Managed Operations

Overview#

Production AI infrastructure rarely fails dramatically. It drifts. Certificates expire on a Sunday. A GPU node's NVLink throughput collapses to half. A Kubernetes upgrade silently breaks the DCGM exporter. A billing job hangs and a workspace overruns its budget by 4x in the time it takes anyone to notice. None of these failures are interesting individually; each of them is the kind of thing that a 24/7 operations rota with the right runbook catches in minutes and that an under-staffed in-house team catches in days. Managed Operations is Yobitel's service for owning that rota on the customer's behalf, against the customer's own estate.

The service is genuinely cross-estate. A customer might run Yobibyte on NeoCloud for production inference, EKS in `eu-west-2` for application services, an on-premise Kubernetes cluster for regulated training, and a bare-metal GPU pod for research. Managed Operations covers all four under one contract, one incident webhook, one monthly review, and one named TAM. The watch is structured around incident severity (P1 through P4), SLA tier, and a runbook library that grows with the customer.

Yobitel runs the rota; the customer keeps every system. The customer's Kubernetes clusters, cloud accounts, on-premise hardware, and data remain customer-owned and customer-controlled. Yobitel reads the customer's existing Prometheus, Grafana, OpenTelemetry, and log-aggregation surfaces; integrates incident webhooks into the customer's PagerDuty, Opsgenie, or ServiceNow; and follows the customer's change-approval workflow alongside Yobitel's ITIL-aligned process. There is no Yobitel-installed agent that has read-write access the customer cannot revoke.

Yobitel Communications, the UK-headquartered AI infrastructure company that delivers Managed Operations, sells the service as a per-node or per-cluster monthly subscription in USD. UK NCSC OFFICIAL alignment is the default posture for the operations team; ISO 27001 and SOC 2 Type II cover the wider control set. Premium and Mission-critical tiers include a named TAM as the single point of contact for both operational matters and capacity planning.

Quick start — onboarding a customer estate#

Onboarding follows a structured four-stage flow that takes most customers four to eight weeks from contract sign to SLA going live on the watched estate. The stages run in sequence and the customer's TAM and operations lead are the joint owners; the customer's existing platform and SRE teams stay in the loop throughout.

Stage one — discovery. The TAM and an operations engineer walk through the customer's estate: which clusters, which clouds, which on-premise hardware, which observability stack, which incident-management tooling, which change-approval workflow, which compliance pin. The output is an estate map and a baseline runbook list pulled from Yobitel's existing library, annotated with what the customer already has covered and what the watch will need.

Stage two — baseline. Yobitel reads the customer's Prometheus and Grafana, integrates the incident webhook into the customer's PagerDuty or Opsgenie, registers a Yobitel pager rotation, and walks the joint team through the standard P1-P4 severity definitions. The baseline stage produces the first month of run-of-rota data — actual on-call load, the noisiest alerts, the most-fired runbooks — without yet pulling the trigger on the SLA.

Stage three — watch. The SLA goes live. Yobitel's NOC takes the on-call rotation; the customer's team stays available for warm hand-off during the first month of live watch. Monthly service reviews start at this stage and continue for the life of the engagement.

Stage four — incident response. The first real incident under SLA is handled jointly; the post-mortem is co-owned. By the end of the second month under SLA, the customer's team is typically able to step out of warm hand-off and rely on the NOC for the watched envelope.

Concepts#

Managed Operations exposes a small set of concepts that match how an operations leader thinks about a 24/7 watch. The mental model is incident severity at the centre, with watch tier, runbook library, and on-call schedule around it.

• Watch tier — the SLA tier the engagement is signed at. Standard (business-hours watch with on-call escalation), Premium (24/7 watch with named TAM), or Mission-critical (24/7 watch with redundant on-call rotation, sub-15-minute P1 response, and quarterly tabletop exercises).
• Incident severity — the P1-P4 classification applied at the moment an alert fires. P1 is customer-impacting and time-critical (e.g. production inference down, billing pipeline halted). P2 is degraded but contained (e.g. one node out of a cluster). P3 is non-customer-impacting (e.g. certificate expiry warning). P4 is informational.
• Runbook — the documented response to a specific alert or incident class. Yobitel maintains a runbook library that the customer's specific runbooks extend; runbooks are versioned in the customer's Git and Yobitel's reviewed and approved through the standard change-approval workflow.
• On-call schedule — the rota and escalation path. Yobitel's NOC primary, customer's secondary (optional), customer's escalation owner (required), TAM (Premium and Mission-critical only). The schedule is published in the customer's incident-management tool.
• Capacity planning cycle — the monthly review that surfaces utilisation, drift, and capacity recommendations. Drives reservation, scale-out, and right-sizing decisions; for Yobibyte and NeoCloud customers, the cycle integrates with the same FOCUS export the customer sees.
• Scope envelope — the explicit list of clusters, clouds, on-premise hardware, and software surfaces under SLA. Anything outside the envelope is best-effort or excluded. The envelope changes through the standard change-management process, not in-flight.

