0

Editor’s note: preceding the AI Infrastructure Forum, now available on-demand, Shawn Rosemarin, vice president of R&D – Customer Engineering at Pure Storage, spoke to RCR Tech about the difference between AI factories and general-purpose data centers — and why AI data-readiness is the critical prerequisite for both.

From general-purpose data centers to AI factories
Unlike general-purpose data centers, AI factories are AI-optimized — designed specifically for high-volume data ingestion, training, and AI inference at scale. Think of them as end-to-end production lines for AI, where data flows from raw input to trained models to real-time predictions.

Companies that offer and leverage AI factories include hyperscalers and infrastructure providers such as Dell, HPE and Pure Storage. Google Cloud, Microsoft, AWS, Nvidia, Dell, HPE, Coreweave, Siemens, IBM, and Intel. Their number is set to grow rapidly as the specialized demands of generative AI outstrip what traditional data centers can handle. But AI factories will increase in number as the specialized demands of generative AI exceed what general-purpose data centers can handle.

By 2030, data centers equipped to handle AI processing loads are going to require a staggering $5.2 trillion in Capex, as compared to the $1.5 trillion that will power traditional IT applications, according to McKinsey and Company research. That kind of investment raises the stakes: if the data isn’t ready, the AI factory won’t deliver.

The use cases that are most likely to drive the need include:

• Sovereign/national infrastructure (water, roads, telecom)
• Advanced robotics
• Autonomous vehicles
• Drug discovery
• Personalized medicine
• Telecommunications network management
• Customer service
• Secure financial services

“Today, sovereign cloud is gaining steam, as major governments are leaning on hyperscaler and public/private partnerships to build out AI Pods ‘in country,’ with strict data sovereignty considerations,” said Shawn Rosemarin, vice president of R&D – Customer Engineering at Pure Storage. The company is working with governments to build localized data centers and infrastructure, with pre-validated, modular computing units that integrate compute, networking, storage, and GPUs, along with software designed for training, fine-tuning, and inferencing. “Underneath all of it sits a common requirement: the data must be ready—classified, governed, and accessible—before the AI factory can produce anything of value,” added Rosemarin.

In the private sector, businesses are increasingly involved in POC/POV testing to evaluate regulatory demands and key aspects of managing sensitive AI data. “Enterprises’ employees are excited about the potential for AI, given their own experiences with consumer LLMs. They seek similar access and capabilities in their internal content.”

The challenge: internal content is often scattered, inconsistent, and poorly governed. Before an AI factory can be justified, organizations must address AI data readiness:

• Where is the data?
• Who owns it?
• Is it cleaned, labeled, and governed?
• Can it be safely exposed to AI models given regulatory and privacy constraints?

Without answering these questions, even the most sophisticated AI factory risks sitting idle—or worse, generating incorrect or non-compliant outputs. Rosemarin believes the next phase of innovation will be coming in the “next round of scale,” which, he explains, includes:

• GPUs as the core processors for AI computations and training
• Networking to rapidly feed the GPUs and AI factoryconnect the GPUs and other components so data moves rapidly
• Storage at Exabyte scaleof vast amounts of data used to train and run AI models
• KV Cache for LLMs tohat accelerate AI inference
• Security and Governance to protect AI infrastructure and the sensitive data it supports. 

To this list, he adds a non-negotiable pillar: “None of this works efficiently without AI data readiness—the ability to reliably feed these systems with the right data, in the right format, under the right controls.”

Reality Check: LLM is not AI, and Machine Learning is not genAI

Three of the biggest obstacles right now are AI data readiness, finding the financial business cases that make the most sense (buildout versus benefit), and operational challenges, such as finding skilled workers. 

“In many cases, the Boards of many companies are telling CIOs to stand-up AI Infrastructure even before use cases and data readiness is established,” explained Rosemarin. “They are pressured to put in a data center and call it ‘AI.’”

Rather than infrastructure coming last, it is coming first, and often with an intent to improve agent productivity or accelerate application development, reduce customer churn or improve customer service as a starting point.

Because the urgency push to implement AI infrastructure often comes before companies have fully crafted their business case or articulated business challenges, AI can fail. That fact can have serious consequences: 

“The half-life of a GPU is six months, so by the time a company spins it up and feeds is AI-ready data, it could be two generations behind. In the realm of innovations in platform efficiency, edge analytics and alternative GPUs, a company could risk over-investment or obsolescence be even further behind.” 

Instead, Rosemarin recommends starting with the problem and the realistic data requirements defining the problem area as the starting point, such as building for:

• Operational efficiency focus on automating repetitive workflows where data is already structured and available through automation of repetitive tasks;
• Risk management such as detecting fraudulent financial transactions. Invest first in consolidating, labeling, and governing the underlying data streams, or managing risk through the detection of fraudulent financial transactions.
• Customer experience (chatbots and virtual assistants) to answer FAQs and route complex queries to human agents, making sure customer interaction data is captured and structured so the models can improve over time. Ultimately, that can lead to personalized product recommendations around customers’ needs and previous purchases and behaviors.

The point is that in each case, the readiness of the data should dictate the pace and scale of the AI factory investment.

AI as a foundation: from “killer app” to integrated value

While an AI factory is ultimately how some larger organizations are going to stand up the infrastructure necessary to support the whatever “killer apps” they believe will make their investments worthwhile, Rosemarin argues it’s more effective to consider AI as a foundational capability, with smaller integrated wins over time. 

The “impatience waiting for the killer app” mindset is common, according to Rosemarin, “but success requires a well-defined strategy that aligns AI initiatives with specific, measurable business goals. Productivity gains will come when AI capabilities are integrated with existing workflows.”

Critically, that strategy should start with a blunt question: Is our data ready for this? If the answer is no, the priority should be data readiness programs—centralizing, cleaning, governing, and securing data—before scaling into full AI factory deployments.

To learn more about how to sensibly and methodically build an AI factory foundation, check out the on-demand version of “AI factory: turning infrastructure into scalable AI production,” a panel discussion in which Rosemarin explores the foundational pillars and biggest challenges to AI factories with Thomas Nadeau, co-chair Evenstar Workstream, Open Compute Project, and Ted Weatherford, vice president of business development at Xsight Labs.