Webinar · Took place online on Thursday 10 September 2026
About 50 minutes with Andrei Kvapil, the creator of Cozystack: turn the GPUs you already own into a multi-tenant platform teams can use — or customers can buy — from bare metal to a metered inference endpoint, assembled entirely from open source.
Every team wants AI infrastructure, and the market sells it as a product you buy whole. In reality a GPU platform is a stack of layers — sharing, tenancy, inference, metering — and each one already has a mature open-source project waiting. The gap is not the hardware. It is the assembly, and that is exactly what this session walks through.
Expensive cards run at twenty or thirty per cent while one team waits on another, because there is no sharing, no quotas and no self-service. Or the budget bleeds into hyperscaler GPU instances for work that could run on hardware in your own racks, with the data staying home. Or the platform team is asked to "hand the GPUs out" and discovers there is no clean way to do it.
None of these calls for new hardware. They call for a way to turn the cards you already own into a platform teams can use — or customers can buy.
Cozystack is an open-source cloud platform and CNCF Sandbox project that turns a node with GPUs into GPU-ready tenant clusters, managed databases, S3-compatible storage and inference — each one a Kubernetes resource behind one API. Teams get their own cluster, not a namespace, with the cards visible inside.
On top of it, Ænix modules turn the platform into a business: GPU usage measured per tenant, tenant quotas, a self-service panel, and integrations that pass consumption to your billing system (WHMCS or your own), where charging happens.
Not pilots. Production platforms serving real users, built on the same foundation the session walks through.
A European academic-computing platform moved off a public hyperscaler onto owned bare metal, with fractional GPU sharing across jobs and a single Cluster API spanning bare metal, a hyperscaler and a sovereign Swiss cloud — bursting to rented GPUs only when demand spikes.
A Swiss cloud provider sells GPU beside virtual machines, managed Kubernetes and databases from one platform — whole-card and shared, metered and billed with its own panel. Its own engineers became Cozystack maintainers.
Both are anonymised at the customer's request. Andrei walks through what each of them actually did with GPUs — and what he would do differently on your stack.
Live demos, not slides — then your questions.
From bare metal to a GPU-ready tenant. A node with GPUs joins the cluster, drivers roll out, and a team gets its own cluster — not a namespace — with the cards visible inside.
Ways to allocate a GPU. One whole card (passthrough) or a share of it — vGPU and fractional (HAMi) — side by side, with the utilization graph that turns idle silicon into money.
From a model to an endpoint. vLLM in a tenant, then NVIDIA Dynamo — disaggregated prefill/decode and KV-cache-aware routing — with before/after latency under concurrent load on the same hardware.
Three doors to one API. Order a GPU cluster from a UI, from the CLI, or with a Git commit — the same resource every time, no portal on top of a portal.
Charging for GPU use. The session discussed GPU-hours for renting capacity and input/output tokens for selling inference as pricing patterns. In the platform, GPU usage is measured per tenant and charging happens in your billing system (WHMCS or your own).
When your service isn't in the catalog. Package it once and deploy it with a button — plus where queues (Kueue, Volcano, KAI) fit for distributed training.
The session ended with a live Q&A with the speaker.
A layer-by-layer map of a GPU platform — and the open-source project that fills each one.
A one-page decision matrix for sharing GPUs: passthrough vs vGPU vs fractional, and when each wins.
A clear-eyed view of metering — per GPU-hour and per token — and where the honest limits are.
A reference runbook for the whole path, so your team can reproduce it on your own cluster.
Enterprise teams handed a pile of GPUs and told to make them useful, and cloud, telecom and GPU providers building a sovereign AI offering. If you're weighing an open-source build against Run:ai, OpenShift AI, or simply renting from a hyperscaler, the session was built around your situation.
Especially the people who own the GPUs and the decision:
Andrei created Cozystack, the open-source cloud platform and CNCF Sandbox project, after more than fifteen years of building clouds and high-load infrastructure. He contributes to Kubernetes, KubeVirt, Cilium and LINSTOR, and speaks at KubeCon and other industry events. At Ænix, he helps teams turn their GPUs and hardware into commercial cloud services.
It ran online on Thursday 10 September 2026. The recording is not published yet.
The team can go through a GPU platform for your own setup. For hands-on sessions coming up, see the workshops.
Webinar · Took place online on Thursday 10 September 2026
The webinar took place on 10 September 2026, and the recording is not published yet. Talk to the team about a GPU platform for your own setup.