Yobitel Marketplace

Overview#

Customers adopting AI inside an enterprise spend a surprising amount of time on what should be a five-minute question: which model, which application, for which workload, under which licence, in which region, at what cost? Open registries are vast and unfiltered; cloud-marketplace catalogues are anchored to a single hyperscaler and a single billing surface; internal evaluation projects expand to fill the calendar. The marketplace exists to collapse that into a filterable shopfront where a buyer in procurement, a platform team, and a downstream developer can all see the same shortlist for the same workspace.

Yobitel Marketplace is the catalogue surface of the Yobibyte platform. It is not a separate marketplace — it is the shopfront customers see when they sign into their workspace, with three layers in one view: open-weight models with verified deployment recipes, AI Applications (Yobitel-published and partner-published), and customer-private items scoped to the workspace. The same filters apply to all three; the deploy action lands in the same workspace; the same identity, sovereignty, audit, and billing surface holds across them.

Compared with the Hugging Face Hub, the Yobitel Marketplace is curated rather than open: every entry has been reviewed for a verified deployment recipe, declared sovereignty eligibility, and explicit licence terms, and deployment is in-workspace rather than 'download and operate yourself'. Compared with AWS Marketplace or Azure Marketplace, the catalogue is anchored to the Yobitel platform rather than a single hyperscaler; deployments land on the workspace's bound sovereignty region, not the cloud-specific region taxonomy. Compared with Bedrock Catalog or Vertex AI Model Garden, the catalogue is multi-publisher and multi-runtime, and AI Applications sit alongside raw models in the same shopfront.

Yobitel Communications — a UK-headquartered AI infrastructure company and NVIDIA Inception partner — operates the marketplace as the customer-facing surface of Yobibyte. Curation, sovereignty attestation, and licence review run on a published cadence. Pricing is in USD throughout; rank is sourced from InferenceBench; deployment lands in the customer's Yobibyte workspace.

Quick start#

Browsing the marketplace is a workspace activity. Sign in to the Yobibyte console with your corporate identity provider; the marketplace shopfront opens to the catalogue scoped to your workspace's sovereignty pin (a UK-bound workspace sees only entries eligible for UK NCSC OFFICIAL). Use the top-level toggle to switch views: Models, AI Applications, or Customer-Private.

Apply the filters that matter to the workload. Sovereignty narrows the catalogue to entries eligible for your pin; licence narrows to commercial-use clearances; family narrows to Llama, Qwen, Mistral, Phi, Gemma, and other open-weight families for models, or to industry (clinical, finserv, manufacturing, telco, retail, agriculture, public sector) and category (decision support, vision QA, agentic workflows, document understanding) for AI Applications. The result list shows the InferenceBench rank for models, the deployment profile for applications, the USD price band, and the smallest GPU configuration the workload fits on.

Click Deploy on the entry you want. For models, the deploy dialog asks for replicas, autoscaling target, and a spend cap; Yobitel populates the engine, SKU, and serving flags from the verified deployment recipe. For AI Applications, the deploy dialog opens the configuration wizard (data sources, RBAC mappings, branding, policy bindings). Either flow lands in the workspace with the same identity, audit, billing, and KMS controls the workspace already enforces.

Concepts#

The marketplace introduces a small set of concepts that sit on top of the workspace primitives. The mental model is: a catalogue holds three classes of entry; filters narrow the catalogue to what is eligible for a workspace; rank orders the remaining set; deploy lands the entry inside the workspace.

• Catalogue — the unified shopfront covering Models, AI Applications, and Customer-Private items. Every entry carries sovereignty, licence, family/industry, and footprint metadata. The catalogue is the single source of truth for what a workspace can deploy.
• Models — open-weight models with verified deployment recipes (engine selection, GPU SKU shape, serving defaults). Customers do not select the engine or SKU; they choose the model and the recipe drives the runtime.
• AI Applications — Yobitel-published first-party applications (MediQuery and the wider vertical suite) and verified partner-published applications. Configuration is the customer surface; runtime is operated by Yobitel.
• Customer-Private Items — workspace-scoped uploads (bring-your-own model weights or packaged customer-built applications) that live inside the catalogue but are visible only to that workspace. The catalogue registers them; the storage stays in customer-owned buckets.
• Sovereignty Filter — the primary catalogue lens. Eligible sovereignty pins are declared per entry (`ncsc-official`, `eu-data-boundary`, `fedramp-equivalent`, `hipaa-attested`). The default view scopes to the workspace's bound region; widening is explicit and audited.
• Licence Filter — narrows by commercial use, redistribution, attribution, and per-call or per-seat obligations. Each entry's licence terms are surfaced in the listing and again in the deploy dialog; acceptance is recorded.
• Verified Recipe — the deployment profile attached to a model or application: engine selection, GPU SKU shape, serving flags, retrieval and evaluation behaviour, and footprint envelope. The recipe is the publisher's contract with the platform; the customer sees its shape but does not edit its internals.
• InferenceBench-Sourced Rank — the public economic ranking applied to model entries: price per million tokens and tokens per second at representative input/output shapes, refreshed from inferencebench.io. Rank is honest by construction because it is computed against public benchmarks.

