NEXUS
Sovereign AI Cloud Orchestration Platform
Turns bare-metal NVIDIA GPU infrastructure into a branded, multi-tenant sovereign AI cloud: GPU-as-a-Service, Model-as-a-Service and Fine-Tuning-as-a-Service with built-in FinOps.
- Three services on one engine: GPUaaS, MaaS and FTaaS
- OpenAI-compatible API gateway for open-weight models
- Fully white-labelled portal under the operator's own domain and pricing
- Per-request metering, quotas and spend caps for every tenant
Overview
Data centres that rent raw floor space and power capture little of the value created by AI workloads, while GPUs sit idle between large batch jobs. NEXUS is an orchestration platform that lets a data centre or State Data Centre operate its own branded AI cloud on its own hardware.
It provides three services on a single engine — GPU-as-a-Service with multi-tenant slicing and fair-queue scheduling, Model-as-a-Service through an OpenAI-compatible gateway, and Fine-Tuning-as-a-Service with LoRA and QLoRA pipelines — plus a FinOps control tower for metering, quotas and white-label billing. Customer data and fine-tuning stay inside the data centre.
Key capabilities
-
GPU slicing and time-sharing
Partitions GPUs into isolated instances with dedicated memory and compute, or time-shares them, to pack multiple tenant workloads without noisy-neighbour interference.
-
Cluster telemetry and failover
Monitors GPU temperature, power draw, PCIe bandwidth and VRAM residency, with automated workload re-routing and zero-downtime node evacuation.
-
Fair-queue scheduling
A fair-share queue in front of every model, with visible queue position for clients and priority pre-emption for enterprise SLAs.
-
OpenAI-compatible model gateway
Applications built for the OpenAI API, LangChain or LlamaIndex switch by changing the base URL.
-
High-throughput inference
Integrates vLLM, Ollama and llama.cpp backends with PagedAttention, continuous batching, streaming token routing and persistent VRAM residency.
-
LoRA and QLoRA fine-tuning
Distributed multi-GPU LoRA and 4-bit QLoRA runs with DeepSpeed ZeRO-3 / FSDP integration, without data leaving the data centre.
-
Dataset curation
Validation pipelines turn documents, JSONL conversations and SQL queries into tokenised training splits, with de-duplication and PII scrubbing.
-
Checkpoint and adapter manager
Version control for model weights; deploy trained adapters into the gateway with one-click rollback, and evaluate against benchmark suites.
-
Per-request metering
Captures prompt tokens, completion tokens, execution latency and GPU-seconds for every API call.
-
Quotas and spend caps
RPM / TPM rate limits and daily or monthly spend caps per tenant, with automatic throttling or pausing.
-
White-label portal
Custom domain, logos, brand palette and billing currency — the operator keeps the customer relationship.
-
Multi-organisation RBAC and SSO
Tenant organisations with role-based access control and single sign-on.
Architecture
NEXUS is partitioned into a control plane — IAM, tenants, models, FinOps, guardrails and telemetry — and a data plane: an OpenAI-compatible inference gateway and an intelligent router with failover across a local GPU pool, private VPC models and optional external providers.
