Token Factory turns GPU capacity into tokens — it does not create that
capacity. In the NeoCloud, the layer that does is k0rdent AI: Mirantis'
infrastructure product that turns racks of NVIDIA-certified hardware into
consumable clusters. This page records what k0rdent AI provides, how Token
Factory relates to it, and — just as important — what Token Factory
deliberately does not do.
Source: k0rdent AI ↔ Token Factory alignment, 2026-07-24. The boundary is
still being defined on both sides; assumptions below are marked as such.
| Capability | Shape |
|---|---|
| Kubernetes as a service | A cluster definition (node count range × machine type) instantiated per customer: you get a vanilla Kubernetes cluster with GPU nodes and retrieve the kubeconfig via REST API |
| Bare metal as a service | A loosely-coupled group of bare-metal GPU nodes; SSH keys retrieved via API |
| Multi-tenancy + quotas | Providers carve out per-tenant maximum quotas; tenants self-provision up to them |
| Network fabric | InfiniBand support (built for NVIDIA-class deployments) and RoCE support (built for IREN) — fabric provisioning is handled below the cluster line |
| Atlas | The provider-only UI (onboarding hardware, data-center management). Customer-facing UIs are expected to be provider-branded builds against the customer APIs; a reference customer portal exists but is paused |
| Offtake-aligned APIs | The MVP API surface is designed around the NVIDIA offtake spec (Kubernetes aaS + bare metal aaS); IREN's first customer is NVIDIA |
The delivery pipeline k0rdent AI optimizes: hardware comes off the truck →
plugged in → discovered → added to the provisionable pool.
Working assumption — not yet a settled contract
The expected model is that Token Factory runs as one tenant of k0rdent
AI's Kubernetes-as-a-service: we receive a Kubernetes cluster with GPU
nodes (a cluster-definition instance) and deploy every Token Factory
capability on top of it. Where exactly "k0rdent AI stops and Token Factory
begins" is not 100% defined yet on either side; treat this section as the
current best understanding, not an agreed interface.
Consumption model:
Explicit non-responsibilities (so nobody builds them twice):
Conversely, everything above the cluster line — gateway, key governance,
metering, portal, identity, chat, model serving orchestration — is Token
Factory and is not k0rdent AI's job.
Ideas discussed across the boundary, tracked as research:
Capability gating (2026-07-31): Token Factory never requires k0rdent
— it runs on any conformant Kubernetes (vanilla, OpenShift, MKE, …). But
where k0rdent APIs are present, some capabilities may light up that only
work with them — dynamic scaling of GPU capacity, deeper tokenomics
optimization. Same platform, richer behavior on a k0rdent substrate; never
a hard dependency (ADR-003 stands).