Formerly known as Wikibon

Enterprise AI Requires a Full-Stack Cloud Infrastructure Strategy

Enterprise AI is entering a new phase. The initial wave of investment focused heavily on GPU availability, model training, and the infrastructure required to support large-scale AI experimentation. But as organizations move toward deploying agentic AI systems that execute workflows, interact with enterprise applications, and perform operational tasks, the infrastructure requirements are changing.

Running a model is no longer enough. Enterprises need the surrounding cloud infrastructure to support the entire application lifecycle, including compute, networking, storage, security, and the execution environments where AI agents operate. They also need the ability to deploy those workloads across distributed environments without sacrificing performance, cost efficiency, or data sovereignty.

During a live interview at KubeCon + CloudNativeCon North America 2026, I spoke with Kevin Cochrane, CMO of Vultr, about the changing requirements of enterprise AI infrastructure and why organizations need to rethink how they build, deploy, and operate agentic systems. Our conversation explored the limitations of standalone GPU infrastructure, the importance of balancing CPU and GPU resources, the economics of running customized open-source models, and why platform engineering teams will become central to mainstream enterprise AI adoption.

Enterprise AI Is Outgrowing Standalone GPU Infrastructure

One of the most important changes happening across the AI infrastructure market is the transition from infrastructure built primarily for model training to infrastructure designed for production enterprise workloads. In the early stages of the market, organizations could deploy bare-metal GPU clusters and hand them over to AI labs, hyperscalers, or well-funded startups. Current enterprise requirements are different. Organizations deploying AI into business operations need more than GPUs. They need integrated infrastructure that supports application development, data pipelines, networking, storage, security, and operational management.

According to Cochrane, this is where the distinction between specialized AI infrastructure and a full cloud platform becomes increasingly important. “We’ve always said that we’re the tortoise in AI infrastructure,” Cochrane said, describing Vultr’s deliberate approach to building for enterprise workloads rather than pursuing GPU capacity alone. Enterprises deploying agents to streamline operations, improve customer experiences, and create new revenue opportunities need infrastructure capable of supporting the entire application environment. “They need a full cloud platform,” Cochrane explained. “They need integrated cloud compute with specialized AI compute and AI services for networking and storage.”

A GPU cluster may provide the compute required to train or run a model, but it does not automatically provide the enterprise services necessary to operationalize that model. Organizations still need to integrate security controls, network architecture, application services, and cloud-native tooling. Cochrane identified three persistent barriers to enterprise adoption: infrastructure complexity, shortages of specialized skills, and the inability of standalone neocloud infrastructure to fit existing enterprise network, security, and application architectures. Vultr’s approach is to combine specialized AI compute with the broader cloud services enterprises need, reducing the number of separate infrastructure components organizations must assemble and manage.

Agentic AI Is Bringing the CPU Back Into the Infrastructure Conversation

The growing focus on GPUs has overshadowed another important part of the AI infrastructure equation: CPUs. As enterprises move toward agentic AI, CPU resources are becoming increasingly important because agents do more than generate responses from a model. They execute code, interact with tools, access repositories, run scripts, process data, and perform tasks across application environments.

Those activities require execution infrastructure separate from the GPU resources powering model inference and reasoning. “Not all workloads need to be on a GPU,” Cochrane said. CPUs remain essential for databases, application code, and data pipelines. They are also becoming important for smaller, optimized open-source models performing specific inference tasks closer to the edge.

But the emerging requirement goes further. AI agents need secure environments in which they can execute work without creating unnecessary risk for the underlying enterprise infrastructure. Cochrane highlighted the growing importance of agent sandboxes, including new projects being incubated within the Cloud Native Computing Foundation (CNCF). These environments provide controlled execution boundaries where agents can operate, access approved resources, and perform tasks without unrestricted access to production systems.

For infrastructure providers, this creates a new category of demand. Enterprises must increasingly provision CPU resources for agent execution and application orchestration alongside GPUs for model inference. Vultr’s full-stack positioning is designed to accommodate both sides of this architecture. “At the end of the day, the ability to have one single platform that you can build and deploy and run with safety and security your entire agentic AI system” is an important enterprise proposition, Cochrane explained.

The economics of agentic AI cannot be measured solely through GPU utilization or inference throughput. Organizations must also account for the compute required to execute agent workflows and the operational overhead associated with managing multiple providers.

Data Sovereignty Is Becoming a Core AI Infrastructure Requirement

As AI moves into sensitive enterprise workflows, data sovereignty is becoming more than a regional compliance consideration. It is increasingly an architectural requirement. Organizations are beginning to use proprietary operational data to post-train and customize open-source models for specific business functions. Those models may ultimately support customer interactions, internal operations, and other workloads that depend on sensitive information. The infrastructure supporting those models must provide clear boundaries around where data resides, how it is accessed, and where inference takes place.

Cochrane emphasized that enterprises need to treat data residency and sovereignty as fundamental requirements when deploying production AI systems. “If you’re actually using highly sensitive data to tune a model and run an infer model in a region, you need to make certain that data is owned by you, controlled by you, and run and only used in that region,” he explained.

Vultr addresses this through regional infrastructure isolation. According to Cochrane, each of the company’s 34 data center regions operates independently. Data placed in a region remains resident there and is not accessed by Vultr services outside that region. Importantly, the isolation extends beyond data storage. Cochrane described an architecture in which an application deployed in Vultr’s French region can continue operating even if the global control plane and other Vultr regions become unavailable. “If the rest of the world went dark, the global control plane’s down, every other Vultr data center region goes down. What happens to your production application in France? It just runs,” he said.

