FLEXPOD AI AFF A90 Sep 23, 2026
The disclosed reference stack
CANCOM describes a configuration with four NVIDIA RTX PRO 6000 GPUs, two Intel Xeon 6 processors, Cisco Nexus networking at up to 400GbE, and a NetApp AFF A90 with 69 TB usable capacity. The software layer includes Red Hat OpenShift AI and NVIDIA AI Enterprise. CANCOM positions it for on-premises generative AI, machine learning, RAG, and vision AI, with preinstalled or customer-provided workloads available for evaluation.
A further Sovereign & Secure AI Factory offering adds governance, security, and sovereignty positioning. The public material does not publish the AFF drive layout, ONTAP release, network topology, storage protocol, validated component firmware matrix, benchmark method, redundancy design, price, or delivery region.
Why the pre-purchase test matters
- It exposes bottlenecks across the whole pipeline. Token rate alone can hide data-loader stalls, small-file metadata pressure, index-build time, checkpoint bursts, and network oversubscription.
- It makes 69 TB a starting point, not a capacity promise. Calculate raw source data, transformed copies, vector indexes, model artifacts, checkpoints, snapshots, replicas, and required free-space headroom.
- It can validate the support boundary. Freeze the tested GPU driver, CUDA, container platform, NIC, switch, firmware, ONTAP, and application versions as one reproducible matrix.
- It can turn sovereignty into evidence. Record the location and administrators of primary data, replicas, telemetry, keys, support bundles, and every external model or update endpoint.
A useful acceptance report
- Publish the workload shape: dataset, file count and sizes, concurrency, model, precision, batch size, and run duration.
- Report GPU utilization and data-loader wait beside storage throughput, IOPS, latency percentiles, and CPU/network saturation.
- Test degraded operation and recovery: controller, link, switch, node, and site failure where applicable.
- Time a restore and a clean rebuild from protected data; do not infer recoverability from snapshots alone.
- State which result applies only to the tested four-GPU configuration and what must be revalidated when it scales.
Bottom line: CANCOM has published a tangible, test-before-purchase FlexPod AI offer. Its value is the integrated validation path; buyers should resist treating the component list as proof that every RAG, vision, or training workload will meet its target.