Reserve the fabric
your model
actually needs.
We are building dedicated NVIDIA B200, H200 and H100 clusters — InfiniBand-connected, liquid-cooled, zero egress. Reserve capacity before it is racked and we build the topology around your workload, not around a catalogue.
What we commit to
Why reserve before we rack
GPU model, node count, fabric ratio and storage tier are set to your run — not inherited from a catalogue SKU someone else designed.
The reserved price is locked when you sign, so you are not exposed to the spot market on the day you actually need the capacity.
A reservation costs nothing to hold. Billing starts the day you accept the cluster against its benchmarks — not the day you sign.
01 / The economics
One product,
priced like
one product.
Hyperscalers run hundreds of services and price GPU rental to subsidise the rest. We will run one thing: dense accelerated compute. No abstraction tax, no bundled services you never open, no egress bill at the end of the quarter.
- ✓ Published hourly rates — no "contact us for pricing"
- ✓ $0 ingress, $0 egress, $0 inter-node transfer
- ✓ Reserved discounts to 58% — cancel-for-convenience clause
Indicative launch pricing, derived from the H100 rate card. Fixed for you at reservation.
02 / The hardware
The platforms we build on.
Vendor specifications are NVIDIA's published figures for these platforms. None of this hardware is in service yet. Rates are indicative launch pricing in USD, exclusive of tax — not a quotation and not an offer capable of acceptance.
03 / Reference architecture
Interconnect is the
whole ballgame.
A thousand GPUs on a slow fabric is a thousand GPUs waiting. That is why we are specifying rail-optimised from day one rather than retrofitting later — non-blocking NDR InfiniBand, NVLink within the node, and parallel storage that keeps the accelerators fed. This is the design every COCO cluster gets built to.
04 / Rental model
A single card,
or the whole machine.
Take one GPU for an afternoon of debugging, or a full bare-metal node with the NVLink mesh intact. Same hardware, same fabric — you pick the granularity.
A dedicated GPU passed through to your instance. Provisioned in minutes, metered per second, stopped whenever you like.
- ✓ Dedicated GPU passthrough — no time-slicing
- ✓ Hourly, daily or monthly billing
- ✓ Persistent volume survives instance restarts
- ✓ Jupyter, SSH and API access from the console
The entire node, no hypervisor in the way. Full NVLink + NVSwitch mesh across all eight accelerators and dedicated InfiniBand ports for scaling out.
- ✓ Bare metal — root access, your own OS image
- ✓ 900 GB/s all-to-all NVLink inside the node
- ✓ 8× NDR 400G ports for multi-node training
- ✓ No noisy neighbours, no virtualisation overhead
05 / How long you hold it
Three ways to hold capacity.
Mix them. Most customers reserve a training baseline and burst on-demand for evals.
06 / How we're building it
Designed for the
audit you'll run later.
We would rather commit to these controls in writing now than retrofit them after a customer's security review. Formal certifications follow once clusters are in service — we will publish them when we hold them, and not before.
Regions are being finalised now. Tell us where your data has to sit and we will confirm whether it is in scope for the first build.
07 / Get a quote
Tell us the shape
of the workload.
A solutions engineer — not a bot, not an SDR script — comes back within one business day. We will tell you plainly what is in scope for the first build and what is not.
Prefer email? sales@cococloud.biz
08 / Questions