-
Edge
- Sovereign TLS reverse proxy / WAF (TLS 1.3)
-
Control plane
- Nexus Control: tenants, RBAC, audit
- Nexus FinOps: token ledger, rate cards, budgets
- Nexus Guard: PII / credential DLP
- Telemetry
-
Data plane
- Nexus Gateway: OpenAI-compatible API
- Nexus Router: cost / latency routing, circuit breaking, failover
-
Compute
- Nexus Fabric: GPU telemetry, VRAM balancing, driver recovery
- Local GPU pool (NVIDIA L4 / A100 / H100)
- Nexus Forge: LoRA fine-tuning
Technical specifications
| Version | NEXUS v1.0.0 |
|---|
| Gateway endpoints | OpenAI-compatible /v1/chat/completions, /v1/embeddings, /v1/models with streaming |
|---|
| Targets | Local GPU pool, private-VPC custom models, optional external providers (e.g. OpenAI, Anthropic) |
|---|---|
| Policy | Cost- and latency-optimised routing, circuit breaking, zero-downtime failover |
| GPUs | NVIDIA L4, A100, H100 |
|---|---|
| Sharing modes | MIG hardware partitioning (on GPUs that support it) and time-shared slicing |
| Inference engines | Ollama, vLLM, llama.cpp |
|---|
| Nexus Forge | Parameter-efficient LoRA fine-tuning with checkpointing and GPU lease locking |
|---|
| Metering | Immutable token ledger, multi-tier rate cards, cost forecasting, budget governors |
|---|
| Nexus Guard | Zero-trust perimeter, PII and credential DLP sanitisation, abuse prevention |
|---|
| Components | Python control and data plane, Go TLS gateway, React / TypeScript management console |
|---|
| Stack | Linux kernel 5.15+, NVIDIA driver 535.129+ or 550.54+, CUDA 12.0+ (12.4 recommended), Docker CE 24+ / Compose v2.20+, NVIDIA Container Toolkit 1.14+, Python 3.11+, Go 1.22+ |
|---|
Deployment
-
Docker Compose
Containerised deployment with NVIDIA GPU acceleration.
-
Native host
Systemd services managed with the nexus-ctl CLI.
-
Operator data centre / State Data Centre
Branded multi-tenant AI services on the operator's own GPUs.
Requirements
| Operating system | Ubuntu 22.04 / 24.04 LTS or RHEL 9+ |
|---|---|
| CPU | Minimum 8 cores (x86_64); recommended 32+ cores (AMD EPYC / Intel Xeon) |
| RAM | Minimum 32 GB ECC; recommended 128 GB+ ECC DDR5 |
| GPU | Minimum 1× NVIDIA GPU with 16 GB+ VRAM (e.g. L4, T4); recommended 4–8× L4 / A100 / H100 |
| Storage / network | Minimum 100 GB NVMe and 1 Gbps; recommended 2 TB+ NVMe Gen4/Gen5 and 10–100 Gbps |
| Rollout plan | Days 1–3: cluster provisioning · Days 4–8: gateway and white-labelling · Days 9–14: FinOps and commercial go-live Source-reported |
Supported integrations
-
OpenAI SDKs, LangChain, LlamaIndex
Client compatibility
Existing applications work by changing the base URL.
-
vLLM, Ollama, llama.cpp
Inference engines
-
DeepSpeed, FSDP
Distributed training
-
SSO
Identity
Multi-organisation single sign-on.
Security & governance
-
Hard multi-tenant partitioning
Tenant workloads are isolated at the hardware or scheduler level.
-
Data stays in the data centre
Fine-tuning and inference run on the operator's own infrastructure.
-
PII scrubbing
Training datasets are de-duplicated and scrubbed of personal data.
-
Tenant governance
RBAC, SSO, rate limits and spend caps per organisation.
Use cases
-
Data centres
Branded sovereign AI cloud
Launch GPUaaS, MaaS and FTaaS under the data centre's own domain, brand and pricing.
-
Enterprises (BFSI, technology)
Private fine-tuning
Train custom domain models with LoRA / QLoRA without data leaving the data centre.
-
State Data Centres
Public-sector AI services
Offer hosted open-weight models to government departments on state infrastructure.
Documentation & developer resources
Cluster Provisioning
Architecture
Downloads
- NEXUS — Technical & Customer Deck
NEXUS technical and customer briefing (PowerPoint, 10 slides).
Frequently asked questions
Can existing OpenAI-based applications use NEXUS?
Yes. The gateway is OpenAI-compatible; applications built with the OpenAI SDK, LangChain or LlamaIndex change only their base URL.
Which GPUs can be shared between tenants?
NEXUS supports hardware partitioning (MIG) on GPUs that provide it and time-shared slicing otherwise. GPU models are confirmed during sizing.
Whose brand do customers see?
The operator's. The portal is fully white-labelled with the operator's domain, logo, colours and billing currency.
Items marked Source-reported are as stated in VedaCLOUD.AI product material and have not been independently verified.