.png){kind=logo}
Cyril Connan runs a production platform that looks cloud-native from the developer’s side but sits on hardware his company controls. As Head of Infrastructure at Mobula, he designed the company’s bare-metal infrastructure and migrated its platform away from Google Cloud Platform. Today, the environment includes multiple Kubernetes clusters across three availability zones, a 45TB Postgres database, and infrastructure supporting an average of roughly 4,000 to 5,000 HTTP requests per second, not including WebSocket traffic. For Connan, the significance is not simply that the system runs without a hyperscaler. It is that the company can still deploy and operate like a modern software organization while owning the machines underneath it.
“Cloud-native describes how you build and operate the platform,” Connan says. “It does not mean you have to rent the underlying infrastructure from AWS or GCP.”
That idea has shaped much of his career. Before joining Mobula, Connan spent three years in professional services at OVHcloud, where he worked with companies migrating away from AWS and GCP. The recurring issues included cloud bills, lock-in, and sovereign-cloud requirements. By the time he arrived at Mobula, he had already worked with customers dealing with the cost, control, and sovereignty consequences of earlier cloud decisions.
“Public cloud solves a real problem at the beginning,” Connan says. “It is fast and easy to start there. The tradeoffs become more important later, when cost, performance, control, or lock-in start affecting what the company needs from its infrastructure.”
At Mobula, his response was not simply to replace one hosting vendor with another. Connan designed dedicated infrastructure and then built a platform layer intended to preserve the advantages engineering teams expect from modern cloud environments. That includes Kubernetes, GitOps through ArgoCD, internal GitHub runners, VPN access, SSO, and 2FA. The aim is to let product teams deploy and work without managing the underlying machines while keeping the servers, network, databases, and security architecture under the company’s control.
The migration began with installing and configuring the bare-metal servers, followed by deploying Kubernetes and the platform tools on top of them, including CNI, CSI, and GitOps. Connan then created a new Postgres cluster and replicated the existing database from GCP, deployed the applications, and ultimately shifted traffic through DNS. He personally led that path from the physical infrastructure through the platform layer.
The migration was phased rather than handled through a single cutover. Mobula built the new platform in parallel with the GCP environment and gradually redirected clients to the new clusters through DNS. The transition was completed without downtime because traffic could be moved incrementally. The architecture also allows several fully autonomous clusters to operate at the same time, with traffic load-balanced or redirected between them when necessary.
Those details matter because moving infrastructure away from a hyperscaler is not simply a purchasing decision. A platform running on dedicated hardware has to replace functions that managed cloud services previously handled. Compute has to be provisioned, storage has to be designed, networks have to be operated, and deployments still need to remain predictable for developers. Connan’s work spans those layers rather than ending at the Kubernetes cluster itself.
“The goal is not to make engineers think about servers all day,” he says. “The platform should let them ship normally while I still understand and control what is underneath.”
The 45TB Postgres environment makes that responsibility particularly concrete. At that scale, the database affects decisions about networking, storage, latency, availability, and the surrounding platform. Connan’s infrastructure work therefore has to account for how the database interacts with the rest of the system rather than treating it as an interchangeable managed service. The machines, network, storage, and data layer become connected parts of the architecture.
The move also changed Mobula’s infrastructure economics. On GCP, the company was spending approximately $50,000 to $60,000 per month while running one GKE cluster, dedicated Postgres nodes, and less traffic than the platform handles today. The bare-metal environment costs approximately $30,000 to $35,000 per month while supporting three Kubernetes clusters, including staging and production in Paris and another production cluster in Singapore, with substantially more compute and workload capacity. Database performance has also improved significantly. A team member working closest to that layer estimates an improvement of at least five times, although Connan treats that figure as an operational estimate rather than a controlled benchmark.
Connan does not reduce the argument to a claim that bare metal is automatically cheaper or faster. His position is more specific. Public cloud offers speed and convenience, particularly early on, but dedicated infrastructure can give teams more control over hardware placement, networking, database configuration, and where workloads physically run. Whether that trade makes sense depends on the workload and the organization operating it.
That qualification is important to the way Connan talks about infrastructure. He is not interested in treating hyperscalers as a mistake. In fact, he describes Google as a source of inspiration and does not frame his work as a campaign against major cloud providers. His criticism is directed at the assumption that choosing cloud-native architecture requires giving up the ability to own or choose the underlying infrastructure.
“What people sometimes confuse is convenience with architecture,” he says. “Kubernetes, microservices, GitOps, rolling upgrades, and modern deployment practices are architectural choices. Where the machines live is another decision.”
His experience at OVHcloud reinforced that separation. Connan worked directly with customers who had already encountered problems with cloud cost, lock-in, or sovereignty. Those migrations gave him firsthand experience evaluating infrastructure after those constraints had become operational concerns rather than abstract possibilities. That experience now informs the way he evaluates infrastructure choices at Mobula.
In representative professional-services engagements at OVHcloud, Connan worked with companies already running on AWS or GCP because of concerns around cost, lock-in, or sovereignty, as well as organizations operating environments built on VMware, Docker Swarm, or other stacks. His role was to understand the existing environment, determine which infrastructure approach fit the customer’s actual requirements, and help move the workloads onto that platform. The work did not follow a single migration recipe because the technical constraints and starting architecture differed from one organization to another.
Connan’s work at Mobula reflects a broader principle he has carried through his career: infrastructure choices should remain technical decisions that engineers understand well enough to question. In practice, that means running a modern software platform on dedicated hardware while continuing to provide the deployment and security systems developers expect. Connan does not argue that every company should leave the public cloud. Mobula instead gives him a production-scale example of what another architecture can look like.
For Connan, the question is whether an organization understands its workload well enough to make the choice deliberately. His work at Mobula shows what that can look like when a company keeps cloud-native software practices while taking responsibility for the machines beneath them.
