Your AI Strategy Is Waiting on Provisioning: How Cisco Stack Automation Closes the Gap

The problem nobody talks about in AI strategy meetings

Every boardroom conversation about AI eventually arrives at the same uncomfortable question: why does it take so long to stand up the infrastructure? Organizations that can approve an AI initiative in a single meeting often wait six to twelve weeks for the underlying platform to become usable. That gap between decision and capability is where momentum dies, budgets erode and competitors move first.

The delay is not a people problem. It is a complexity problem. A modern AI or data platform deployment touches somewhere between 25 and 40 tool configurations, generates 300 to 700 discrete infrastructure actions and requires 40 to 100 handoffs and approvals across compute, network, storage, virtualization and platform teams. Every handoff is a queue. Every manual step is an opportunity for drift, misconfiguration or rework. GPU driver versions must align with acceleration libraries and toolkits, which must align with Kubernetes versions, which must align with container runtimes, which must align with the fabric configuration underneath it all. Get one layer wrong, and the failure often does not surface until validation, weeks into the project.

Public cloud solved this problem years ago by hiding the assembly behind a service catalog. On-premises infrastructure never got that treatment, even as the case for running AI workloads on-prem grew stronger on data gravity, sovereignty and economics. That is the gap Cisco is now closing.

What Stack Automation actually is

At Cisco Live in June 2026, Cisco introduced Stack Automation by Quali, a deployment automation platform co-engineered with Quali and offered exclusively through Cisco. The premise is direct: bring a cloud-like consumption experience to Day 0 and Day 1 infrastructure and software deployment on-premises, compressing cycles that took weeks into hours.

Quali is a name worth knowing, even if it has not crossed your desk before. The company has spent years powering the Cisco DevNet Sandbox, the environment that provisions complex, real-world lab topologies on demand for hundreds of thousands of developers and network engineers. That is not a proof-of-concept relationship. It is a production automation engine that has been battle-tested at scale, now productized and pointed at enterprise data centers.

The platform presents a Solutions Hub, essentially a curated catalog of deployable blueprints spanning Cisco and third-party software. An administrator or platform engineer selects a solution, provides a handful of inputs and the platform orchestrates everything beneath the surface: bare-metal OS installation, hypervisor deployment, Kubernetes clusters, GPU enablement, application installation and validation. Day 2 operations then hand off cleanly to the element managers enterprises already run, including Cisco Intersight and Nexus Dashboard, so nothing about existing operational tooling gets disrupted.

Two deployment scopes matter here. Software-only automation provisions and configures individual layers, whether that is Nexus Dashboard, Splunk, Red Hat OpenShift, VMware ESXi, Qumulo or dozens of others, on top of existing hardware or cloud. Full-stack solutions go further, orchestrating entire pre-validated architectures such as Cisco AI PODs, including networking, compute, storage and the software above them, in a single unified workflow. Software-only capability is targeted for general availability in August 2026, with full-stack deployments including AI PODs following in October 2026.

Why this is a strategy conversation, not a tooling conversation

It is tempting to file this under IT operations and move on. That would be a mistake, for three reasons.

First, deployment speed is now a business metric. When infrastructure provisioning takes ten weeks, every downstream initiative inherits that latency. Data science teams wait. Application teams wait. The business case that justified the investment ages before the first workload runs. Compressing Day 0 and Day 1 from weeks to hours changes the tempo at which the entire organization can experiment, fail and iterate. That is the same argument that drove cloud adoption a decade ago, now applied to the on-prem estate where an increasing share of AI workloads will actually live.

Second, embedded expertise reduces operational risk in a way that headcount cannot. Stack Automation bakes Cisco validated configurations and security baselines directly into every deployment. The security settings, policy enforcement and best practices that normally depend on your most senior engineer being in the room are applied consistently and automatically from the moment of provisioning. In an environment where skilled infrastructure engineers are scarce and expensive, encoding that expertise into the platform is a genuine risk transfer.

                                                                                  -Jim Stucky - Director, GS&A Automation Practice
 
Third, this reframes what the infrastructure team sells internally. The platform's two routes to market, consumer and producer, mirror how mature IT organizations already think. Consumers select curated solutions from the catalog and deploy them self-service. Producers, typically platform engineering teams, build and customize their own blueprints, integrating existing Terraform, Ansible, Helm and Python automation into governed, reusable services published to an internal catalog. Your infrastructure organization stops being a ticket queue and starts operating as an internal service provider with a product catalog.

The licensing model deserves attention

Cisco has structured this in a way that removes the usual excuse for not trying it.

The Essentials tier is free and includes 100 concurrent compute units, providing access to the solution catalog for standard deployments. Support is not included at this tier but can be added at cost. For a meaningful number of use cases, OS provisioning, VM deployment, LLM hosting, GPU-as-a-service, the free tier is genuinely usable, not a crippled trial.

The Advantage subscription is billed per concurrent compute unit, or CCU, a unified metric measuring the scale of resources being actively automated at any given time. The conversion ratios are straightforward: a bare metal OS deployment consumes 2 CCUs and requires Intersight Advantage licensing, a concurrent VM consumes 1 CCU, a container consumes 0.25 CCU and a serverless function consumes 0.125 CCU. You pay only for capacity under active management, and the metric scales sensibly from heavyweight bare metal down to ephemeral functions. Advantage unlocks the capabilities that matter to producer teams: the Blueprint Designer, GitOps integration, advanced customization, parameters, scheduling, approval workflows through Webex, Slack, ServiceNow and email, cost show back and access to agentic design tools, including Claude Cowork.

