NVIDIA's RTX Spark is a small form-factor desktop module with 128GB of unified memory and 273 TOPS of AI compute. It runs AI models up to 120 billion parameters — models in the same capability bracket as some of what was accessible only via expensive cloud APIs eighteen months ago. The Spark is not a cloud server. It sits on a desk.
What 120 Billion Parameters Means in Practice
Model parameter counts are an imperfect proxy for quality, but a useful one. The frontier models most businesses access via ChatGPT or Claude sit at several hundred billion parameters in their larger variants. A 120B dense model is not at that level, but it handles the workload that accounts for the majority of real business use: writing, summarising, coding assistance, document analysis, answering questions about internal information.
The gap between "good enough for business" and "frontier performance" is real, and it matters for hard reasoning tasks. For volume work — drafting, sorting, extracting — it mostly does not.
Why On-Premise AI Is Interesting for South African Businesses
The data stays in the building. For accountants, legal firms, medical practices, and anyone handling information covered by POPIA, that is not a nice-to-have. For some clients and some regulators, sending files to a US cloud provider to be processed by an AI, even with a solid data processing agreement, creates a chain of custody question they will not accept.
The second point is cost structure. Cloud AI is priced in USD and billed monthly. The rand-dollar rate makes API fees expensive in local terms. Hardware has a one-time cost. Once amortised over the device's lifespan, the per-query cost approaches zero.
The Honest Limits
The RTX Spark is not cheap. It will land in South Africa at a price that makes it a considered purchase. It is a fit for businesses large enough to justify dedicated AI infrastructure and with a specific reason to keep processing on-premise.
For most small SA businesses, the cloud APIs remain the practical choice. But the Spark's arrival shows the direction of travel: capable AI hardware is shrinking, the models that once required a data centre are running on a device that fits in a carry-on bag, and the cost will keep coming down. The businesses thinking about on-premise AI today are early, not wrong.