Ascend NPU fleet

Six surfaces over one control plane, one per persona. Everything they show came back from a real control plane running on real Ascend hardware — an Atlas 800I A2 with eight 910B3 — captured through each persona's own token, refusals included. Nothing on these pages is invented, and every page says at the top whether it is replaying a recording or talking to a live API.

Recorded, not mocked. Two sessions: one healthy fleet, and one where a node was genuinely taken off its certified stack and detected by the controller.

The five people this is for

P1 · Fleet operator — served

Operator console

What capacity exists, who holds it, which nodes are on their approved stack, and what has drifted. The half that matters most — compliance and drift — only has an HTTP endpoint because building this page is what showed it was missing.

healthy fleet a drifted node as a tenant (refused)

P2 · Tenant admin — served

Tenant portal

One customer's entitlement, what they hold, what they have used and what it costs. It exists only because the API gained an identity and a scope policy; before that, one unauthenticated GET returned every tenant's spend.

as Acme as Globex as an engineer

P3 · Tenant engineer — served

Engineer workspace

Get a slice and run work on it without writing a PodGroup. The value is the refusals: every rule the form applies is one the hardware or the namespace policy turned out to impose, discovered by failing against real silicon.

as an engineer as Globex

P4 · Finance — served

Finance

Device-seconds from observations actually made, priced per extended resource name, attributed to a billing account that outlives the tenant. It closes a billing period by reading the meter rather than resetting it, and exports one as CSV. What it cannot do is deliver an invoice or take a payment.

fleet-wide one tenant as an engineer (refused)

P5 · Support / SRE — served

Support

Why a job is pending, in the scheduler's own words; why a node is out, by drift phase; and a merged timeline of platform and cluster records that states before any number how far back it is complete. A window with no records never reads as a window in which nothing happened.

the fleet during drift as a tenant (refused)

What can be demonstrated, on hardware, today

A bare Atlas 800I A2 brought to two isolated tenants holding real vNPU slices of real Ascend silicon, each scheduled through its own Volcano queue with gang semantics.
Work submitted entirely through an HTTP API by a token carrying one tenancy — no kubectl, no YAML, and no way for that token to reach another customer's rows.
Usage metered to the device-second, priced per extended resource name, invoiced per billing account, and surviving a restart of the platform.
A node taken off its certified stack and detected within a minute, with the exact field that differs.
Every refusal in the scope policy, demonstrated against a live cluster: 41 checks, including that one customer's response does not mention the other.

What must not be claimed

That it heals itself. Remediation runs end to end and what it runs is a template with REPLACE_ME in it, because the Ascend Deployer invocation and the reboot path are still unanswered. It detects, and refuses to touch the node.
That it spans clusters. One cluster exists. The cross-cluster gang path is deliberately refused rather than untested.
That a customer can log in. Keycloak is deployed and the realm is right; the pilot currently runs on static demo tokens, and every page says via=static-token because the API does.
That the prices mean anything. They are placeholders nobody has agreed to.
That whole cards work. As of 2026-10-29 a whole-card request fails on this node at kubelet admission, identically under Volcano and the default scheduler, while slices run fine. Unexplained, and it contradicts the day before.

How to read these pages

Each surface has a provenance bar at the top. Recorded means it is replaying a captured session; live means it is talking to a control plane on the same origin. The persona selector changes which token's recording is replayed, which is why a tenant token on the operator console shows refusals rather than an empty page: the refusal is the product working.

The recordings carry node names, tenant names and billing account identifiers from the pilot. They contain no credentials, and they were read before being published.