SLA tiers and commitments#

Three SLA tiers cover the bulk of customer needs. The tier defines the watch window, the response commitments, the recovery commitments, and the included surfaces; the customer's actual scope envelope sits inside the tier.

Supported infrastructure#

Managed Operations covers a broad envelope of customer-built and Yobitel-built infrastructure. The supported list below is the envelope under SLA today; anything outside the envelope is handled as Professional Services rather than ongoing operations.

Engagement size by infrastructure scale#

The sizing below comes from production engagements at the small, mid-market, and enterprise scale. Node counts are aggregate across the watched estate; the relevant tier is driven primarily by criticality, not node count.

Scope envelope — in and out#

Managed Operations is a service contract, so the relevant question is not 'what is the limit?' but 'what is in scope?'. The table below is the standard envelope; customer-specific extensions and exclusions are documented in the signed scope statement and reviewed quarterly.

Pricing#

Managed Operations is priced per-node or per-cluster per month in USD on a tiered base rate. The base rate covers the SLA tier; per-incident surcharges apply only to engagements outside the standard scope envelope. Pricing is delivered as a FOCUS 1.1 line item alongside the rest of the Yobitel stack so spend rolls up into the customer's FinOps pipeline without manual reconciliation.

Security and compliance#

Managed Operations is delivered by a UK-headquartered team operating under NCSC Cloud Security Principles. UK NCSC OFFICIAL alignment is the default posture for the watch; EU and US engagements layer GDPR, EU AI Act high-risk-system obligations (where the watched estate is in scope), HIPAA, and SOC 2 Type II on top. The operations team's own controls (background checks, BYOD restrictions, separation of duties, immutable audit logging) sit under ISO 27001 and SOC 2 Type II.

Access to the customer's estate is read-only by default. Any write access required for incident response is requested through the customer's change-approval workflow with the relevant runbook attached; the customer can pre-approve named runbooks for in-incident write access, or require live approval for every write. Every action Yobitel takes on the customer's estate is logged to the customer's audit surface, not a Yobitel-only one.

• NCSC Cloud Security Principles — default posture for the operations team and the customer-watched envelope.
• G-Cloud — listed under Cloud Support (Lot 3); orderable through the Crown Commercial Service framework.
• Cyber Essentials Plus — current certificate for the operations team.
• ISO 27001:2022 — current certificate covering the operations team and its tooling.
• SOC 2 Type II — annual third-party audit covering security, availability, confidentiality.
• ITIL-aligned change management — integrating with the customer's existing approval workflow.
• GDPR / UK DPA 2018 — DPA, sub-processor list, EU SCCs available.
• EU AI Act — for customers running high-risk AI systems, Managed Operations provides the operational-resilience evidence layer.
• HIPAA — BAA available for healthcare-customer engagements.
• Read-only-by-default — write access requested through the customer's change-approval workflow; pre-approved runbooks available.

Alternatives#

Managed Operations is one option for running a 24/7 watch on AI infrastructure. The honest read: an in-house SRE team gives full control but takes 6-18 months to staff and burns continuously on rota cost; a hyperscaler-managed service covers the cloud-native primitives well but loses depth on GPU, fabric, inference engines, and ML pipelines, and is single-cloud by definition; a general MSP can cover availability but rarely has AI-infrastructure depth or AI-specific runbook libraries. Managed Operations sits in the middle as the contract that covers AI-infrastructure breadth across cloud, on-premise, and hybrid estates, with UK NCSC OFFICIAL as the default sovereignty posture and Yobibyte/NeoCloud bundled-discount economics where the customer already runs on Yobitel surfaces.

Common incident classes#

Managed Operations is a service contract rather than a product, so 'troubleshooting' is reframed as the most common incident classes the watch handles. The classes below cover the bulk of paged incidents on a typical customer estate; the runbook library covers them with documented response, fix, and post-mortem templates.

Where Managed Operations fits in the Yobitel stack#

Managed Operations is the day-two operations layer that wraps the rest of the Yobitel stack. Professional Services delivers the day-one build; NeoCloud Operations delivers the partner-build sovereign facility; Yobibyte is the managed inference surface; NeoCloud is the sovereign capacity layer; Customer Excellency owns the strategic relationship. Managed Operations is the contract that keeps the result running.

Most customers buy Managed Operations alongside one or more of the other Yobitel surfaces. A customer running Yobibyte on NeoCloud often adds Managed Operations to cover the customer's own application services and any non-Yobitel infrastructure in the same envelope. A customer running a partner-built NeoCloud often contracts Managed Operations as the day-two operate phase. A customer running entirely on hyperscaler Kubernetes can contract Managed Operations without ever adopting another part of the Yobitel stack — the watch covers their estate as-is.

The boundary with Professional Services is deliberate. Professional Services is the engineering engagement for net-new builds, migrations, and bespoke implementations; Managed Operations is the ongoing watch on what is in production. Most engagements that start as Professional Services convert at least partially into Managed Operations at production hand-off; the same engineers in the consulting engagement are not the same engineers running the rota, but the runbook library carries through from one to the other.