XPERTAI Request access
Infrastructure

Reserved, isolated,
metered at the edge.

The pool sits under a long-term capacity agreement and is dedicated to production traffic. Isolation is a property of the deployment, not a configuration setting.

100%

Of the pool is reserved for a single tenant.

Capacity is contracted, not rented by the hour. The pool exists for one customer's production traffic and nothing else runs inside the boundary.

That arrangement is unusual because it removes the two things that make shared AI capacity hard to forecast: a neighbour whose traffic shape changes, and a price that moves with the spot market.

Foundations

Three properties the deployment depends on.

01

Contracted capacity

Reserved under a 60-month term, sized for production rather than for a benchmark run. The ceiling is known before the first request.

02

Physical isolation

Separate infrastructure, not a logical partition of a shared cluster. No other tenant is admitted inside the pool boundary.

03

Metering at the edge

Consumption is counted where the pool meets the network, so what is reported is what crossed the boundary.

Architecture

Where the boundary sits.

A request crosses four points. Only the first and the last are visible to the customer.

Request path through the isolated pool A customer application sends a request through a private gateway into a reserved single-tenant pool, where it passes metering, the inference runtime and GPU nodes before a response returns. Customer application RESERVED POOL — SINGLE TENANT Gateway private ingress Metering per request Runtime serving stack GPU dedicated nodes RESPONSE — TOKENS COUNTED ON THE WAY OUT
At a glance

The deployment in one table.

Capacity modelDedicated pool, reserved ahead of demand
IsolationPhysical — separate infrastructure, single tenant
Commitment60-month capacity term
MeteringPer request, counted at the pool boundary, input and output split
AccessAPI endpoint, private endpoint or embedded surface
DeploymentDesignated data centre, contracted through the Singapore entity
ReportingConsumption by tenant, model and billing period
Operations

Who watches the pool.

Monitoring

Throughput, latency and error rates are watched continuously at the gateway. Capacity headroom is reviewed against consumption rather than against a fixed forecast.

Incident response

Incidents are handled by the team that operates the runtime, not escalated to a hardware vendor. One point of accountability from symptom to resolution.

Change control

Model versions and runtime upgrades are rolled out behind the pool boundary. The customer's integration surface stays stable across upgrades.

Request access

Ask about the current deployment.

Capacity, isolation and delivery can be discussed in detail under NDA.

Request access