Enterprise adoption of artificial intelligence is creating new requirements for enterprise AI infrastructure. While much of the early AI investment centered on access to powerful GPUs, organizations moving AI applications and agents into production are discovering that accelerators are only one part of the equation.
Agentic AI systems need environments where they can execute code, access data, connect to enterprise applications and operate securely across distributed infrastructure. That requires more than bare-metal GPU capacity. It calls for a complete cloud platform that combines compute, networking, storage, security and the services needed to manage applications at scale.
That shift is creating an opportunity for neocloud providers, but it is also changing what enterprises expect from them. Infrastructure providers increasingly need to support the full application lifecycle rather than simply supply specialized hardware.
During a live interview at KubeCon + CloudNativeCon North America 2026, theCUBE Research’s Paul Nashawaty spoke with Kevin Cochrane, chief marketing officer at Vultr, about the changing requirements of enterprise AI infrastructure, the return of CPU-based computing, data sovereignty and the growing role of platform engineering.
“They need a full cloud platform,” Cochrane said. “They need cloud compute, specialized AI compute, networking, storage and an open ecosystem.”
Enterprise AI needs more than GPU infrastructure
Vultr entered the cloud market in 2014 with a developer-focused platform built around performance, cost efficiency and accessibility. Since then, the company has expanded into AI infrastructure, offering both traditional cloud services and specialized GPU resources.
Cochrane said Vultr took a deliberate approach to the AI infrastructure market, prioritizing enterprise workload and platform-engineering requirements rather than focusing exclusively on GPU availability.
“We’ve always said that we’re the tortoise in AI infrastructure,” he said.
That approach reflects a broader change in how enterprises are approaching AI. Early infrastructure demand largely came from hyperscalers, AI research organizations and startups building large-scale models. Those customers needed access to GPU clusters, often through bare-metal infrastructure.
The next phase is different. Enterprises are looking to use AI agents to streamline operations, improve customer experiences and create new revenue opportunities. Those applications must work alongside existing enterprise systems, comply with security requirements and integrate with established network and application architectures.
Bare-metal GPU infrastructure alone does not address those needs. Cochrane identified complexity and a lack of specialized skills as major obstacles to enterprise AI adoption. Organizations may have access to powerful accelerators, but they still need the surrounding infrastructure and services to build, deploy and operate applications.
Vultr’s strategy is to provide a more integrated environment, including traditional cloud compute, GPU infrastructure, networking, storage and access to cloud-native projects. Cochrane said the platform supports more than 60 Cloud Native Computing Foundation projects that customers can deploy directly.
The goal is to reduce the work required to assemble an AI environment. Rather than making enterprises integrate every component themselves, providers can offer preconfigured services and composable infrastructure that developers can deploy as needed.
That distinction is becoming important as AI moves beyond experimentation. Enterprises need infrastructure that supports production applications, not just model training and inference.
Agentic AI is bringing the CPU back into the conversation
The growing focus on GPUs has sometimes obscured the role of traditional computing resources in AI architectures. But agentic AI is expanding the range of workloads that infrastructure providers must support.
An AI agent does more than send a prompt to a model and receive a response. It may execute scripts, interact with repositories, run builds, access databases and invoke external tools. Those activities require compute resources beyond the GPU.
“Not all workloads need to be on a GPU,” Cochrane said.
CPUs remain essential for databases, application code and data pipelines. They can also support smaller, optimized open-source models and inference workloads that do not require dedicated accelerators.
Another emerging requirement is the environment in which agents execute tasks. As agents gain the ability to run code and interact with systems, enterprises need secure sandboxes that isolate those activities from production environments and sensitive resources.
Cochrane pointed to work underway in the cloud-native ecosystem around safe and secure agent sandboxes as an indication of how infrastructure requirements are changing.
This creates a more complex infrastructure model. GPUs may handle demanding model reasoning and inference, while CPUs support the applications, pipelines and execution environments surrounding those models.
For organizations, the implication is that enterprise AI infrastructure decisions should account for the entire agentic workflow. Buying GPU capacity without considering the compute, storage, networking and isolation requirements around it can create another infrastructure silo. A full-stack approach gives platform teams more flexibility to match resources to individual workloads while maintaining a consistent operational environment.
Data sovereignty is becoming an infrastructure architecture decision
As enterprises expand AI deployments across regions, data sovereignty is becoming more than a question of where information is stored.
