Clastix Raises €2.9 Million as Mistral Bets on Kubernetes

Clastix, a Naples Kubernetes firm, has raised €2.9 million in seed funding with Mistral as both a customer and an investor. CDP Venture Capital Sgr led the round that closed on 24 September 2026, with Vertis SGR joining through its Vertis Venture 6 Digital Sud fund, which is partly financed by the European Union’s NextGenerationEU programme.

The cheque is seed-sized. The wager is that Europe’s next AI fight is won on the control plane, and Mistral is already training on that layer.

Mistral Is Buying the Pipes It Already Runs

Most seed stories ask a venture firm to believe a slide. This one asks a flagship European model lab to write a cheque for software it already runs in anger. Timothée Lacroix, co-founder and CTO of Mistral, said Clastix is how the lab keeps training and inference clusters up at the density frontier work needs, and he tied the equity to a wider European stack rather than to a tooling nice-to-have.

Clastix is the backbone of our Kubernetes infrastructure. Their work on multi-tenancy and managed control planes helps us to enable our training and inference clusters at scale, with the efficiency that frontier AI demands. By backing Clastix, we’re investing in Europe’s AI stack. A sovereign AI ecosystem isn’t just about models, it’s about the infrastructure layer that makes them run.

Timothée Lacroix, co-founder and CTO, Mistral

Clastix said the same thing in plainer language on the day of the round: Mistral runs training and inference on its control planes, and now holds equity. That double role is the bet. A lab that already feels the cost of every wasted control-plane box is backing the team that tried to delete those boxes.

€2.9M seed, led by CDP Venture Capital, with @MistralAI and Vertis SGR.

Mistral runs their training and inference clusters on our control planes. Now they're investors.#Kamaji stays Apache 2.0. #kMetal is where we take it next.#Kubernetes pic.twitter.com/UplxF07vOg

— CLASTIX (@clastixio) September 24, 2026

Adriano Pezzuto, co-founder and CEO, put the spend on the layer under the models rather than on another training run. “The next phase of AI won’t be decided by models alone but by who controls the infrastructure layer beneath them,” he said. “This funding goes toward building that layer for what’s coming: expanding our R&D to better support AI-scale workloads, and growing our team to bring it into new markets as they build their AI infrastructure.”

A €2.9 Million Seed With State and EU Money

The round is the company’s first external funding in its history, after six years of selling from open source. Dealroom’s tally puts the cheque in the top 5% of all-time Italian seed deals by size, which says as much about how small most Italian seeds are as it does about Clastix.

WHO WROTE THE CHEQUES

Alessandro Scortecci, chief investment officer for direct investments at CDP Venture Capital, called the physical AI build-out a hunt for infrastructure that stays efficient when the fleet gets large, and said the cheque fits a brief to back Italian technology that can strengthen Europe’s position. Renato Vannucci, partner and vice president at Vertis, pointed to customers of the calibre of Nvidia and Mistral and described the deal as a way to strengthen European AI infrastructure from Southern Italy.

WHAT THE MONEY PAYS FOR

• Product work: Further build-out of kMetal, the commercial fleet platform, including governance and production operations.
• Customer muscle: Customer success and solutions engineering so existing production accounts do not outrun a small team.
• Go-to-market: A wider sales motion as the company tries to take kMetal into new markets.
• AI-scale research: More engineering on the open-source Kamaji engine and on workloads that look like training and inference fleets.

Clastix said it expects to grow headcount over the next three years as those accounts get larger. It did not publish a current payroll figure or a hiring target.

How Kamaji Turns 90 Control-Plane Machines Into 3

Pezzuto and Dario Tranchitella, the CTO, founded Clastix in 2020 after running Kubernetes for large enterprises. The observation they kept repeating is brutally physical: a production cluster still burns three machines on API servers and etcd before a single application pod lands. An operator who stands up 30 tenant clusters in the usual way therefore parks 90 boxes on the control plane.

The Kamaji hosted control plane engine moves those components into pods on a small management cluster. Each tenant still gets a real, CNCF-conformant Kubernetes API, a dedicated datastore, full cluster-admin rights and its own lifecycle. The operator sees pods that reschedule when a management node dies, instead of a private trio of machines per tenant.

THE CONTROL PLANE TAX

Clastix calls that a 97% cut on the example, and it is the arithmetic that makes “a real cluster per tenant” affordable. In the March 2026 technical note on kMetal, Pezzuto wrote that a 30-tenant fleet on 3 shared nodes provisions in about 20 seconds, with control-plane upgrades finishing in about 16 seconds through blue and green pod replacement and no downtime.

Kamaji stays Apache 2.0. Clastix lists 2,000 GitHub stars on its homepage and presents the project as an operator you drop onto an existing cluster, not as another Kubernetes distribution. Worker nodes can join from bare metal, virtual machines or a public cloud, which is how a management plane in one site can drive tenants somewhere else.

FROM CAPSULE TO THE SEED

1. 29 June 2020: First commit lands on Capsule, the multi-tenancy project that groups namespaces into tenants.
2. 2020: Pezzuto and Tranchitella form Clastix in Naples around that multi-tenant problem.
3. 13 December 2022: The CNCF accepts Capsule at sandbox maturity, and the code leaves the company proper.
4. 11 March 2026: Clastix publishes the first kMetal technical note, folding hosted control planes into a bare-metal fleet product.
5. 24 September 2026: The company takes €2.9 million of outside capital for the first time.

Capsule was the namespace-level product. Kamaji is the cluster-level one. kMetal is the attempt to sell the rest of the stack without taking Kamaji private.

kMetal Keeps Kamaji Open and Sells the Fleet

Hosted control planes remove the three-machine tax. They do not, on their own, isolate tenant kernels, carve up networks, or keep a thousand clusters on the same profile. That is the gap kMetal is meant to fill, and it is where a seed round of this size will be spent or wasted.

kMetal embeds KVM through KubeVirt so each tenant’s workers run in their own virtual machines, then uses Kube-OVN to give each tenant a separate virtual network, its own IP space and its own DNS. Pezzuto’s March note is blunt about the alternative: namespace filters and shared overlays are fine in a lab and a poor fit for a cloud provider, a regulated estate or an AI factory that hands tenants cluster-admin. Fleet profiles then define version, storage, networking, add-ons and quotas, and a reconciliation loop puts mandatory pieces back if a tenant deletes them.

The same note lists what the product refuses to become. Storage stays bring-your-own through any CSI driver. The KVM layer is for tenant isolation, not a general-purpose hypervisor for Windows estates. GPU passthrough is in; training job orchestrators such as Kubeflow, Ray or Run:ai are the tenant’s problem. Idle tenants can sleep the control plane and drop toward zero, which is the detail that matters if the paying customer is an AI lab with bursty jobs rather than a telco with flat traffic.

Kamaji remains the proof that the architecture works. kMetal is the commercial wrap for isolation, autoscaling, multi-site HA, self-service and support contracts. Engineers can still stand Kamaji up alone and decide later whether they want to pay.

NVIDIA, Fastweb and OVHcloud Already Run It

Clastix says the stack is already in production with infrastructure teams in Europe and North America across AI labs, telecommunications, cloud services and the public sector. Named logos around the round and on the company’s own site include Mistral, NVIDIA, Fastweb, OVHcloud and Rackspace. In an April 2025 essay, Clastix said NVIDIA selected Kamaji to power its DOCA platform; that claim sits in the company’s own voice, not in a separate NVIDIA filing.

We are proud of the step forward we have taken with the new architecture, and we are very pleased to have partnered with the Clastix team for their deep expertise and unique approach to Kubernetes.

Luca Tagliavia, Head of ICT Multicloud BU, Fastweb

Jérémie Monsinjon, head of containers at OVHcloud, said Kamaji “works exactly as expected: it’s simple, efficient, scalable,” and that Clastix stays available for the ugly technical arguments. On the Kamaji site, Kevin Carter, a director at Rackspace, said the open-source project is running inside the Rackspace Spot platform. TINEXT Cloud, a Swiss operator, is a later production name on the same homepage.

That customer mix cuts two ways for the sovereignty story Lacroix and CDP are telling. The code is European-engineered and the new capital is Italian and EU-flavoured. The production footprint already includes a US cloud provider and NVIDIA. If the bet is “Europe owns the pipes,” the pipes are already carrying other people’s water, which is how an infrastructure company actually gets paid.

HyperShift and vCluster Already Occupy This Fight

Hosted control planes are not a secret Naples invention. Red Hat ships the same broad idea inside OpenShift as hosted control planes, still built on the HyperShift operator even after the product name moved on. vCluster Labs sells a different isolation model that virtualises a Kubernetes API for tenants and now pitches hard tenancy for GPU clouds. Platform teams who live in Cluster API also have k0smotron and a growing list of control-plane providers.

Clastix’s claim against that field is narrow. Kamaji speaks Cluster API, emits upstream kubeadm clusters rather than a wrapped distro, and keeps the engine on Apache 2.0 so a cloud provider can build a managed Kubernetes offer without taking a new control-plane religion. kMetal then tries to productise the hard multi-tenancy those providers still have to assemble: kernels, VPCs, profiles, sleep, GPU passthrough. Pezzuto’s own note draws a line against “virtual clusters” that only project a filtered view of a host API, calling that fine for dev and a problem when the tenant is a bank or an AI factory.

A €2.9 million seed will not out-spend Red Hat or a well-funded virtual-cluster vendor. It can keep a 2,000-star engine alive, put more people on kMetal, and give Mistral a supplier that now shares the downside if training clusters get more expensive to operate. That is a real wager. It is also still a seed.

The next test is boring and measurable. Either kMetal turns the 90-to-3 arithmetic into contracts that outgrow a seed payroll, or Kamaji stays a respected operator that other people’s platforms quietly embed. Mistral, CDP and Vertis just paid to find out which.

Frequently Asked Questions

What is Kamaji and is it a Kubernetes distribution?

Kamaji is an operator you install on a management cluster so tenant control planes run as pods, and Clastix is explicit that it is not a new distribution: clusters are built with upstream Kubernetes binaries and kubeadm so they stay CNCF-conformant. The name comes from the boiler-room character in the film Spirited Away. GitHub documentation puts a tenant control plane in a position to serve traffic in seconds once the management cluster is healthy, which is why the company sells it as managed Kubernetes-as-a-service plumbing rather than as another opinionated stack.

When did Clastix hand Capsule to the CNCF?

The Cloud Native Computing Foundation accepted Capsule as a CNCF sandbox on 13 December 2022, two and a half years after the project’s first commit on 29 June 2020. CNCF’s public insights page now lists 984 contributors and 192 contributing organisations, which is the community footprint Clastix gave up when it donated the namespace-level tenancy tool and moved its commercial centre of gravity to Kamaji and kMetal.

Who founded Clastix and who else is in the leadership team?

Adriano Pezzuto (CEO) and Dario Tranchitella (CTO) founded the company in 2020 after years as engineers running Kubernetes inside large enterprises, a history Pezzuto now summarises as more than 10 years on Kubernetes and more than 25 years in infrastructure. Gaetano Pastore is chief revenue officer. The public leadership picture the company used with the seed announcement is those three together, which is the scale of firm that just took its first outside capital.

How is a hosted control plane different from a virtual cluster?

A Kamaji tenant gets its own API server, scheduler, controller manager and dedicated etcd, plus cluster-admin on a CNCF-conformant cluster that can run a different Kubernetes version from its neighbour. A virtual cluster typically presents a Kubernetes-like API that is a translation or projection over a shared host cluster, which is faster to spin up and weaker as a tenancy boundary when two untrusted parties share a kernel and a network. Clastix’s kMetal design assumes the stronger boundary, which is why it adds KVM workers and per-tenant OVN networks on top of the hosted control plane rather than stopping at a filtered API.