← Back to the six layers
Software & operations

Software & Operations.

Between a running cluster and a business sit the software plane and the operating team: provisioning, scheduling, metering, billing, and people on call around the clock. This layer is what turns your compute into revenue.

PROVISIONED · METERED · SOLD
Plan your operating layer →
What this layer is

Where compute becomes revenue.

Once the fabric is lit you own a working supercomputer. It starts earning when a customer can request a slice of it, use it, and receive an invoice for exactly what they consumed.

It is the last layer, and the one customers interact with directly: the control plane that provisions and isolates tenants, the scheduler that keeps expensive GPUs full, the metering that turns usage into a bill, and the operations team that responds when a job stalls overnight.

The prize this layer unlocks

What decides the margin.

$180 B
projected neocloud and GPU-as-a-service revenue by 2030, all of it flowing through this layer.
24/7
the operating standard a paying tenant assumes: monitored, staffed and answerable at every hour, not a cluster someone checks on in the morning.
Utilization
the single number that decides whether the site earns its capital back. Idle GPUs draw power and depreciate while earning nothing.
From cluster to service

What a live cluster still needs.

Four capabilities separate a working cluster from a service a customer can buy and rely on.

01

A real control plane

Customers need to request capacity and stay isolated from one another, down to the network and the storage. On shared GPUs that is a hard problem in its own right.

02

Scheduling that keeps GPUs full

An orchestrator, Slurm for training runs or Kubernetes for services, places jobs, queues them, and packs the cluster so idle silicon is the exception.

03

Metering you can bill against

Every GPU-hour, every byte moved and stored, has to be captured accurately enough to put on an invoice a customer will pay without dispute.

04

SLAs and the team behind them

An uptime commitment is only as real as the monitoring, on-call rotation and day-2 operations standing behind it.

Why it is hard to assemble alone

Where new operators lose the margin.

Utilization is the whole margin

A GPU fleet at half utilization often earns nothing at all, because capital and power costs do not shrink with the idle hours. Keeping the cluster sold and scheduled is continuous work.

Multi-tenancy is a trust problem

Renting the same hardware to strangers means guaranteeing that one tenant can never see, starve or touch another. Isolation failures are rare but existential for a provider, so the architecture has to be right from the first tenant.

Day-2 operations never sleep

Nodes fail, jobs hang and drivers drift for as long as the cluster runs, and a paying customer expects each incident handled before they notice it. That takes staffing, monitoring and runbooks in place from day one.

What we bring

An operating layer that turns compute into revenue.

  • The right software stack for your model: orchestration, provisioning and multi-tenancy matched to whether you sell bare training capacity or a managed AI platform.
  • Metering and billing that hold up: usage captured to the GPU-hour and turned into invoices your tenants pay without argument.
  • An operations plan before you sign a tenant: monitoring, SLAs, on-call and day-2 runbooks, costed and staffed.
  • A go-to-market that fills the fleet: the packaging and pricing that keep utilization high.
Design your operating layer →
SCHEDULED · METERED · SOLD
Who we bring to this layer

The names that run the cloud.

Orchestration, GPU-cloud software and operating tooling are specialist work. We select the stack and the partners and integrate them with the rest of the build.

Anyone can own the hardware. The margin belongs to whoever can sell it and keep it running.

How will your compute actually get sold?

Tell us the cluster you are standing up and who you want to sell it to. We'll come back with an honest read on the software and operations layer it needs.

Start a conversation →