Sovereign cloud · Bare metal · GPU · FR
Infrastructure designed, built and operated in France. Four data centres in a dual ring at 400 Gbit/s, 3.8 ms p99 between sites. Throughput guaranteed by contract, up to 10 Gbit/s per server. 99.99% SLA, penalties credited without you having to claim.
isengard@edge — live reading
Automatic reading. Full history and post-mortems are shared with clients under contract.
Key figures
4sites
Normandy, Île-de-France, Grand Est, South. No site outside the country.
99.99%
Monthly, measured at the edge. Penalties credited without a claim.
30s
From detection to neutralisation. Total capacity: 6 Tbit/s.
10Gbit/s
Per server, guaranteed by contract and never overcommitted. 1 Gbit/s baseline.
Reading of 31 December 2025. Availability and mitigation are measured continuously; the incident history is shared with clients under contract.
Three engineering principles
Every claim on this site corresponds to a metric we take ourselves, in production. No value is lifted from a manufacturer’s data sheet.
We publish percentiles, not averages. Inter-data-centre latency at 3.8 ms p99, 1.24 million IOPS on 4K random reads across our NVMe volumes, a bare server provisioned in 112 seconds through the API.
Six independent layers, each sized to hold on its own: edge filtering, volumetric mitigation, L2 segmentation, application firewall, system hardening, SOC correlation. No layer assumes the one before it held.
Hardware, network, hypervisor, orchestration and operations stay in our hands, in France. No administrative access, no backup and no log leaves the country. This is verified in operations, not in a clause.
Service lines
Cloud, bare metal, GPU, backup, security and colocation share the same API, the same network and the same on-call rota. You are not wiring six suppliers together.
A cloud of your own, operated at ours.
Dedicated hypervisor, dedicated hardware, dedicated network. You get a console and an API; we keep responsibility for the substrate. No resource is shared with another client, at any level of the stack.
No hypervisor, no neighbour, no standard deviation.
For workloads that pay for every layer of abstraction: transactional databases, single-threaded compute engines, trading platforms. You get the bare machine, delivered by API in under two minutes.
Train and infer without queueing.
GPU nodes reserved or billed by the hour, interconnected at 200 Gbit/s and backed by a parallel file system. Capacity is guaranteed by reservation: you do not compete with other clients for your cards.
Client feedback
p99: 210 ms → 74 ms
340 instances migrated in six weeks. Our application p99 went from 210 ms to 74 ms, without a line of code changed. The gain came entirely from storage and network.
Karim B. · CTO
Logistics SaaS platform, 90 staff
Observed RTO: 11 min
Annual failover test: DR triggered at 09:02, production running normally on the standby site at 09:13. Eleven minutes, the record signed the same day. Our audit committee had nothing to add.
Nathalie F. · CIO
Industrial group, 2,400 staff
Provisioning: 40 min → 90 s
The Terraform provider was incomplete when we arrived: per-project quotas were not exposed, and we waited five weeks for the fix. Since then a complete staging environment comes up in ninety seconds against forty minutes before. What I take from it is that they shipped the feature rather than explaining why we did not need it.
Yanis M. · Lead developer
Payments software vendor, 45 staff
Our clients speak under a first name and a surname initial. Company names are published only with written consent; most prefer not to.
Pre-sales
Describe your volumes, your latency constraints and your target RPO. A pre-sales engineer replies within two working hours with a diagram, a sizing and an estimate.