- Storage, security, networking and culture all require rethinking as enterprises transition AI from training-centric to globally distributed inference, according to the newest Fierce Network Research report.
- Data architecture is the critical first step to optimizing performance, particularly precious GPU time
- Agentic AI generates 100,000 requests per hour versus one for a human, drastically increasing demand for every infrastructure layer
Twenty years of enterprise cloud investment isn't a liability in the age of AI. It's the foundation AI runs on — if you know how to extend it.
That's the central finding of a new Fierce Network Research report, "AI doesn't replace. It extends: A practical guide to retooling enterprise infrastructure and teams for the AI age." The independent report draws on interviews with technology leaders at Eli Lilly, Nasdaq, Nature Fresh Farms, the National Hockey League and Sheetz, profiling how each is retrofitting existing stacks for agentic AI workloads.
Agentic AI changes the demand curve — and breaks existing infrastructure assumptions
The fundamental driver is demand. Traditional enterprise applications were built for human users who initiate a request, pause, consider the response and initiate another. Agents eliminate that pause. A human might generate one request per hour; an agent generates 100,000, according to Vultr CMO Kevin Cochrane. Every layer of the stack has to be sized for that machine-scale tempo.
That transition is already underway in enterprise deployments. "AI agents are going to be the dominant force, with a very small subset being human interaction," said Michael O'Rourke, Nasdaq SVP and head of AI and emerging technology.
Training and inference also impose different architectural requirements. Training favors centralized GPU clusters. Inference — which will account for an estimated 80-85% of AI workloads within two years — needs to be globally distributed, close to the customers, employees and data sources agents serve. The virtual private clouds (VPCs), Network Address Translation (NAT) gateways and load balancers already in enterprise IT become load-bearing elements of inference infrastructure, not peripheral plumbing.
Data architecture is the first job — before you buy a single GPU
Solving the data problem is the biggest challenge for AI readiness, and it has to come before the first GPU is provisioned. When Nasdaq began its AI-driven fraud detection project, the first three to four months were almost entirely data architecture work. Siloed data starves AI before it can deliver anything useful.
The investment paid off. Nasdaq cut false positives on pump-and-dump schemes by 60% in Saudi trading.
Why CPUs and storage matter more than ever in an AI infrastructure stack
Infrastructure conversations focus on GPUs, but the CPU orchestrates and saturates the GPU, and a poorly tuned CPU creates an expensive bottleneck. Eli Lilly treats every AI cluster as a paired GPU-and-CPU design.
And storage is an essential AI requirement. The NHL consolidated 16 petabytes off a 4,500-slot tape archive and is now running vision-language model pipelines against a century of footage. Nature Fresh Farms moved to all-flash storage to handle AI across its greenhouse operations. Eli Lilly is raising the bar on both speed and volume enterprise-wide.
Security is AI infrastructure, not a compliance afterthought — and culture is harder than the tech
Security has to be treated as infrastructure from day zero, not a bolt-on. An agent operating 50,000 times a week generates 50,000 instances of authentication and data access — each a potential attack surface. Nasdaq treats cybersecurity, auditability and compliance as shared services that every AI capability inherits automatically, letting the company deploy AI into regulated financial crime workflows without rebuilding security for each use case.
Cultural change proved at least as demanding as the technology work. The fastest path to AI readiness, the report finds, is reskilling engineers who already understand the existing stack — not building a parallel AI organization alongside it. Storage, network and platform engineers already know most of what AI infrastructure demands. The gap is narrower than it looks.
To learn more about how enterprises are extending their existing infrastructure to leverage AI, download our free report: "AI doesn't replace. It extends: A practical guide to retooling enterprise infrastructure and teams for the AI age."