Building production AI workloads on a single GPU
XRobotix set out to answer a practical question: could a South African team run modern AI workloads entirely on locally hosted GPU infrastructure without relying on overseas AI providers?
Using a single NVIDIA RTX PRO 4000 Blackwell GPU hosted in xneelo’s Samrand Data Centre, the team spent several weeks building, testing and optimising a wide range of AI workloads.
What began as an infrastructure trial evolved into StudioX, an AI-powered video production platform capable of supporting:
- Large language models
- Image generation
- Video generation
- Music generation
- AI-assisted SaaS advert creation
- Automated video production workflows
Why GPU infrastructure mattered
For XRobotix, the goal wasn’t simply to access GPU compute. The team wanted infrastructure that could support production AI workloads while providing:
- Predictable monthly pricing
- Low-latency local inference
- Full control over AI workloads
- Data sovereignty for sensitive information
Many AI development workflows still rely on overseas cloud platforms and usage-based APIs. XRobotix wanted to explore whether locally hosted GPU infrastructure could provide a practical alternative for South African businesses building AI-powered products.
How it works:
XRobotix deployed its AI environment on a single NVIDIA RTX PRO 4000 Blackwell GPU with 24GB of VRAM.
The solution combines:
- Ollama behind a LiteLLM proxy for local language model inference
- ComfyUI image generation workflows
- SDXL and Z-Image models
- LTX video generation
- Django, DRF, HTMX and Celery orchestration
- Containerised AI services and rendering pipelines
During the trial, the team also identified several software and rendering optimisations that improved GPU utilisation and reduced rendering times, highlighting how infrastructure and optimisation work together to maximise AI performance.
Business impact
Running on a single locally hosted GPU enabled XRobotix to build and operate a production AI platform while achieving several operational benefits.
The project delivered:
- Fast local AI inference with low-latency access
- Predictable infrastructure costs compared with usage-based API billing
- Reduced reliance on overseas AI providers
- Greater control over proprietary data and generated content
- A production-ready AI video generation platform developed and tested entirely in South Africa
“What surprised us wasn’t simply that it worked. It was how much we could build on a single GPU. We ended up creating an entire AI-powered video production platform on top of it.”
What the team learned
The project also highlighted some practical considerations for organisations exploring AI infrastructure.
According to XRobotix:
- A 24GB GPU is a strong entry point for many production AI workloads.
- Software optimisation can significantly improve AI rendering performance.
- Larger frontier-scale models and highly concurrent workloads require additional GPU memory and compute resources.
- AI video and content generation workflows are now commercially viable on relatively compact GPU infrastructure.
The experience demonstrated that many AI applications no longer require large-scale enterprise infrastructure to become operational, while also highlighting where larger GPU deployments become beneficial.
Planning to extend into perception and control
The success of Project Onyx laid the foundation for XRobotix’s next phase of development: extending its locally hosted GPU infrastructure beyond content generation into perception and control.
The team is planning to extend its xneelo-hosted environment beyond generation and into perception and control, using vision language models (VLMs) to help robots understand what they see, and reinforcement learning (RL) to teach them how to move.
The same 24GB NVIDIA RTX PRO 4000 Blackwell GPU that powers StudioX will also serve as the development bench for an autonomous robot fleet, hosting VLMs locally to interpret camera feeds and evaluate robot behaviour, while driving the RL iteration loops that shape how each robot walks, balances and handles objects.
“The video platform proved the infrastructure. Now we’re pointing it at the physical world. The same GPU that generates our adverts will help our robots learn to see and to move.”
For this next phase, the 24GB VRAM footprint remains a strong starting point. It can comfortably host production-grade VLMs for low-latency local inference and support rapid reinforcement learning experimentation and policy evaluation, while keeping sensitive robotics and customer data on local infrastructure.
As the fleet grows and training scales up, XRobotix expects to pair this local development bench with larger GPU deployments for full-scale policy training, following the same pattern of high performance and data sovereignty at greater scale.
Key focus areas include:
- VLMs hosted locally for robot perception and automated evaluation of training runs
- Reinforcement learning for locomotion, balance and manipulation policies
- Data sovereignty by keeping sensitive robotics and customer data on South African infrastructure
- A clear path to scale from a local development bench to larger GPU deployments for frontier-scale training
“The question was never just ‘can it generate video’. It was ‘can this become the backbone of everything we build next’, and increasingly, the answer is yes.”