Organizations operating in regulated industries or under national data requirements need to understand where data resides, which services can access it and what happens if infrastructure outside a particular region becomes unavailable.
Cochrane described Vultr’s regional architecture as an isolation model in which each of its 34 global data center regions operates independently. According to him, data within a region remains resident there and is not accessed by Vultr services outside that region.
The distinction extends beyond data residency to application operations. Cochrane used France as an example. An application running in a French region should continue operating even if Vultr’s global control plane and its other regions become unavailable.
“If the rest of the world went dark, your production application in France? It just runs,” he said.
That approach is intended to reduce dependencies on global services that could otherwise affect regional application availability. It also gives organizations a clearer boundary for managing sensitive workloads and maintaining operational continuity.
For enterprises, this is an important architectural consideration. A cloud provider may offer regional data storage, but that does not necessarily mean every service required to operate an application is isolated within the same geography.
As AI applications become more deeply integrated into business operations, organizations will need to evaluate sovereignty across the full application stack, including data, compute, management services and operational dependencies.
AI economics are shifting toward infrastructure ownership
The cost of running AI applications is becoming another major consideration as organizations move from pilots to sustained production workloads.
Inference costs can increase quickly when agents repeatedly call models, retrieve information and execute multi-step workflows. As usage grows, enterprises must balance model performance, throughput and infrastructure expense.
Cochrane believes some organizations will increasingly turn to open-source models that can be adapted to their own data and operated on infrastructure they control. Rather than relying entirely on large commercial models for every task, enterprises can post-train specialized models using proprietary information and deploy them for operational workloads where that approach makes sense.
He described an example involving a platform engineering organization that reduced inference costs by 80% by post-training an open-source model and running it on its own infrastructure. The example illustrates a potential economic advantage, although results will depend on the workload, model, infrastructure and operational requirements involved.
The broader point is that organizations are beginning to evaluate the economics of enterprise AI infrastructure across the entire stack. Model selection matters, but so do infrastructure utilization, throughput, data movement and the cost of operating agents over time.
For some workloads, owning more of the infrastructure and model lifecycle may provide better cost control than depending entirely on external inference services.
That does not mean every enterprise should build and operate its own models. It does suggest that organizations need more flexibility to choose where models run and how workloads are distributed.
Platform engineering will shape the next cloud-native stack
The changing infrastructure requirements of agentic AI are also expanding the responsibilities of platform engineering teams.
Cochrane expects platform engineers to become central to enterprise AI adoption, working more closely with business units to understand which agents and preconfigured application environments are needed across the organization.
That moves platform engineering beyond infrastructure provisioning and developer enablement. Teams will increasingly need to help define the building blocks that allow developers to deploy governed AI applications without having to assemble every component independently.
Cochrane also sees an opportunity for the cloud-native community to evolve how it shares infrastructure patterns. The CNCF ecosystem has helped developers adopt individual open-source projects across networking, storage, observability and other infrastructure categories. But as architectures become more complex, organizations need guidance on how those components fit together.
Cochrane predicted the emergence of marketplaces where developers can share opinionated, prebuilt stacks for specific workload types. Instead of selecting and configuring individual components, teams could start with a tested architecture and adapt it to their requirements. That could help reduce the complexity of building agentic applications while encouraging more consistent deployment practices.
He also expects consolidation in the neocloud market. Providers focused primarily on bare-metal GPU capacity may face pressure as enterprises demand broader cloud capabilities. Those able to expand into full public cloud platforms could be better positioned to serve production workloads.
Vultr’s existing cloud platform gives it a foundation for that transition, Cochrane said, with its investments in both traditional compute and specialized AI infrastructure.
The bottom line
Enterprise AI is changing the definition of cloud infrastructure. GPUs remain important, but agentic applications require a broader foundation that includes CPUs, secure execution environments, networking, storage, data sovereignty and integrated cloud-native services.
That creates an opportunity for providers that can deliver more than specialized accelerator capacity. It also puts platform engineering teams in a more strategic position as enterprises work to standardize how AI applications are built, deployed and governed.
The next phase of the neocloud market may therefore be determined less by who can offer the most GPU capacity and more by who can provide the complete environment enterprises need to run AI in production.
As Cochrane’s comments suggest, the infrastructure challenge is no longer simply making AI compute available. It is making the entire system around AI practical to operate at enterprise scale.
A message from John Furrier, co-founder of SiliconANGLE:
Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities.
- 15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more
- 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network
Are you an AWS customer? Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/
Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations.