HUMAIN Brings MI355X and 800G Ethernet AI Cluster Online but Does Not Disclose Current GPU Count or Power Capacity

AMD, Cisco, and HUMAIN, a company owned by Saudi Arabia’s Public Investment Fund (PIF), announced that the first batch of Instinct MI355X infrastructure has entered production and begun serving customers inside and outside the country. The publicly disclosed architecture combines MI355X GPUs, EPYC CPUs, and Cisco Nexus 9000 Series networking powered by Cisco Silicon One. Inter-node connectivity is provided by 800G optical links, while HUMAIN packages the resources as GPU-as-a-service for training and inference. Unlike announcements limited to procurement or power-capacity targets, the material difference in this update is that customer workloads are now running.

The design also represents another production deployment of a large AMD cluster using Ethernet scale-out networking. Distributed training and tensor parallelism for large models frequently perform collective communications such as all-reduce and all-gather. If switch queuing, congestion control, or optical-link configuration is poorly balanced, utilization can decline rapidly even when ample theoretical GPU compute is available. Cisco’s public AI POD documentation shows that Nexus 9000 switches can form lossless 400G/800G RoCEv2 networks. However, this announcement does not disclose the actual topology, external bandwidth per GPU, RCCL efficiency, tenant-isolation approach, or end-to-end job completion time. It therefore provides insufficient evidence to conclude that the cluster has achieved any specific level of training efficiency.

For the next phase, the three companies plan to begin deploying facilities of up to 250 MW in 2027 using the MI400 Series and ROCm. The joint venture’s longer-term target remains up to 1 GW by 2030. These figures should not be conflated with the capacity currently online: the announcement provides no current GPU count, data-center power capacity, number of locations, or customer utilization rate. It also does not explain how the 250 MW plan relates to the previously proposed 100 MW first phase based on MI450. In other words, what has been confirmed is the launch of MI355X services—not the delivery of hundreds of megawatts of capacity.

Engineering teams should next focus on three categories of data: reproducible ROCm and RCCL performance over this Ethernet fabric; the GPU service’s failure domains and multi-tenant QoS; and whether migration to MI400 will require changes to images, core libraries, and network configuration. Only if HUMAIN publishes model throughput, latency percentiles, collective-communication efficiency, and availability metrics will there be enough evidence to assess whether this sovereign AI cloud can deliver stable production performance in addition to expanding capacity.