Yobitel GPU Cloud

Overview#

Teams that need GPU capacity for AI workloads have three structurally different decisions in front of them: which SKU (H100 vs H200 vs B200 vs A100 vs the smaller inference-optimised cards), which region (driven by latency, sovereignty, and price), and which term (on-demand for elastic, reserved for steady, spot for batch). The Yobitel GPU Cloud exists to make those three decisions a SKU-level consumption surface inside the Yobitel Infrastructure pillar — pick the SKU, pick the region, pick the term, and the instance lands with the documented sovereignty, networking, and storage attached.

Yobitel GPU Cloud is the SKU-level surface. It is not the operated-service contract Yobitel delivers to partners (that is NeoCloud — full Tier-III facility, hardware, NOC, SLA) and it is not the platform layer above (that is Yobibyte — the workspace, the marketplace, the OpenAI-compatible API, the fine-tune service). It is the layer that customers either consume directly when they want raw GPU capacity, or consume implicitly when Yobibyte provisions inferences and fine-tunes against it.

Compared with hyperscaler GPU offerings (AWS P5, Azure ND, GCP A3), the differentiator is multi-vendor SKU breadth (NVIDIA and AMD across the current generation), admission-enforced sovereignty across UK NCSC OFFICIAL, EU Data Boundary, and US FedRAMP-equivalent, and per-SKU pricing published in USD with the same rates regardless of whether the customer consumes directly or through Yobibyte. Compared with specialist GPU clouds (CoreWeave, Lambda Labs), the differentiator is the integrated platform above (Yobibyte) and the vendor-neutral discovery layer (Omniscient Compute) — Yobitel GPU Cloud is one good answer, but it does not pretend to be the only answer.

Yobitel Communications — a UK-headquartered AI infrastructure company and NVIDIA Inception partner — operates the cloud. Regions are a mix of Yobitel-operated facilities and Yobitel-managed partner facilities, with the operational discipline and SLA identical across them. Pricing is in USD throughout; sovereignty is admission-enforced; the consumption surface is consistent whether the customer is buying a single GPU or reserving an 8-way H100 cluster.

Quick start#

Provisioning a GPU instance is a workspace activity. Sign in to the Yobibyte console with your corporate identity provider — the same console you use for workspaces, the marketplace, and Yobibyte resources also fronts the GPU Cloud SKU surface. Open the GPU Cloud view; the SKU catalogue opens scoped to your workspace's sovereignty pin.

Pick the SKU you want. The shopfront groups by generation (Hopper, Blackwell, Ada, Turing) and by family (training, inference, edge-class), with the per-SKU price, available regions, and current capacity surfaced on the card. Click Configure: pick the region from the workspace's eligible set, pick tenancy (`shared`, `dedicated`, or `confidential`), pick term (on-demand, 1-yr reserved, 3-yr reserved, or spot), and pick the storage and networking profile. The dialog shows the live USD rate before commit.

Click Provision. The instance reaches Ready in 30–120 seconds for shared on-demand SKUs in well-stocked regions, longer for dedicated or reserved provisions. The instance exposes a standard SSH endpoint, optionally an attached managed storage volume, and the workspace's audit pipeline records the provision event. For most customers, this is also the moment to attach the instance to a Yobibyte workspace and let the platform layer drive the workload rather than connecting directly.

Concepts#

The GPU Cloud surface introduces a small set of concepts that sit beneath the Yobibyte platform layer. The mental model is: an SKU is a hardware shape, a region is a placement choice, a tenancy is an isolation choice, and a term is a pricing choice. The instance carries all four.

• GPU SKU — a specific accelerator + node shape (H100 SXM 80GB, H200 141GB, B200 192GB, A100 80GB PCIe, L40S 48GB, etc.). Each SKU has a fixed node template (CPU, memory, local NVMe, network).
• Region — the placement choice. UK regions (uk-london, uk-manchester, uk-edinburgh, uk-cardiff, uk-belfast) for sovereign UK; EU regions (eu-frankfurt, eu-paris, eu-amsterdam, eu-dublin) for EU Data Boundary; US regions (us-ashburn, us-portland) for FedRAMP-equivalent.
• Tenancy — `shared` (multi-tenant nodes with cgroup and MIG isolation), `dedicated` (single-tenant nodes; billed per node-hour), or `confidential` (NVIDIA confidential-compute mode with TEE attestation; supported on H100 and H200).
• Term — `on-demand` (per-second billing, no commitment), `reserved-1yr` and `reserved-3yr` (committed term with discount), `spot` (interruptible at a floor price, suitable for fault-tolerant workloads).
• Networking Profile — `inference` (standard 100G), `training` (InfiniBand NDR 400G or RoCEv2 400G with topology-aware placement), `edge` (10G with high availability).
• Storage Profile — `block` (network-attached encrypted block storage), `object` (managed S3-compatible object storage in-region), `parallel-scratch` (Lustre or WEKA for training-class workloads).
• Sovereignty Eligibility — declared per SKU and per region. A workspace pinned to UK NCSC OFFICIAL sees only UK-eligible SKUs in UK regions; widening is explicit and audited.
• Capacity Reservation — committed capacity bookings tied to a call-off; surfaces in the workspace's reservation tab and consumes against the same SKU rate at the reserved discount.