Reference — filters and entry metadata#

The table below is the full filter set and entry metadata schema the marketplace exposes. Models, AI Applications, and Customer-Private items share the same shape with category-specific extensions.

Workload patterns#

Three deployment shapes cover most of what customers do with the marketplace — a UK trust deploying an NCSC-eligible Llama for in-region inference, a regulated finance customer deploying a commercial-licence model for advisor copilots, and a media customer deploying an AI Application with a sovereignty pin and a per-application spend cap.

• NCSC-eligible Llama 3.1 deployment — a UK government workspace is bound to NCSC OFFICIAL. The shopfront defaults to UK-eligible entries; the customer filters by family (Llama) and modality (text), sees Llama 3.1 70B Instruct ranked at the top by InferenceBench economics for in-region UK serving, clicks Deploy, sets replicas 1–6, autoscale target 1,800 tokens/sec/replica, and a $5,500/month spend cap. The verified recipe drives the engine, SKU, and serving flags; the endpoint reaches Ready inside three minutes and is OpenAI-compatible from the workspace's domain.
• Commercial-licence model deployment — a regulated finance customer needs a model with explicit commercial-use clearance and per-seat redistribution rights. The licence filter narrows the catalogue to commercial-cleared entries; the customer compares two of them on InferenceBench rank and per-million-tokens pricing, accepts the licence terms in the deploy dialog (acceptance is recorded with the workspace owner's identity), and deploys with EU Data Boundary sovereignty alongside FCA operational-resilience controls inherited from the workspace.
• AI Application deployment with sovereignty pin — a UK NHS trust deploys MediQuery; the marketplace pre-filters the shopfront to NCSC OFFICIAL + HIPAA-attested entries, the deploy dialog opens the configuration wizard, the trust binds the EHR and imaging archive, maps clinician / pharmacist / auditor roles, points audit at the trust's SIEM, and the application reaches Ready in the same session with one Yobitel-operated runtime behind it.

# PREVIEW - the marketplace search API is in active development;
# this is the planned shape and is not runnable today. Customers
# browse the catalogue through the Yobibyte console UI in the
# meantime.
#
# apiVersion: yobibyte.yobitel.com/v1
# kind: MarketplaceSearch
# metadata:
#   workspace: trust-london-uk
# spec:
#   layer: models
#   filters:
#     sovereignty: [ncsc-official]
#     licence: commercial-cleared
#     family: [llama]
#     modality: [text]
#   rankBy: inferenceBench.economics
#   limit: 10

Catalogue depth by category#

The shopfront's breadth is deliberate rather than maximised. Every entry is curated; volume is a side effect of fit. The table below summarises the categories surfaced at the time of publication; new entries land on a published cadence and existing entries are re-reviewed on a rolling schedule.

Limits and quotas#

Default per-workspace limits exist to protect shared infrastructure during onboarding; almost every limit is raisable on request. Limits below apply to the marketplace surface itself; deployment limits inherit from the workspace's Yobibyte limits.

Observability#

The marketplace emits a stable `yobitel_marketplace_*` metric set covering catalogue browse, filter usage, deploy funnel, and customer-private upload activity. Platform teams use these to instrument adoption — which sovereignty pin is most-used, which families dominate search, where the deploy funnel drops users — and procurement teams use the deploy-funnel data to plan budgets.

Three signals matter most: catalogue health (are entries reviewed on cadence, are stale entries flagged?), discovery health (which filters are used, where do users abandon?), and deployment funnel (browse → filter → entry → deploy → Ready, with per-stage timing).