Full-stack services include a one-time service fee for the automated deployment of curated solutions, such as Cisco AI PODs, which will be available once full-stack support reaches general availability.

One clarification that matters for heterogeneous environments: only bare metal OS deployment is exclusive to Cisco UCS servers. Everything else, including VM services, LLM services and Cisco and third-party applications, runs on any underlying server that meets the requirements of the hypervisor or container platform. This is not a lock-in play disguised as automation.

The agentic angle is the most forward-looking piece

The most interesting architectural decision in Stack Automation is how deeply it embraces AI-assisted operations and how honestly it distinguishes between the two approaches.

An embedded Copilot lives inside the Blueprint Designer for straightforward work: creating basic custom blueprints from existing building blocks, wiring inputs and outputs and getting a working deployment definition quickly. It is the right tool when inputs are static, agentic integrations are not required and the blueprint can live locally.

For serious platform engineering, the Claude Cowork integration operates at a different level. Through an MCP server and purpose-built skills, teams use Claude to author complex blueprints that chain multiple automation frameworks, Terraform into Ansible into Python, with dynamic inputs, pre-flight validation checks, conditional logic and full repository management with commit control over blueprint YAML in Git. The workflow is conversational: describe the automation you need, let the agent scan existing blueprints as reference patterns, generate the YAML, validate it and push it to the repository. When a test run breaks, feed the error logs back into the conversation and iterate. Additional agentic design tools extend the same pattern as the ecosystem matures.

For executives, the significance is not the novelty. It is that infrastructure-as-code authorship, historically gated behind a small population of automation specialists, that becomes accessible to a much wider set of engineers. The scarce skill shifts from writing the automation to reviewing and governing it, and the platform provides the governance scaffolding: dedicated dev and production spaces, QA workflows before catalog publication, versioned lifecycle management for automation assets and control over exactly what gets published to the internal catalog.

What this looks like in practice

Consider four scenarios that map directly to what the platform supports today or at the time of general availability.

An enterprise AI factory. A financial services firm greenlights an internal AI initiative built on a Cisco AI POD: UCS compute, dense GPU acceleration, high-performance storage and a Kubernetes platform running a validated enterprise AI software stack. Historically, this is a multi-month program spanning four teams, with GPU driver alignment, fabric tuning and GPU licensing each representing their own mini-project. With full-stack deployment, the entire architecture is orchestrated as one validated workflow, and the conversation with the business shifts from provisioning timelines to model roadmaps.

GPU-as-a-service for data science. A healthcare organization owns a pool of training and inference class GPU capacity that data science teams contend for through tickets and spreadsheets. Using the GPUaaS blueprints, the platform team publishes a catalog item that provisions Jupyter environments with on-demand GPU allocation, along with scheduling and duration policies that automatically reclaim idle resources. Utilization becomes visible, contention becomes governable and the shadow cloud spend that data scientists quietly accumulate when internal capacity is too hard to access starts coming back on-prem.

Sovereign LLM hosting. A public sector agency needs generative AI capability but cannot send data to external APIs. The catalog includes blueprints for deploying open models on existing infrastructure, enabling the agency to have a private LLM endpoint in an afternoon rather than a quarter, with security baselines applied at provisioning time rather than bolted on afterward.

Observability at scale. An industrial company standardizing on Splunk across a dozen sites turns what is normally a site-by-site professional services engagement into a repeatable catalog deployment, with each instance configured identically and validated on arrival. The same pattern applies across the broader supported portfolio, which at GA spans Cisco APIC, Catalyst Center, Nexus Dashboard, ISE, ThousandEyes, Secure Firewall and a continuously refreshed third party list including OpenShift, Veeam, Qumulo, NetBox, Dataiku and more.

The honest caveats

A thought leadership piece that reads like a brochure serves nobody, so three cautions are in order.

Timelines are explicitly subject to change. Software-only GA is targeted for August 2026 and full-stack for October 2026, and Cisco's own materials note that features ship on a when-and-if-available basis. Organizations should plan pilots against the free tier now and structure full-stack commitments with appropriate flexibility.

Automation platforms are only as good as the operating model around them. A catalog nobody governs becomes sprawl with better tooling. The organizations that will extract real value are the ones that treat this as a platform engineering practice, with clear ownership of blueprint standards, publication controls and lifecycle discipline, not as a faster way to do the same ungoverned work.

And CCU-based consumption pricing requires the same financial hygiene as cloud spend. The show back capabilities are built in. Use them from day one, because concurrent consumption models reward organizations that actively manage what is running and quietly penalize those that do not.

Where to start

The pragmatic entry point is low risk by design. The free Essentials tier with its 100 CCUs is enough to run a meaningful pilot: pick two or three high-friction, high-frequency deployment patterns, put them through the catalog and measure elapsed time and touch labor against your current baseline. If the numbers hold, the Advantage conversation and the full-stack roadmap discussion become straightforward.

The harder and more valuable work is validating how this fits your environment before it touches production, and that is where WWT is uniquely positioned to help. The Advanced Technology Center gives organizations hands-on access to the Cisco compute, GPU and storage platforms that Stack Automation orchestrates, in an environment built for exactly this kind of evaluation. And for organizations working through the broader AI infrastructure question, the AI Proving Ground provides a purpose-built environment to test full-stack AI architectures, including AI PODs and the automation that deploys them, against real workloads before committing capital. If deployment velocity is the bottleneck between your AI strategy and your AI results, that is a conversation worth having now, while the platform is in beta and the learning curve is cheapest.