Platform

Three products. One sovereign platform.

GPU compute and managed inference with the developer experience of a public cloud — minus the foreign jurisdiction, and with the option to run entirely disconnected.

Product 1

Managed models

A curated catalog of open-weight models served on vLLM and SGLang behind an OpenAI-compatible API — deployed to a production endpoint in one click.

Curated open-weights catalog

Mistral 7B as the default, plus Mistral NeMo, Mixtral and Qwen-Coder — served on vLLM and SGLang, tuned for production latency and throughput.

Per-project API keys

Every project issues its own scoped keys with configurable rate limits. Keys are revocable instantly and every use is written to the audit log.

Live NOK metering

Token counts and endpoint GPU time metered per project, per endpoint and per key — priced in Norwegian kroner, with hard quotas and no opaque credit systems.

Product 2

Your models

Bring your own trained or fine-tuned models. Upload the weights into your own isolated storage and deploy them one-click — they never cross your isolation boundary.

Your isolation boundary

Weights upload into your own isolated storage and stay there. They never cross your isolation boundary — not to a shared registry, not to another tenant, not to ArcemCloud.

One-click deployment

Deploy an uploaded model to a production endpoint in one action, served on the same vLLM/SGLang stack and the same OpenAI-compatible API as the catalog.

Built for proprietary models

Ideal for defense contractors and others running trained or fine-tuned models that must not leave their control — with the same metering, keys and audit trail.

Product 3

GPU compute

Rent dedicated GPUs when you need raw capacity — for training, fine-tuning or anything else that needs a shell and a CUDA driver.

Dedicated GPUs, SSH access

Whole GPUs, never shared, with direct SSH access and curated base images for training, fine-tuning and batch workloads.

Stop/start billing per GPU-second

Billing runs only while the instance does. Stop it and it costs nothing — no reservations you forget about, no rounding to the hour.

Fine-tune, then serve

Train or fine-tune on your data, then serve the result as a managed endpoint on the same platform — one workflow from experiment to production.

An API you already know

Endpoints speak the OpenAI wire format — whether they serve a catalog model or one you brought yourself. Existing SDKs, agents and evaluation harnesses work unmodified: swap the base URL and the key.

arcemcloud — inference

curl https://api.arcemcloud.no/v1/chat/completions \
  -H "Authorization: Bearer $ARCEMCLOUD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-nemo",
    "messages": [
      {"role": "user", "content": "Summarize NSM guidance on cloud sovereignty."}
    ]
  }'
$

Across all three

One platform underneath

OpenAI-compatible API

Point existing SDKs, agents and evaluation harnesses at your ArcemCloud endpoint by changing the base URL. No proprietary client, no rewrite, no lock-in.

Grafana observability

Dashboards for latency, throughput, token counts and per-GPU telemetry: utilization, memory, temperature and power. The same stack ships with air-gapped deployments.

Sovereign by construction

Norwegian jurisdiction, hard namespace isolation, default-deny networking, tamper-evident audit and air-gap capability — detailed on the security page.

How it works

1
Create a project
A project is the unit of isolation: its own namespace, quota, API keys and usage ledger.
2
Deploy a model or rent GPUs
Pick a catalog model, upload your own, or provision dedicated GPUs. Live and metered within minutes.
3
Call the API
Use any OpenAI-compatible client with your project key. Rate limits and usage are enforced and visible in real time.

Isolation, audit and air-gap details are on the security page. Prices are on the pricing page.