Just the core — Compute Nodes, Supervisor, API Gateway, and Model Storage — deployed on your own infrastructure and wired into the identity, logging, and monitoring you already run. No platform services to rip out, no operational handover, no ongoing vendor dependency. Annual runtime license with a Software Support SLA on ARK components.
ARK Core is the streamlined tier for organisations with an existing platform stack that want to adopt the ARK inference engine without pulling in the full managed platform. You operate it. You integrate it. We license the software and support the ARK components.
Keycloak, ELK, and Prometheus/Grafana are production-hardened. You need the inference engine — not another platform to manage.
Public-sector, defence, and regulated deployments where vendor footprint must be minimal. Air-gap supported. Source-of-truth for identity, logging, and telemetry stays with you.
Sovereign cloud providers, MSPs, and SaaS platforms integrating ARK as the inference layer inside an existing platform product — not bolting on a second control plane.
ARK Core ships the proprietary inference engine and the components that make it run — nothing more. You bring your own identity provider, log aggregator, and metrics stack; the ARK runtime plugs into them through standard interfaces. If you ever need the managed platform services, the upgrade path to ARK Tailored is a configuration change, not a migration.
ARK Runtime — inference engine with heterogeneous GPU support, elastic hot-scaling, and session-level KV isolation
Compute Nodes — execution runtimes bound to specific devices; any mix of GPU generations and vendors
Supervisor Node — orchestration, shard routing, master-master replication
API Gateway — edge proxy + OpenAI v1 / Anthropic-compatible API service
Model Storage — Hugging Face-format shared repository (use ARK's or point at your own backend)
Identity — bring your own Keycloak, Okta, Entra, or OIDC-compatible IdP for tenancy and API-key management
Logging — stream structured logs into your existing ELK, Splunk, Datadog, or SIEM
Telemetry — Prometheus-compatible metrics (including nvidia-smi) consumed by your Grafana, Datadog, or observability stack
Monitoring — wire alerts into your own on-call tooling (PagerDuty, Opsgenie, Splunk On-Call, Datadog)
Portal — not included. Add via ARK Tailored if you need a chat & embeddings UI for non-technical teams
Standard integration is the same as you'd build with any LLM provider — ARK just changes where the inference happens, and what crosses your perimeter to get there.
ARK Core is designed for teams that want minimal vendor involvement in day-to-day operations. Install using ARK-supplied artefacts and manuals — or bring us in for an installation guidance package. Either way, once it is running, your team owns the operational loop; ARK supports the software.
Annual runtime license sized by GPUs under license. Pick the modalities and model catalogue that match your workload; upgrade at any time.
Your team installs using ARK's reference architecture, container images, and manuals. Optional installation guidance package available if you want ARK engineers alongside.
Wire ARK into your existing IdP, logging pipeline, metrics stack, and API gateway. Standard interfaces; no bespoke glue required.
Your team runs the platform. ARK provides a Software Support SLA covering the ARK runtime components — response times and escalation paths scale with GPU license volume.
ARK Core is sold as an annual runtime license, sized by GPUs under license. No managed-services fee, no platform-services surcharge — you only pay for the runtime. Optional add-ons cover extended model catalogues, additional modalities, and one-time installation guidance.
Priced per GPU under license. Base allocation includes 10 models and text modality.
Software Support SLA on ARK runtime components. Standard / Premium / Enterprise, auto-selected by license size.
Upgrade from 10-model base to a 10–20 model catalogue.
Per additional modality (image, vision, embeddings). Text is always included.
One-time installation guidance package available on request. Professional services for integration or custom configuration priced separately.
Everything that makes ARK defensible against vLLM, TensorRT-LLM, and Ollama is in the runtime itself — not the surrounding platform. ARK Core gives you that engine, on your infrastructure, without the rest.
Run H100, H200, B200, A100, MI300X, MI325X, and Gaudi 3 in the same cluster. No identical-hardware requirement. No re-provisioning when a generation rolls over.
Multi-host inference runs at around 5 Mbit/s per session. No NVLink or InfiniBand fabric required — ARK runs on the network you already own.
Add or remove compute at runtime. No reload, no session drops, no maintenance window. Continues operating after 90–99% GPU failure.
Session-level KV isolation lets multiple tenants or workloads share the same GPU fleet without leaking context — the audit story regulators expect.
Choose ARK Core when you have a mature platform team and only need the runtime. Choose ARK Tailored when you want the full managed platform on your infrastructure. Choose ARK Cloud for zero-ops serverless inference.
Per-token pricing. Best-effort 99% availability. Zero ops.
Annual license. Integrate with your own platform stack.
Runtime + platform services, deployed and supported by ARK.
Tell us about your GPU footprint, existing platform stack, and integration requirements. We will come back with a license quote, reference architecture, and installation plan.