
Kubernetes-native GPU orchestration on Moroccan soil. H100s from $1.80/hr. 90-second provisioning. 5 hubs. 99.98% SLA. Data stays in Africa.
HarchOS is the operating system for sovereign AI compute — built and operated in Morocco, powered by renewable energy, and open-source from day one. Every workload runs under Moroccan data sovereignty law (Loi 09-08). No US CLOUD Act exposure. No GDPR conflicts. No data replication outside the continent. Just clean, fast, carbon-aware GPU compute — accessible from a single CLI command.
Every inference call, every training run, every dataset leaves the continent by default. The current model is broken — technically, legally, and economically.
HarchOS operates on a three-layer architecture inspired by autonomous systems — observe the grid, predict the optimal path, execute with precision. Every layer is open-source and auditable.
Workload arrives via SDK or CLI
THINK layer scores every hub
ACT layer schedules and deploys
Carbon footprint reported per job
Continuous ingestion of carbon intensity data from national grid APIs, energy spot pricing, GPU utilization metrics, and inter-hub network telemetry.
ML-powered scheduling engine that predicts the optimal hub, time window, and GPU configuration. Optimizes the Pareto frontier of cost, carbon, and performance.
Deploys, migrates, and scales workloads automatically. Supports live migration between hubs when carbon conditions shift — moving training from Tanger to Ouarzazate when solar peaks.
OpenAI-compatible inference with built-in carbon tracking. Every response includes gCO2 emitted, renewable percentage, and the serving hub. Async workloads, SSE streaming, and a context manager for carbon-aware batch processing.
from harchos import HarchOS
# Sovereign by default. Carbon-aware by default.
client = HarchOS(
region="morocco",
sovereignty="strict",
carbon_aware=True,
)
# Train on 32 H100s — auto-routed to Ouarzazate solar hub
job = client.train(
model="darija-13b",
gpus=32,
gpu_type="h100",
dataset="s3://harchos/darija-corpus",
)
print(f"Trained on {job.hub}"
f"Carbon: {job.carbon_gco2} gCO₂"
f"Renewable: {job.renewable_pct}%")Scale from 1 to 1,024 GPUs in 90 seconds. NVLink, InfiniBand, liquid cooling. Powered by 100% renewable energy.
| GPU | Memory | On-demand | Spot | Best for |
|---|---|---|---|---|
H100 SXM5Popular | 80 GB HBM3 | $1.80/hr | $0.72/hr | LLM training |
H200 SXM5 | 141 GB HBM3e | $2.40/hr | $0.96/hr | Long-context inference |
A100 80GB | 80 GB HBM2e | $1.10/hr | $0.44/hr | Fine-tuning |
L40S | 48 GB GDDR6 | $0.65/hr | $0.26/hr | Inference edge |
Submarine cable connectivity via Marea, ACE, and AA-1. Redundant mesh routing. Sub-50ms latency from African edge. No data leaves the continent.

Real-time cluster telemetry — from the Ouarzazate hub. Refreshed every 5 seconds.
Drag the sliders. See your monthly bill. Compare to AWS p5.48xlarge. All prices in USD.
Estimates only. Spot pricing available at 40% discount. Reserved capacity commits 1-year for additional 20% off.
Built different. Built sovereign. Built in Africa.
HarchOS | AWS | GCP | Azure | |
|---|---|---|---|---|
| Sovereignty enforced at infra level | ||||
| No US CLOUD Act exposure | ||||
| Moroccan Law 09-08 compliant | ||||
| Carbon-aware scheduler | ||||
| Per-job carbon footprint | ||||
| Open-source SDKs + CLI | ||||
| Zero egress between African regions | ||||
| Terraform provider with sovereignty HCL |
Data sovereignty isn't a feature — it's the architecture. HarchOS enforces sovereignty at three levels, and once set, it can never be downgraded.
Full compliance with Morocco’s data protection law. CNDP registered. Personal data never leaves the kingdom.
No US corporate presence. We do not respond to US extraterritorial data requests. Sovereign by jurisdiction.
Workloads, weights, datasets, logs — everything stays on African soil. Replication is continental, never intercontinental.
HarchOS doesn't just run workloads — it understands the carbon impact of every compute decision and actively minimizes it.

Deploy in one command. Pay per token. Sub-50ms latency from the African edge.
Moroccan dialect LLM trained on HarchOS. Best for North African NLP.
Sahel-region LLM covering 12 West African languages.
East African flagship model. Sub-50ms inference from Abidjan hub.
First production-grade Amazigh language model. Indigenous AI matters.
No DevOps team required. Just a terminal.
Each hub is strategically located near renewable energy sources. Combined, they form Africa's largest planned GPU mesh — 1,798 GPUs (target) across solar, wind, and hybrid installations.





“HarchOS let us train a 13B parameter model on Moroccan soil, under Moroccan law. No US exposure. No GDPR gymnastics. Just compute.”
“We cut our inference cost by 73% migrating from AWS p5 to HarchOS. Same latency. Carbon-aware by default. Game-changer.”
“For phosphate processing analytics, sovereignty is non-negotiable. HarchOS enforces it at the infrastructure level — we audited the code.”
Every workload deploys with region=morocco, sovereignty=strict, and carbonAware=true — without you setting anything. To run outside Africa, you would have to explicitly opt out. The default is the secure path.
Everything is open-source. Everything is documented. Everything is auditable.
Full REST + gRPC API. OpenAPI 3.1 spec. SDKs in Python, TypeScript, Go, Rust.
OpenCarbon tracking, async workloads, SSE streaming, batch processing. Production patterns.
OpenSENSE/THINK/ACT deep dive. Hub topology, scheduler internals, sovereignty enforcement.
Open
Free $500 credit. H100 instances in 90 seconds. Data stays in Africa. No credit card required.