The PromQL block below is the alert most platform teams add first: it catches the 'users are looking at entries but no one is actually deploying' failure mode that signals either a recipe regression, a licence-acceptance friction, or a workspace-quota block.

groups:
- name: yobitel-marketplace
  interval: 30s
  rules:
  - alert: MarketplaceDeployFunnelCollapse
    expr: |
      sum by (workspace) (
        rate(yobitel_marketplace_entry_view_total[15m])
      ) > 5
      and
      sum by (workspace) (
        rate(yobitel_marketplace_deploy_success_total[15m])
      ) == 0
    for: 30m
    labels: { severity: page }
    annotations:
      summary: "{{ $labels.workspace }} marketplace deploy funnel has collapsed"
      runbook: https://docs.yobitel.com/runbooks/marketplace-funnel

  - alert: MarketplaceLicenceAcceptanceStuck
    expr: yobitel_marketplace_licence_pending_total > 5
    for: 1h
    labels: { severity: warn }

Cost and FinOps#

Browsing the marketplace is free; cost flows through Yobibyte at deployment time. Models bill per inference (per-million-tokens, per-second-GPU-time, or per-call depending on the recipe); AI Applications bill at their published tier in USD plus pass-through inference and fine-tune consumption; customer-private items bill at the workspace's standard rate.

Every line item carries the FOCUS 1.1 columns (BilledCost, EffectiveCost, ListCost, ChargePeriod, ServiceCategory, SubAccountId, Tags), with marketplace-deployment tags attached so spend pivots cleanly by catalogue entry, by workspace, or by business unit.

Security and compliance#

Catalogue eligibility is enforced at the marketplace level. A workspace pinned to UK NCSC OFFICIAL sees only entries whose declared sovereignty includes that pin; widening the filter to entries outside the pin is an explicit, audited action that does not change what the workspace can actually deploy without an additional admission step at Yobibyte.

Licence acceptance is recorded against the workspace owner's identity and frozen at deployment time; subsequent licence changes by the publisher require explicit acceptance before they propagate. Customer-private items inherit the workspace's KMS and audit configuration; their weights or packages stay in customer-owned buckets and are encrypted at rest with the customer's key.

• NCSC Cloud Security Principles — catalogue entries declare sovereignty eligibility; UK OFFICIAL-tier audited annually.
• G-Cloud framework — Yobitel listings procurable via G-Cloud Lots 2 (Cloud Software) and 9 (Cloud Software for AI); marketplace deployments inherit the listing.
• Cyber Essentials Plus — maintained for the platform.
• ISO 27001 / SOC 2 Type II — current certificates available under NDA.
• GDPR / UK DPA 2018 — DPA, sub-processor list, and EU SCCs available; data residency enforced at admission.
• HIPAA — BAA available for clinical applications and HIPAA-attested model entries; encryption, logging, access-control controls audited.
• EU AI Act — risk classification and disclosure obligations recorded against each entry; partner-published applications include the partner's declared classification.
• DSP Toolkit (NHS) — NHS-eligible applications publish their DSP Toolkit responses on the listing.

Alternatives and customer-owned baseline#

Without the Yobitel Marketplace, most teams build the equivalent shortlist in a spreadsheet — they pull entries from open hubs, manually attach licence and sovereignty notes, manually compute rank from benchmark CSVs, and re-do the exercise for every workspace and every region. The comparison below positions the marketplace against the closest alternatives.

Troubleshooting#

The errors below cover the catalogue-surface and deployment-funnel failure modes seen most often. The full runbook library is at docs.yobitel.com/runbooks.

Where the Yobitel Marketplace fits in the Yobitel stack#

Yobibyte is the platform underneath the marketplace; every deployment from the catalogue lands in a Yobibyte workspace, with the workspace's identity, audit, billing, and KMS controls applied. The marketplace is the discovery surface; Yobibyte is the runtime surface. Omniscient Compute sits below Yobibyte — when a marketplace deployment lands, Yobibyte calls Omniscient at admission and reconciliation time to ground the workload's placement in current capacity and pricing.

InferenceBench is the public ranking surface the marketplace consumes. When a customer sorts the Models layer by rank, the numbers come from the public InferenceBench tables — the same numbers anyone can verify at inferencebench.io. That makes rank honest by construction; the marketplace cannot quietly favour Yobitel-aligned entries because the rank is computed and published independently.

The AI Applications Suite is the application layer the marketplace surfaces alongside models. Customers do not need to know which layer (model or application) they are operating in; the shopfront is unified, and the workspace pulls the runtime, identity, audit, and billing surface together regardless of catalogue layer. The marketplace is, deliberately, the single front door customers walk through.