That level of operational independence addresses a potential weakness in distributed cloud architectures: hidden dependencies on centralized services that can create broader failure domains. For enterprises operating across multiple jurisdictions, regional isolation can support both data sovereignty and operational resilience. It also becomes increasingly important as AI models are customized using region-specific enterprise data.

Token Economics Are Shifting Toward Infrastructure Ownership

The discussion around AI economics is also evolving. Much of the current market conversation focuses on maximizing tokens per dollar from third-party model providers. While that remains important, enterprises deploying differentiated business workloads may need to evaluate a different approach.

Cochrane argued that organizations running sensitive operational workloads should consider post-training open-source models on proprietary data and operating those models on infrastructure they control. “Enterprises for sensitive operational workloads that are actually running the business and differentiating their business, need to be running bespoke open-source models that are trained and tuned for their business operations,” Cochrane said. He referenced a platform engineering case study in which an organization reportedly reduced inference costs by 80% after taking an open-source model, post-training it, and running it on its own infrastructure. While that example is not a universal benchmark, it illustrates an important economic consideration.

The lowest-cost path for an enterprise is not necessarily the provider offering the cheapest individual inference request. Depending on workload characteristics, volume, model requirements, and operational constraints, owning more of the infrastructure stack may provide better long-term economics. This is where tokenomics becomes a broader infrastructure decision. Organizations need to evaluate the relationship between model performance, token throughput, GPU utilization, CPU requirements, data movement, and the cost of operating the entire application environment.

For some workloads, third-party model APIs will remain the most efficient option. For others, particularly those involving sensitive proprietary data or predictable high-volume inference, customized open-source models running on controlled infrastructure may provide a more attractive economic model.

Platform Engineering Will Shape the Next Cloud-Native Stack

Looking ahead, Cochrane identified three developments that he believes will shape the next phase of cloud-native infrastructure. The first is the expanding role of platform engineering teams. As business units begin identifying opportunities for AI agents, platform engineering teams will increasingly need to translate those requirements into reusable, governed infrastructure and application environments. “Platform engineering teams are going to be at the center of mainstreaming enterprise adoption of agentic AI systems,” he said. This represents an expansion of the traditional platform engineering mandate. Rather than simply providing infrastructure abstractions, platform teams will increasingly work with lines of business to identify the agents, services, and preconfigured technology stacks developers need. That includes defining approved execution environments, integrating security and governance controls, and standardizing access to models and reusable application components.

Cochrane’s second prediction is the emergence of an entirely new cloud-native stack designed around AI workloads. He pointed to the growing number of CNCF projects being developed to support the AI ecosystem. Vultr currently supports more than 60 CNCF projects, according to the company. However, an expanding ecosystem introduces another challenge: integration complexity. Cochrane described Vultr’s vision for addressing this through a marketplace of opinionated stacks, where developers and platform engineering teams can share prebuilt recipes for deploying complete systems supporting specific workload types. The concept extends the traditional open-source community model. Instead of sharing individual projects and components, developers would increasingly share complete infrastructure configurations and deployment patterns that others can reuse.

His third prediction concerns consolidation in the neocloud market. Cochrane expects enterprises to work with a smaller number of providers as their infrastructure requirements expand beyond specialized GPU capacity. “We’re going to see a consolidation and majority of enterprises working with a smaller number of neoclouds that make the transition to becoming full public cloud platforms,” he said. That reflects a broader maturation of the AI infrastructure market. Access to GPUs may attract customers, but long-term enterprise adoption requires the broader compute, networking, storage, security, and cloud-native services associated with a full cloud platform.

Analyst Take

The enterprise AI infrastructure market is moving beyond the GPU race. GPU availability remains important, but the more difficult challenge is building an integrated environment capable of supporting production AI applications. As agentic systems become more autonomous, enterprises need infrastructure that can support model inference, application execution, agent sandboxes, data pipelines, networking, storage, and governance without introducing additional operational complexity. Vultr’s positioning reflects that shift. Its combination of traditional cloud compute, AMD and NVIDIA GPU options, regional infrastructure isolation, and cloud-native ecosystem support addresses a broader set of enterprise requirements than specialized accelerator capacity alone.

Three developments stand out. First, CPUs are becoming increasingly important as agentic AI introduces new requirements for secure execution environments and application orchestration. Second, data sovereignty and token economics are becoming interconnected infrastructure decisions. Organizations customizing open-source models with proprietary data will increasingly evaluate where those models run and whether owning more of the infrastructure can improve long-term economics. Third, platform engineering is moving closer to the center of enterprise AI strategy. The ability to provide reusable, governed infrastructure stacks will become an important factor in determining how quickly organizations move agents from experimentation into production.

The larger takeaway is that enterprise AI is becoming an infrastructure integration problem. Organizations do not need more disconnected pieces of AI infrastructure. They need a coherent platform that brings compute, models, data, networking, security, and execution environments together while preserving the flexibility to deploy workloads where they make the most sense. The next phase of AI infrastructure will be defined less by who can provide the most GPUs and more by who can make the entire AI operating environment work.

Article Categories

Join our community on YouTube

Join the community that includes more than 15,000 #CubeAlumni experts, including Amazon.com CEO Andy Jassy, Dell Technologies founder and CEO Michael Dell, Intel CEO Pat Gelsinger, and many more luminaries and experts.
"Your vote of support is important to us and it helps us keep the content FREE. One click below supports our mission to provide free, deep, and relevant content. "
John Furrier
Co-Founder of theCUBE Research's parent company, SiliconANGLE Media

“TheCUBE is an important partner to the industry. You guys really are a part of our events and we really appreciate you coming and I know people appreciate the content you create as well”

Book A Briefing

Fill out the form , and our team will be in touch shortly.
Skip to content