What Tech Leaders from Zayo and Cisco Want Every Enterprise to Know:

When most executives think about their AI strategy, they picture models, prompts, and dashboards. What they rarely picture is a fiber conduit running underground, a liquid-cooled rack drawing 60 kilowatts of power, or a supply chain delay holding up their entire data center buildout.

That mismatch — between where leaders are investing their attention and where the actual constraints live — was the central tension in a recent panel that brought together two of the most experienced infrastructure minds in enterprise technology. Troy Lupe, who oversees network strategy for one of North America’s largest fiber providers, and Jason Andrews, who manages engineering operations across Cisco’s global footprint, didn’t hold back. What emerged was a frank, vendor-neutral conversation about what it actually takes to build an enterprise that won’t be outpaced by the demands of the next three to five years.

1. The Problem with Calling AI a “Software Problem”

There’s a statistic worth sitting with: more than 50% of Americans have now used a large language model, and over 20% use tools like ChatGPT on a daily basis. Inside the enterprise, that number climbs even higher — roughly 60% of employees are incorporating AI into their day-to-day work in some form.

That kind of adoption curve would be remarkable for any technology category. But it creates a challenge that most AI roadmaps aren’t prepared to address: at this scale, AI stops being a software challenge and becomes a physical infrastructure challenge.

“The conversation has moved. The pressure points are no longer about which model to fine-tune. They’re about power availability, fiber density, and cooling capacity.”
— Troy Lupe, EVP & Chief Network Officer, Zayo Group

As AI workloads grow, the underlying constraints are increasingly physical — and physical infrastructure has timelines, lead times, and costs that behave very differently from software. Data centers are already moving to non-traditional locations — former industrial sites, repurposed crypto-mining facilities — in search of cheap power and existing conduit. The race isn’t for GPU access anymore. It’s for megawatts and fiber. 

2. The Network Has Fundamentally Changed — Have Your Enterprise Designs?

Traditional enterprise network design was built around North-South traffic flows: data moving from users to servers, from branch offices to the core, from endpoints to the cloud. AI workloads break that assumption entirely.

Training and inference at scale depend on massive GPU clusters communicating simultaneously — East-West traffic, node-to-node, running at volumes that legacy architectures weren’t designed to carry. The industry is sprinting from 800 gigabits per second toward 1.6 terabits and 3.2 terabits within the next three to four years. That’s not incremental growth. That’s a categorical change in what the network needs to do.

Resilience is evolving just as rapidly. Where “dual diversity” was once the enterprise gold standard for redundancy, AI-scale operations increasingly require “traversity” and “quadversity” — three and four independent routing paths, purpose-built to ensure no single failure can interrupt workloads running continuous inference across hundreds of nodes.

“For most enterprise network teams, this isn’t a configuration change. It’s a design-level rethink.”
— Troy Lupe, EVP & Chief Network Officer, Zayo Group

3. Power and Cooling: The Numbers That Should Alarm You

Legacy rack deployments were designed around power densities of 1.5 to 2 kilowatts per rack. Today’s GPU-dense configurations routinely demand 30 to 40 times that figure — a rack that once drew 2 kilowatts is now expected to draw 60 to 80.

Air cooling cannot keep up. Liquid cooling, once reserved for the most specialized HPC environments, is now a practical requirement for any enterprise serious about hosting AI workloads. Retrofitting cooling infrastructure is expensive, time-consuming, and in many cases requires rebuilding from the floor up.

Supply chain adds another layer of friction. Both panelists flagged six-month lead times for high-power connectors as an active constraint on buildout schedules. For teams accustomed to hardware arriving within weeks, this is a significant planning adjustment — one that rewards organizations who start procurement well ahead of when they expect to need the capacity.

4. The Data Hygiene Problem Nobody Talks About at the All-Hands 

It’s tempting to treat data quality as a backend problem — something IT resolves quietly before the business ever sees it. In AI environments, that approach fails loudly and publicly. 

Without clean, well-governed data flowing into AI systems, the outputs are unreliable. Hallucinations aren’t just a model problem — they’re frequently a data problem. Fragmented CMDBs, inconsistent inventory records, and unresolved duplicates don’t just create minor inaccuracies; they undermine the confidence teams need to act on AI-generated recommendations. 

The payoff for getting it right is substantial. Real-time monitoring, automated fault isolation, and self-healing systems — when built on clean operational data — can compress root-cause analysis from multi-hour war room sessions into automated remediation workflows. The ROI materializes quickly, and it compounds.

“Software agents that take actions need to be logged, governed, and managed with the same discipline as any other infrastructure component.”
— Jason Andrews, VP Engineering Operations, Cisco

As AI agents become more deeply embedded in operational workflows, they can no longer be treated informally. Agents that provision, reroute, or respond to incidents need service catalogs, audit trails, and access policies — accounting for automated actors alongside human ones. 

5. Build vs. Partner: The Question Every CTO Is Getting Wrong

Build only when it creates genuine, defensible competitive differentiation. Everywhere else, the math favors partnership.

This isn’t about cost alone, though the cost argument is real. Proprietary infrastructure requires proprietary expertise to maintain. It creates integration overhead at every system boundary. It absorbs engineering cycles that could go toward differentiated product development. And when the underlying technology shifts — as it has repeatedly in AI — internally built systems carry technical debt that vendor-supported solutions don’t.

The flip side: don’t over-rotate on partnerships either. Vendor bloat — too many providers, too many platforms, too many data silos — creates its own form of integration tax. The goal is a tight, validated partner ecosystem built around reference architectures that have actually been stress-tested, not a sprawling portfolio of licenses that no one has time to operationalize.

6. What’s Coming: Two Predictions Worth Writing Down

TROY LUPE — ZAYO GROUP
Power availability and fiber capacity will become the defining constraints on AI infrastructure growth in North America. As the construction boom intensifies, the sheer volume of underground work will drive a significant increase in physical fiber cuts — one of the most disruptive and hardest-to-predict failure modes in any network. Enterprises that have built multi-path resilience will weather this period very differently than those still operating on legacy diversity assumptions.

JASON ANDREWS — CISCO
The threat window for cybersecurity vulnerabilities is collapsing. The time between a CVE publication and active exploitation — once measured in weeks — is compressing toward seconds and minutes as adversarial actors use automation to accelerate attack deployment. Human-speed patch management is no longer sufficient. Enterprises need automated patch workflows and must begin preparing now for post-quantum cryptography, which will require foundational changes to encryption infrastructure before quantum computing capabilities mature.

The Strategic Takeaway

For years, the differentiators in technology strategy lived at the application layer — which software you chose, which vendor you standardized on, which platform you ran your cloud workloads on. AI is pushing the determinants of competitive advantage further down the stack: into network architecture, power infrastructure, data governance, and the operational discipline to run autonomous systems at scale.

The enterprises that recognize this shift early — and invest accordingly — won’t just be better positioned for the AI applications of 2027. They’ll have built a foundation that makes every subsequent technology transition faster and less disruptive than it would otherwise be.

That’s what future-ready actually means.

Catch the complete conversation — including live audience Q&A — on YouTube: Engineering Tomorrow: How Tech Leaders Are Building Future-Ready Enterprises