HUMAIN launches an NVIDIA-powered AI cloud with 1.1 megawatts live

HUMAIN has launched an AI cloud in Saudi Arabia running on NVIDIA HGX B300 Blackwell Ultra systems, announced at LEAP 2026. The locally hosted platform is live with 1.1 megawatts of operational AI infrastructure, which the company says is already running at above 90 percent utilisation, and it plans to expand to 35.1 megawatts by 2027. Those two numbers are the ones that matter, and it is unusual to get both. Most AI infrastructure announcements in the region give a target capacity and a date, with nothing to say about what is switched on and whether anyone is using it. HUMAIN has published a live figure and a utilisation figure alongside the plan, which makes the launch far easier to assess than the announcements around it. What it says is that the platform is small and full. At 1.1 megawatts, this is a single-room deployment rather than a data centre campus, in the order of a few racks of Blackwell Ultra systems. HGX B300 is NVIDIA's eight-GPU Blackwell Ultra platform, aimed at large-scale training and high-throughput inference, and each of those systems draws serious power, which is why a megawatt figure at this scale translates into a modest number of nodes. Utilisation above 90 percent on a cloud that has just launched means demand is not the constraint. Capacity is. Against the plan, the live number is very small. The 35.1 megawatt target for 2027 is more than thirty times the current deployment. The wider HUMAIN and NVIDIA partnership, announced in May 2025, set out a projected build of up to 500 megawatts powered by several hundred thousand NVIDIA GPUs over five years, with a first phase described as an 18,000-GPU GB300 Grace Blackwell supercomputer. What is live today is roughly two tenths of one percent of that headline. Both things are true at once, and reporting either without the other gives a false picture. This is also not HUMAIN's only compute track. On the opening day of LEAP, AMD, Cisco and HUMAIN said AMD Instinct MI355X systems running on Cisco networking had gone live in the Kingdom and were serving customers in Saudi Arabia and beyond, with plans to deploy 13,000 AMD GPUs this year, add up to 250 megawatts starting in 2027 and reach as much as one gigawatt by 2030. HUMAIN is therefore standing up NVIDIA and AMD estates in parallel, which is a deliberate choice rather than an accident of procurement. Running both matters for buyers. Multi-vendor silicon gives HUMAIN negotiating leverage and supply resilience at a time when frontier GPU allocation is the binding constraint on every AI cloud in the world. It also means customers need to ask which estate they are being sold, because the software path is not identical. A workload built on CUDA does not move to Instinct hardware for free, and the porting effort is real even with the maturity that ROCm has gained. Anyone signing with HUMAIN should establish which accelerator family their capacity sits on, what the migration story is if they are moved, and how pricing differs between the two. There are things HUMAIN has not published. There is no price list, no statement of instance types or minimum commitment, no published queue or lead time for new customers, and no service level commitment. On a platform running above 90 percent utilisation, availability is the question a buyer would ask first, and the answer is not in the announcement. Nor has the company published which models are served, what the data residency and access controls look like in contractual terms, or how sovereign the sovereignty is when the underlying silicon, firmware and toolchain come from abroad. HUMAIN is owned by Saudi Arabia's Public Investment Fund and chaired by Crown Prince Mohammed bin Salman, and it has been the single most active entity in Saudi AI infrastructure over the past two years, with announced work spanning data centres at NEOM, Arabic language models, an AI PC, an enterprise operating system and partnerships with most of the large Western vendors. The AI cloud is the layer where all of that has to turn into something a customer can buy by the hour. The next proof point is capacity. A jump from 1.1 to 35.1 megawatts inside roughly fifteen months is an aggressive build in any market, and it depends on grid connections, cooling and GPU delivery as much as on demand. If utilisation stays above 90 percent as capacity grows, that is a genuine signal about Saudi AI demand. If capacity arrives and utilisation falls, that is a different and equally useful signal.