one is making them work as one fleet:
The developer interface to all of it.
A Capsule is a dynamic container between the people and agents requesting compute and the silicon underneath. A user asks for a model. OxCapsule finds the hardware, places the work and balances it across the full spectrum of heterogeneous compute.
Helpful Commands All in one place
The capsule CLI orchestrates every layer through one consistent surface: auth, discovery, dev environments, streaming and storage.
capsule auth loginBrowser OAuth, persistent tokens.capsule orbitA live look into your fleet.capsule env set prodSwitch dev / demo / prod on one fleet.capsule list --filterFilter by vendor, VRAM or cores.capsule code a100VS Code over SSH, idle-timeout, auto-clone.capsule cursor rtx-4090Cursor IDE over the same remote-dev flow.capsule docker a100Containers with GPU passthrough and volumes.capsule streamPixel-stream GPU-heavy apps to any thin client.
See every system you own.
Orbit opens a live window into the whole fleet, every system and every GPU, across every vendor, in one place.
Vendor mix · NVIDIA · AMD · Intel · Tenstorrent · Storage
Qwen3.5-397B running across 3 RTX A6000s. One command.
Qwen3.5-397B does not fit on a single box. With Capsule Queue, one command finds every GPU in the fleet that can contribute, disaggregates the model across three RTX A6000s, and runs it as a single inference endpoint. No driver gymnastics, no manual sharding.
A single NVIDIA box runs a model out of the box. Table stakes. Everyone can do it.
A 397B model that fits on no single box, spread across the fleet, from one command.
Three things no one else has.
One fleet, any silicon. NVIDIA and non-NVIDIA, edge to cluster, working as a single pool.
One model across many systems. Capsule Queue finds every GPU that can serve a model and runs it as one, even when it fits on no single box.
Relative performance of your fleet across models, on tap. As far as we know, no competitor offers it.
Join the beta.
We released the beta on November 4, 2025. Hundreds of developers already reach continuous compute across heterogeneous hardware with OxCapsule.
One session. Three machines.
No dropped frames.
At Tenstorrent Deploy Dev Day, we showed what owning your compute stack looks like: one AI workload moving across three different machines without missing a frame.
Three compute targets. One continuous creative session. Zero re-installs, zero driver gymnastics.
Put your whole fleet to work.