AWS has launched AI Factories, a new offering that deploys dedicated AI infrastructure inside customers’ own data centres rather than AWS’s. It’s positioned as a solution for organisations that want AWS’s AI capabilities but can’t or won’t move their data and workloads to the public cloud, usually for sovereignty, regulatory, or compliance reasons. The pitch is simple: you provide the data centre space and power, AWS handles everything else.
AI Factories combine NVIDIA GPUs, Trainium chips, AWS networking, storage, databases, and AI services like Amazon Bedrock and SageMaker into a private AWS Region operated exclusively for a single customer. It’s essentially AWS infrastructure running on your premises, managed by AWS but isolated from the public cloud. The setup lets organisations keep data local while accessing the same tools and services available in standard AWS Regions.
The partnership with NVIDIA is central here. AWS AI Factories include access to NVIDIA’s Grace Blackwell and the upcoming Vera Rubin platforms, along with the full NVIDIA AI software stack and thousands of GPU-accelerated applications. AWS’s Nitro System, Elastic Fabric Adapter networking, and EC2 UltraClusters support these platforms, and future Trainium4 chips will integrate NVIDIA NVLink Fusion interconnect technology for better performance and flexibility.
AWS is really selling speed and operational simplicity. Building AI infrastructure in-house requires capital investment in GPUs, data centres, power, procurement cycles, model licensing, and operational expertise. Most organisations don’t want to deal with that complexity, and AWS claims it can deploy AI Factories faster than customers can build equivalent infrastructure themselves. The trade-off is that you’re still locked into AWS’s ecosystem, just in your own building.
The first major AI Factory deployment is in Saudi Arabia, where AWS is building an “AI Zone” for HUMAIN featuring up to 150,000 AI chips including NVIDIA GB300 GPUs. It’s a massive project and a signal that sovereign AI infrastructure is a real market, particularly in regions where governments want local control over AI capabilities. The Saudi deal is also a multi-gigawatt commitment, which suggests AWS sees AI Factories as a long-term revenue stream.
For governments and regulated industries, AI Factories address a genuine problem. Public sector organisations often face legal or political requirements to keep data within national borders, and industries like finance and healthcare have strict regulatory frameworks. AI Factories let these organisations access AWS’s AI tools without violating sovereignty or compliance rules, which is why AWS is pitching the service hard to government customers.
The catch is cost and commitment. AI Factories require significant capital expenditure for data centre space and power, and they’re only viable for organisations running large-scale AI workloads. Smaller operations won’t justify the investment, and even large organisations need to be confident they’ll use the infrastructure long enough to recoup costs. AWS hasn’t published pricing, but it’s safe to assume AI Factories aren’t cheap.
For South African organisations, AI Factories could be relevant if data sovereignty becomes a regulatory requirement or if local AI development reaches a scale that justifies dedicated infrastructure. The South African government has discussed data localisation policies, and if those materialise, AI Factories might be a practical solution. In the private sector, large banks, telecoms, and mining companies could theoretically benefit, but only if their AI ambitions are substantial enough to warrant the investment.
AWS’s broader play here is obvious: extend its infrastructure into spaces where the public cloud can’t reach. By offering AI Factories, AWS can compete for workloads that would otherwise go to on-premises or sovereign cloud providers. It’s a hedge against regulatory shifts and a way to capture revenue from organisations that want AWS’s capabilities but not its shared infrastructure.