Enterprises are under pressure to turn AI infrastructure investment into measurable value. Models and compute continue to advance, but the performance and economics of AI also depend on how compute, storage, data and users are connected.
That pressure is visible in our The AI Impact Imperatives, 2026, research, which found that the median expected payback period for major AI investments is roughly 18 months. In The Value Edge: Powering the Next Era of TMT, research we carried out with Economist Enterprise, only one-third of 202 C-suite executives surveyed had a framework to measure AI’s commercial impact.
At AI Infra Summit 2026 in Santa Clara, Clayton Wagar, Leader, AI and High Performance Networking at Nokia, explored how foundational network architecture can support AI performance, flexibility and ROI as enterprises scale.
The network is part of the AI system
AI workloads bring together compute, memory, storage and data that may sit within a rack, across a data center or across multiple locations. Moving information between those resources makes the network a core part of how effectively the overall system performs.
“The network itself is what provides that leverage, which gives us the ability to form the technical connections between all of these pieces,” said Wagar.
As AI becomes more distributed, those connections extend beyond individual data centers. Enterprises may need to access infrastructure across private environments, cloud services and geographically distributed facilities, while inference increasingly happens closer to users and devices.
For Wagar, network architecture and orchestration allow those resources to work together as a broader system.
Our work with Nokia in data center networking reflects that focus. Together, we combine next-generation leaf-spine fabrics with intelligent automation and engineering excellence to help enterprises scale infrastructure faster, simplify operations and reduce total cost of ownership across the network lifecycle.
Match network design to the workload
Capacity planning becomes more difficult when AI demand and the underlying technology are changing quickly. Enterprises have to make infrastructure decisions that can support future growth without paying for capability their workloads may not need.
Wagar argued that organizations can make those decisions by starting with the requirements of the workload rather than assuming every deployment needs the newest available technology.
“You don’t always have to buy the latest gadget,” he said. “You can get a long way with the technologies we have today.”
Standardized architectures and mature technologies can offer a practical path for enterprises building or expanding AI infrastructure. Availability, operational expertise, procurement leverage and unit economics can all influence the right design.
Distribution changes the economics
AI infrastructure economics can also vary by location. Power costs, available capacity and access to compute differ across regions and can change over time.
For enterprises operating across multiple sites, connectivity affects where workloads can run and how capacity can be shared. Wagar pointed to shifting workloads between locations as one way network architecture can give organizations greater flexibility as they manage infrastructure economics.
That flexibility depends on interconnection between data centers, clouds and other infrastructure domains. Networks are the connective layer within data centers, between facilities and out to users and devices, allowing distributed resources to operate as part of a broader AI infrastructure.
For enterprises, that creates more options for how infrastructure is deployed and used as demand evolves.
Automation becomes a scaling requirement
Operating these environments introduces another challenge. As AI infrastructure grows, the number of devices, connections and changes that network teams need to manage also increases.
Automation can help teams respond faster and operate at greater scale, but Wagar emphasized that speed also raises the consequences of a bad decision.
“Automation makes things go faster. So, you can amplify mistakes,” he said.
He described a gradual path toward more autonomous network operations. AI can initially assist people with documentation, context and operational tasks before taking on greater responsibility for managing the environment.
“I don’t think we ever fully remove humans because we’re the ones that can discern what the right thing to do for our business is,” said Wagar.
That human oversight becomes even more important as automation takes on a greater role in network operations.
The role of automation is increasingly tied to scale and performance. Wagar noted that network automation conversations several years ago often centered on cost savings. AI infrastructure is expanding the emphasis toward keeping larger environments healthy, responsive and manageable.
Infrastructure choices shape AI ROI
AI ROI depends in part on how efficiently the underlying resources can be put to work. Network architecture affects how compute, storage and data interact, where workloads can run and how effectively infrastructure can be operated as it scales.
That makes network design part of the investment decision. Enterprises need enough capacity and flexibility for the workloads they expect, alongside an operating model that can adapt as those requirements change.
“We can’t just continue to throw money at it. We have to be smart about how we manage these things,” said Wagar.
Foundational choices around network architecture, orchestration and automation will shape how quickly enterprises can scale AI, how efficiently they can use infrastructure and whether growing capital investment can translate into measurable business returns.
