Cisco and Canonical have published a Cisco Validated Design (CVD), a tested guide for building AI infrastructure at edge locations without hand-assembling the whole stack. The premise is straightforward: pair the Cisco Unified Edge hardware with Ubuntu software and hand teams a recipe that already ships proven, so nobody has to fight dependencies, versions and certifications on their own.
What’s in the software base
The foundation is Ubuntu Server 24.04.3 LTS, with Ubuntu Pro on top covering up to 15 years of security maintenance and the backporting of critical fixes. That’s what separates a serious edge deployment from an experiment. Machines scattered across hundreds of stores, factories or branches don’t get patched every week, so they need a long support window and fixes that don’t force a major-version jump.
On that base, the CVD stacks the rest:
- Canonical Kubernetes for cloud-native orchestration.
- LXD for system containers and lightweight VMs.
- MicroCloud, which combines LXD, MicroCeph and MicroOVN into a small, self-contained private cloud.
- The Data Science Stack (DSS) for setting up the model-development environment.
- Charmed Operators that automate the operation of Kubeflow and MLflow, the two MLOps pieces that usually cost the most to keep current.
The Cisco hardware
The chassis is a Cisco UCS XE9305, three rack units and short depth, built for sites with no data centre behind them. It takes up to five Cisco UCS XE130c nodes, each with a 6th Gen Intel Xeon SoC of up to 32 cores and up to 768GB of DDR5. For inference, the design uses NVIDIA L4 Tensor Core GPUs, aimed at energy efficiency rather than training large models.
Management runs through Cisco Intersight, Cisco’s cloud platform, and provisioning is Zero-Touch: from firmware up to the Kubernetes layer, everything deploys via curated blueprints, with no one driving out to configure each node by hand.
Why it matters
The post names five bottlenecks that hold AI back outside the data centre: missing GPUs and dense compute, environments that are too rigid, poor scaling across thousands of dispersed sites, the CapEx and OpEx cost, and software fragmentation (versions that fall behind, unpatched CVEs, single-vendor lock-in). The CVD tackles all of it with a validated combination instead of leaving you to assemble the puzzle.
If you run AI spread across many locations, the value here isn’t any single component but the fact that someone has already done the work of confirming they fit together. That cuts the time between “we have hardware” and “we have models serving in production,” which is where most of the effort tends to disappear. And by leaning on Ubuntu Pro, the security cycle stays with those machines for years, which is exactly what you need when you can’t touch them often.
Source
Original article by Canonical, published 11 June 2026: AI at the edge: simplifying infrastructure with Cisco and Canonical.