Edge and disconnected AI
One decision sits underneath every remote deployment: whether the capability has to work when the link does not.
The decision, not the diagram #
Forward locations, ships, rigs, substations, factory floors, survey vehicles and mobile command posts generate the data that decisions turn on, and they sit at the end of a satellite link, a radio hop or no link at all. That much is well understood. What is usually left unresolved is the commitment it implies.
If the answer has to be available when the link is not, then inference belongs at the site and the weights have to be there before they are needed. Everything else — the hardware, the mirroring, the reconciliation, the operator's view — follows from that single choice, and this page is about making it deliberately rather than discovering it during an outage.
Disconnected operation is a design property, not a degraded mode you discover during an outage.
Signals that the decision is already made for you
- A round trip that is too slow to be operationally useful.
- Bandwidth costs that rule out shipping raw sensor or video data at all.
- A classification or a mandate that forbids the data leaving the site.
- An outage window during which the capability still has to be there.
Four questions that settle it #
Most conversations about the edge stall on the wrong question. These are the four that actually determine the shape of the deployment.
- Does the capability have to work when the link is gone?
- If yes, inference belongs at the site and the weights must be mirrored there in advance. That is the whole decision. Everything after it is sizing.
- Is the constraint bandwidth, or is it permission?
- Bandwidth makes a central round trip slow and expensive. A classification or a residency rule makes it impossible, and also rules out the managed pattern. Both end at a local site, but only one of them leaves you a choice.
- Who operates the site, and what do they see?
- Central operators publish the model state they want at each site and see what each site reports it is running. While a link is down they hold the last reported state. Nobody needs a session at the site itself.
- How long can a site run alone?
- As long as its hardware, its models and its workload allow. We size that with you against a real site rather than publish a figure that would not survive contact with yours.
What you are committing to #
An Agentycs edge is a complete inference site rather than a cache: it hosts its own model workers, serves requests locally, holds mirrored weights and enforces the same tenant policy as the core, so a remote deployment does not quietly become a governance exception.
It stays in sync over an authenticated, signed link built for wide-area conditions, and it keeps serving its loaded models when that link is unavailable. The commitment you are making is to decide in advance what each site holds, because that is what it will have.
The mechanism, in short
The detail of how a site is built and operated belongs on the deployment page. This is the summary, if you want it before you go there.
- Remote autonomy
- An edge keeps serving its loaded models when the core is unreachable, then reconciles state once the link returns, so a link failure is not a capability failure.
- Weights already on site
- Model artefacts are seeded, mirrored and pinned out to workers and remote edges before a model is needed, from a local mirror that works with the link down.
- Central intent, local execution
- Desired model state is published out to each site and what each site is actually running is reconciled back into one central view, so operators keep oversight without needing a live session.
- A bridge built for bad links
- Control and dispatch travel over an authenticated, signed HTTP/3 link designed for wide-area conditions, with request and token-stream recovery so transient loss does not drop an in-flight generation.
- The same policy everywhere
- An edge enforces the same tenant policy as the core, including whether data may leave the site, so a remote deployment does not become a governance exception.
- Hardware to match the site
- The server range runs from racks at a main site to a 6-litre micro server that fits in a backpack, draws 400W, can run on battery and still carries a full 96GB GPU and a complete platform instance. Or use your own hardware where that suits.
- Networking that survives geography
- A secure overlay with mesh-internal DNS and zero-trust group policy lets sites and services resolve and authenticate each other across networks you do not own.
- Degradation you can see coming
- Continuous per-cluster wide-area probing against external anchors makes network degradation visible before it turns into an outage.
Anima
The edge runtime, artefact mirroring, edge dispatch, stream recovery and remote autonomy.
Apex
The whole AI server range, from a rack at a main site down to a 6-litre micro server that fits in a backpack.
Cloud Mesh
Secure site-to-site networking, mesh-internal DNS and zero-trust policy across links you do not control.
Atom
Local data and retrieval at the site, so grounding does not require a round trip either.
Where to go next #
This page is the decision. The mechanism, the hardware and the operating detail live on the deployment pages, which is where to go once the decision is made.
Edge and remote sites
How a site is built, governed and operated. Start here once the decision is made.
Air-gapped
A complete platform with no external route at all, and the trade-offs that come with it.
Agentycs in a box
The micro end of the range: backpack-sized, battery-capable, 400W, running the full platform.
Multi-site
Many sites reconciled into one operational view.
Where this lands hardest #
The same capability answers a different first question in each sector, and this one belongs squarely to a single reader.
Defence and intelligence
Capability that only works when data crosses a boundary is capability you cannot use.
Scope a disconnected site
Tell us the site, the link you can rely on and the longest outage you have to keep working through. We will size what runs there.