The ASUS ExpertCenter Pro ET900N from Altron Arrow

South African organisations have stopped asking what AI might do for them. The questions now are operational: how to deploy it, where it should run, how much compute it needs, and what that means for their data, their budgets and their technology strategy.

For years, AI adoption was framed as a software conversation — which platform to adopt, which model to train, which use cases to prioritise. The bottleneck has moved. GPUs, memory, storage, networking, power, cooling, security and data management now determine whether AI gets from proof of concept into production. Access to the right compute is becoming as important as access to the right model.

The infrastructure challenge in South Africa

Local organisations face a particular set of constraints. High-performance compute is expensive and access to advanced GPU infrastructure is limited. Bandwidth and latency affect cloud-based AI performance. Electricity supply and data-centre capacity complicate large-scale planning: Africa has less than half a gigawatt of active data centre capacity for more than a billion people. Data sovereignty adds another layer for anyone handling regulated information, and specialised AI infrastructure skills are in short supply.

South Africa cannot simply replicate the hyperscale models built elsewhere. The assumption that every organisation needs a large data-centre footprint or a major cloud commitment is no longer viable — and it is no longer necessary.

A huge data centre is not always needed

As AI adoption matures, workloads are becoming more distributed. Developers need compute for model training and testing. Business units need local inference for analytics, automation and computer vision. Operational environments, from factories to retail stores, increasingly rely on edge AI to process data in real time.

That creates a role for compact, distributed infrastructure. Rather than centralising compute in a single data centre or cloud environment, organisations can deploy smaller AI-capable systems at different points in the business, providing meaningful compute without the cost and lead times of a traditional AI data-centre build.

The rise of accessible AI compute

A new generation of compact systems is emerging to meet this need, designed to put AI compute closer to users, developers and operational environments so that organisations can start small, experiment locally and scale with real demand.

The ASUS Ascent GX10 is one example. Built on Nvidia’s GB10 Grace Blackwell Superchip — the same silicon as Nvidia’s DGX Spark — it delivers a petaflop of FP4 AI performance and 128GB of unified memory in a chassis measuring 150x150x51mm, enough to fine-tune models of up to 200 billion parameters on a desk. Two units can be linked over ConnectX-7 networking to double that. It draws up to 180W, which matters in a market where power is a planning constraint rather than a line item.
The ASUS ExpertCenter Pro ET900N G3 brings workstation-class performance to business and professional settings, supporting AI workloads that need to sit closer to the user in engineering, analytics, design, research or operational roles. For many organisations it is a practical middle ground: powerful enough for demanding AI tasks, accessible enough to deploy across teams and departments.

The practical effect of systems like these is local development and testing, faster experimentation cycles, inference close to the point of use, tighter control over sensitive workloads, less reliance on cloud connectivity, and a realistic starting point for an organisation building its own AI capability.

It is no longer cloud versus on-premises

For years, infrastructure decisions were framed as a binary: cloud or on-premises. AI breaks that framing, because different workloads have different requirements, and those requirements decide where a workload should run.

Some AI workloads benefit from the elasticity and scale of cloud. Others need low latency, local processing, data sovereignty, predictable costs or offline capability, which points towards on-premises or edge infrastructure.

So, the future is hybrid. Organisations will blend cloud, on-premises and edge compute according to the nature of each workload, and compact systems such as the GX10 and the ET900N become part of that model, letting them place compute exactly where it is needed. The question is which location suits which workload.

From proof of concept to production

Many proofs of concept succeed technically and fail operationally, because infrastructure planning was treated as an afterthought. Closing that gap means assessing current and future workloads, the split between training and inference, GPU and memory requirements, where data sits and which sovereignty rules apply, security, power and cooling, networking and latency, total cost of ownership, and how the environment scales.

That assessment belongs inside an organisation’s AI strategy rather than beside it. Infrastructure determines what is possible, how quickly AI can be deployed and how economically it can be sustained.

The role of the technology ecosystem

Making AI accessible takes more than selling hardware. It requires an ecosystem: compute, infrastructure, software, cybersecurity, distribution, technical expertise and channel partners.

Altron Arrow, which distributes ASUS AI systems in South Africa, works within that ecosystem to help organisations and partners get hold of what modern AI deployments require. The aim is compute that is realistic, scalable and economically viable for local organisations, rather than a scaled-down copy of a hyperscale build.

The next stage of AI adoption

Accessibility will shape the next phase of AI adoption. The organisations that benefit most are likely to be those that understand their own workloads and build an infrastructure strategy around them, rather than those with the largest budgets. For South African businesses, affordability is no longer the barrier it was; what matters now is how quickly AI moves from experiment to operational advantage.