Mistral's CTO puts managed control planes under the company's training clusters

The cost Kamaji goes after is structural. A Kubernetes control plane runs the components that manage a cluster, and the traditional layout gives each cluster its own dedicated machines for them, so hardware and operational work grow with every new cluster ^[4]. Kamaji runs those components as pods inside a shared management cluster instead ^[5]. Creating a tenant cluster then means scheduling pods on capacity you already operate.

Clastix's technical documentation describes that model, and the company says the claimed efficiency benefit depends on how an operator configures and uses its infrastructure ^[9]. That caveat is honest, and it sets the condition for the saving transferring to anyone else. Per-cluster control-plane hardware has to be a real line in your bill: many clusters, most of them small, each sitting on machines you sized for redundancy instead of load. Run six large clusters and the pods buy you almost nothing.

Mistral invested in the round ^[2], and Timothee Lacroix described the investment as support for Europe's AI infrastructure ^[10]. The customer reference is also a shareholder. Tech.eu did not report a cluster count or the share of Mistral's fleet involved, so the supported statement is a narrow one: a company that trains and serves its own models operates some of that estate on managed, multi-tenant control planes ^[3].

Capsule, the founders' earlier open-source multi-tenancy project, came before Kamaji ^[17]. Kamaji itself stays open source under Apache 2.0, and kMetal is the paid management layer around it ^[6][7]. The split follows the argument Adriano Pezzuto, Clastix's co-founder and CEO, made in March: hosted control planes address only part of the cost of running multi-tenant Kubernetes ^[11][16]. kMetal brings infrastructure management and isolation together with the control plane ^[12], and version 1.0 reached general availability on August 31st ^[8]. Tech.eu reported the seed 24 days later ^[1][18].

Clastix expects to grow headcount over the next three years and did not attach a hiring target to the plan ^[14]. "The next phase of AI won't be decided by models alone but by who controls the infrastructure layer beneath them," Pezzuto told Tech.eu ^[15]. In my view the commercial shape is the right one at this size, because the free component keeps adoption cheap and the paid layer aims at bare metal, where the machines are the expense ^[4]. What has to be proven is that a team already running Kamaji for nothing will pay for the fleet management above it ^[6][7].