NEXUS architecture: three service pillars

GPUaaS, MaaS and FTaaS on one orchestration engine with a FinOps control tower.

NEXUS Updated 1 min read
On this page
  1. GPU-as-a-Service
  2. Model-as-a-Service
  3. Fine-Tuning-as-a-Service
  4. FinOps control tower

NEXUS turns bare-metal GPU clusters into tenant-facing AI services operated under the data centre’s own brand.

GPU-as-a-Service

Multi-tenant GPU partitioning (MIG on supporting GPUs) and time-shared slicing, hardware health telemetry (temperature, power, PCIe bandwidth, VRAM residency), automated failover with zero-downtime node evacuation, and fair-share queues with visible queue position and priority pre-emption.

Model-as-a-Service

An OpenAI-compatible gateway for open-weight models, backed by vLLM, Ollama or llama.cpp with PagedAttention and continuous batching. Clients migrate by changing the base URL:

from openai import OpenAI
client = OpenAI(base_url="https://ai.your-datacenter.example/v1", api_key="YOUR_TENANT_KEY")

Fine-Tuning-as-a-Service

Distributed LoRA and 4-bit QLoRA runs (DeepSpeed ZeRO-3 / FSDP), automated dataset curation with de-duplication and PII scrubbing, and a checkpoint and adapter manager that hot-swaps adapters into the gateway with one-click rollback.

FinOps control tower

Per-request metering (prompt and completion tokens, latency, GPU-seconds), RPM / TPM quotas, daily and monthly spend caps, pre-paid and post-paid billing, multi-organisation RBAC and SSO, and a white-label portal.