Reference — SKU coverage matrix#

The table below is the SKU coverage matrix for Yobitel GPU Cloud. Per-SKU node templates, region availability, and SKU-specific notes are published in the SKU's marketplace listing; the matrix below is the index.

Workload patterns#

Three consumption shapes cover most of what customers do with the GPU Cloud directly — an inference-class workload on a single L40S node, a steady training cluster on reserved H100s, and a spot-tier batch fine-tune on B200s with explicit fault tolerance.

• On-demand inference — a UK customer needs a single L40S to host a model for an internal application. They pick the L40S 48GB SKU in uk-london, shared tenancy, on-demand term, inference networking profile, and a 100 GB object storage attachment. The instance lands in 45 seconds at $1.20/GPU/hr and is attached to the customer's Yobibyte workspace for the platform layer to operate.
• Reserved training cluster — a research lab commits to a 32-card H100 SXM5 cluster for 12 months. They reserve four 8-way H100 SXM5 nodes in uk-london, dedicated tenancy, 1-year reserved term, training networking profile (InfiniBand NDR with topology-aware placement), and 200 TB parallel scratch on Lustre. The reservation is provisioned within 48 hours, billed at $2.45/GPU/hr against the call-off, and surfaces in the workspace's reservation tab.
• Spot batch fine-tune — a media company runs a nightly fine-tune on commodity B200 capacity. They submit a spot-tier provision request for eight B200 192GB nodes in eu-frankfurt with a price ceiling of $3.80/GPU/hr; Yobitel's scheduler matches against available spot inventory and provisions within the ceiling. The fine-tune is checkpointed against the workspace's object storage every 15 minutes so an interruption resumes rather than restarts.

# PREVIEW - the GpuInstance declarative shape is in active development;
# this is the planned shape and is not runnable today. GPU Cloud
# instances are provisioned today through the Yobibyte console.
#
# apiVersion: yobibyte.yobitel.com/v1
# kind: GpuInstance
# metadata:
#   name: training-cluster-uk
#   workspace: research-lab-uk
# spec:
#   sku: h100-sxm5-80gb
#   region: uk-london
#   tenancy: dedicated
#   term: reserved-1yr
#   nodes: 4
#   networking: training-ib-ndr
#   storage:
#     - kind: parallel-scratch
#       size: 200Ti
#       profile: lustre
#     - kind: object
#       size: 50Ti
#   sovereignty: uk-ncsc-official
#   spendCap:
#     amount: 75000
#     currency: USD
#     window: monthly

SKU and region coverage#

The table below summarises practical coverage: per-region SKU availability and the typical role each region plays in the Yobitel-operated footprint. Capacity moves; treat the table as a sizing input and validate live availability in the SKU catalogue before committing to a reservation plan.

Limits and quotas#

Default per-account quotas exist to protect shared capacity during onboarding; almost every limit is raisable on request. Limits below apply per workspace unless noted.

Observability#

Every GPU instance emits the NVIDIA DCGM metric set (or AMD ROCm-SMI for MI300X) over the workspace's standard `yobitel_gpu_*` namespace — so platform teams scrape one stream regardless of vendor. The hosted Grafana dashboard surfaces SM occupancy, HBM utilisation, power draw, NVLink throughput, ECC error counts, and per-instance per-process attribution. Customers ship the same stream to their own Prometheus via the workspace's scrape endpoint or via OpenTelemetry to their observability platform.

