What we hear from early neocloud networking teams.
Four things that come up in almost every conversation.
Tenants are provisioned by standalone scripts, and the information lives in a spreadsheet
Several controllers and consoles, context switched by hand
A reference architecture that keeps getting modified
Network operations and monitoring are not correlated with GPU utilisation
How Aviz solves the complexities of AI factories.
One orchestration layer — network, GPU, and telemetry — integrated upward and abstracting everything below, plus the engineers who run it with you.
Nothing already deployed has to be removed for this to start working.
How you build your AI factory with ONES.
The same three-stage path applies whether you’re standing up a sovereign AI cluster, a private AI environment, or a neocloud. One agentless software appliance — VM or bare metal, out of the data path — replaces manual work with automation and consistency, aligned to the reference architecture.
Design and validate
- RA-driven designs, planned for growth and expansion
- Simulate end-to-end networking before hardware arrives with NVIDIA DSX Air
NVIDIA DSX Air · Aviz ONE Center with FTAS
Operate multi-tenancy
- Network abstraction through the ONES operator and API
- Automatic tenant segmentation, workload isolation via API
Rafay · Red Hat · Spectro Cloud · vCluster
Monitor and operate
- Network monitoring, forecast and prediction
- Lifecycle management and troubleshooting
ServiceNow · Zendesk · Slack
ONES isn’t a one-time setup. It stays connected to your orchestration tools, NVIDIA’s controllers, and DSX Air for as long as the factory runs — end to end, from an empty switch to a tenant running on its own isolated slice of the network.
Operationalizing the factory through telemetry.
Next step: a quick proof of concept.
Start with the ONES OVA, AI-Parser, or AI-NOC — with the partner of your choice. No hardware required to begin.
Deploy
ONES OVA, AI-Parser or AI-NOC, in your environment.
Validate
Your topology proved in the DSX Air twin, before hardware.
Prove the value
One tenant end to end, or one week of real token data.
Licensed per GPU, under the ONES AI Factory license.
Priced by the number of GPUs under management — not by device count, use case, or token volume.
Subscription terms. 1-, 3-, or 5-year terms.
Reference architectures
Aviz supports NVIDIA Spectrum-X RA 2.1 and Aviz RA 1.0.
Aviz RA 1.0 options
Includes Spectrum-X RA 1.3 and InfiniBand designs, for rack- or server-based deployments.
Per-GPU coverage
One license per GPU covers every east–west and north–south network that GPU’s fabric needs.
Technical questions.
What GPU architectures does ONES cover today?
H100, H200, B200, GB200, B300, and GB300. B300 quad-plane, RTX PRO east–west AI Grid, east–west Ethernet on SONiC, custom designs, and Aviz Gateway Service are roadmap items.
Does ONES support 128,000 GPUs?
128K is NVIDIA’s published Spectrum-X multiplane scale, not a published ONES management ceiling.
Which DPU generation does Aviz support?
BlueField-3 or ConnectX, for north–south networking. BlueField-4 (NVIDIA’s published 800 Gb/s platform) is not claimed here.
Where does DSX Air fit?
ONES defines and operationalizes the network design; NVIDIA DSX Air is the simulation environment that validates it before physical deployment.
How is this licensed?
Per GPU under management, under the ONES AI Factory license, covering Spectrum-X RA 2.1 and Aviz RA 1.0 designs, on 1-, 3-, or 5-year terms.
Do we need hardware on-site to start?
No — a POC starts with the ONES OVA, AI-Parser, or AI-NOC, validated in the DSX Air twin, no hardware required to begin.
An integration project becomes a repeatable production capability.
Aviz ONES adds the operational layer for a repeatable, multi-tenant platform.
Maximize GPU performance while simplifying operations.
Shorter path from hardware to productive GPU capacity.
Integrations proved together in DSX Air, before production.
API-driven tenant onboarding, not manual changes.
Consistent segmentation across tenants and workloads.
One model across H100 through NVL rack-scale.