Project Onyx: We Ran Frontier AI on ONE 24GB GPU in Johannesburg

Four weeks ago, @xneeloHosting exclusively loaned us one Blackwell-class GPU server. We ran every kind of AI workload we could throw at it: large language models, image generation, lip-synced video, AI-composed SaaS adverts, and built a full video production platform end-to-end on top. This entire video was generated on that one card.

The whole thing runs in South Africa. No prompts, no audio, no generated content left the country.

📖 Full technical write-up: https://portal.xrx.info/s/ProjectONYX/

━━━━━━━━━━━━━━━━━━━━━━━
🛠 WHAT WE RAN ON IT
Language models (via Ollama + LiteLLM gateway):

Gemma 4 26B at ~76 tokens/sec
GLM 4.7 Flash (MoE, 256K context)
Qwen 3.6 27B

Image and video:

SDXL Base + Z-Image family for stills
LTX-2.3 22B for image-to-video with audio conditioning and lip-sync

Composition and agent stack:

Hyperframes (HeyGen, Apache 2.0) for HTML/CSS/GSAP rendering
Claude Code in a sandboxed Docker container as the AI video composer
Cloudflare Tunnel for secure public access

━━━━━━━━━━━━━━━━━━━━━━━
🎯 WHAT WE LEARNED
A 24GB GPU in South Africa is the sensible entry point for serious on-prem AI today. It replaces 70 to 90 percent of typical hosted-API spend for a small team, with full data sovereignty and single-digit-millisecond local latency.

Two breakthroughs from the trial that changed how we build:

Chrome ANGLE Vulkan flag on Blackwell delivers 3 to 4 times faster headless GPU rendering. Free performance on a single flag.
AI agents composing video directly in HTML, CSS, and GSAP beats diffusion-generated UI imagery in every quality dimension. Diffusion is wrong for legible product screenshots.

━━━━━━━━━━━━━━━━━━━━━━━
🇿🇦 Hosted by Xneelo in Johannesburg → xneelo.co.za
🌐 More from XRobotix → xrx.co.za
💬 Reach out for commercial AI video production or self-hosted AI consulting
━━━━━━━━━━━━━━━━━━━━━━━
#SouthAfrica #AI #MachineLearning #LocalHosting #OnPremAI #DataSovereignty #POPIA #LLM #VideoGeneration #Xneelo #XRobotix #Blackwell #NVIDIA #ClaudeAgent #Hyperframes #LTXVideo #ComfyUI #StudioX

Видео Project Onyx: We Ran Frontier AI on ONE 24GB GPU in Johannesburg канала XRobotix Pty Ltd