Three signals matter most for raw GPU consumption: SM occupancy (is the GPU actually doing work?), HBM utilisation (is the workload memory-bound?), and ECC error counts (is the silicon healthy?). The PromQL block below is the alert most platform teams add first.

groups:
- name: yobitel-gpu-cloud
  interval: 30s
  rules:
  - alert: GpuSmOccupancyLow
    expr: |
      avg by (workspace, instance) (
        yobitel_gpu_sm_occupancy_ratio
      ) < 0.20
      and
      avg by (workspace, instance) (
        yobitel_gpu_power_watts
      ) > 50
    for: 30m
    labels: { severity: warn }
    annotations:
      summary: "{{ $labels.instance }} GPU SM occupancy below 20% under load"

  - alert: GpuEccCorrectableElevated
    expr: rate(yobitel_gpu_ecc_corrected_total[10m]) > 5
    for: 15m
    labels: { severity: warn }

  - alert: GpuEccUncorrectable
    expr: increase(yobitel_gpu_ecc_uncorrected_total[5m]) > 0
    for: 1m
    labels: { severity: page }

Cost and FinOps#

Pricing is per-second for on-demand, with material discount on 1-year and 3-year reservations and a floor-price spot tier for fault-tolerant workloads. The table below uses the same USD rates as the Yobibyte cost table — consuming a GPU through Yobibyte does not change the SKU rate, so the SKU price is the same whether the customer is buying directly or implicitly via the platform.

The FOCUS 1.1 billing export ships per-instance line items with the SKU, region, tenancy, term, and the workspace's tags attached, so spend pivots cleanly by SKU, by region, by team, or by call-off.

Security and compliance#

Tenancy is the primary isolation control: `shared` uses cgroup, namespace, and MIG isolation on multi-tenant nodes and is the right default; `dedicated` pins to a single-tenant node and is billed per node-hour for the strictest isolation without TEE; `confidential` uses NVIDIA confidential-compute on H100 and H200 with TEE attestation so encryption keys never leave the GPU.

Sovereignty is admission-enforced at provision time. An instance request in a workspace bound to UK NCSC OFFICIAL is refused if it targets a non-UK region; an EU Data Boundary workspace refuses placement outside the EU; a FedRAMP-equivalent workspace refuses placement outside the US. The refusal is recorded in the workspace audit stream.

• NCSC Cloud Security Principles — controls mapped per principle; UK OFFICIAL-tier regions audited annually.
• G-Cloud framework — Yobitel published in Lots 2, 3, and 9; sovereign GPU Cloud consumption procurable via the framework.
• ISO 27001:2022 / ISO 27017 / ISO 27018 — current certificates available under NDA.
• SOC 2 Type II — annual third-party audit.
• GDPR / UK DPA 2018 — DPA, sub-processor list, EU SCCs available; admission-enforced residency.
• HIPAA — BAA available; encryption, logging, access-control controls audited.
• FedRAMP-equivalent — Moderate-baseline-aligned controls available in US partner regions.
• CSA STAR Level 2 — published self-assessment plus third-party attestation.
• Cyber Essentials Plus — annual third-party assessment maintained.

Alternatives and customer-owned baseline#

Yobitel GPU Cloud is one option among several at the SKU level. The table below positions it against hyperscaler GPU products and specialist GPU clouds.

Troubleshooting#

The errors below cover the SKU-level provisioning failure modes seen most often. The full runbook library is at docs.yobitel.com/runbooks.

Where Yobitel GPU Cloud fits in the Yobitel stack#

Yobitel GPU Cloud is the SKU-level consumption surface; NeoCloud is the full operated-service contract Yobitel delivers to partners (facility, hardware, NOC, SLA), and Yobibyte is the platform layer above. The SKU surface sits between them: it is what NeoCloud-operated regions ultimately surface as, and it is what Yobibyte consumes implicitly when it provisions inferences and fine-tunes on the customer's behalf.

Practically, three customer types interact with the SKU surface differently. A platform-level customer almost never touches the SKU surface directly — they consume Yobibyte and let it pick capacity. A research lab consumes the SKU surface directly because the workload is a custom training job that does not fit the platform's abstractions. A finance team uses the SKU surface as a costing input — they look at per-SKU rates and reservation discounts to forecast spend.

Omniscient Compute indexes Yobitel GPU Cloud alongside other providers; Yobitel does not preferentially rank itself in the index. If a competing neocloud is the better answer for a workload, that is what Omniscient surfaces — the GPU cloud earns its position by being competitive on price, performance, and sovereignty rather than by being the default. That discipline is what keeps the index honest.