LAS VEGAS — The time period “AI manufacturing unit” got here up in a number of keynote displays, media briefings, and in interviews with executives at this 12 months’s Consumer Electronics Show (CES). (The present came about final week.)
And the idea appears to imply various things to totally different folks.
“The time period is totally being utilized in a number of — and sometimes sloppy — methods proper now,” stated Alex Cordovil, Dell’Oro Group analysis director. In conversations with operators, hyperscalers, and infrastructure distributors, he stated the primary means the time period is used is to imply “specialised knowledge heart.”
That’s additionally how Nvidia CEO Jensen Huang used the term. ”The explanation I don’t name it a knowledge heart is {that a} knowledge heart is a spot the place folks retailer knowledge,” Huang stated throughout a CES keynote.
“These AI factories are fully totally different from knowledge facilities,” Siemens CEO Roland Busch informed reporters on the present. “They eat extra vitality, want liquid cooling, and can’t be managed with regular constructing controllers — they want industrial-level controls.”
Booz Allen Hamilton CTO Invoice Vass is on board with that specific definition. “An AI manufacturing unit is a knowledge heart that’s designed particularly for AI-specific {hardware},” he defined. For instance, it might need particular concrete to assist the bodily weight of the server racks, and makes use of AI, digital twins and simulations to optimize energy, cooling, and server placement.
Others, nonetheless, appear to make use of the phrase to imply one thing smaller than a knowledge heart, referring extra to the servers, software program, and different programs used to run AI.
For instance, the AWS AI Factory is a mixture of {hardware} and software program that runs on-premises however is managed by AWS and comes with AWS providers resembling Bedrock, networking, storage and databases, and safety.
At Lenovo, AI factories seem like packaged servers designed for use for AI. “We’re trying on the structure being a hard and fast variety of racks, all working collectively as one design,” stated Scott Tease, vp and normal supervisor of AI and high-performance computing at Lenovo’s infrastructure options group.
That variety of racks? Something from a single rack to tons of, he informed Computerworld.
Every rack is a little bit larger than a fridge, comes totally assembled, and is usually totally preconfigured for the client’s use case. “As soon as it arrives on the buyer website, we’ll have service personnel join energy and networking,” Tease stated.
For others, the AI manufacturing unit idea is extra in regards to the software program.
Some consultants use the phrase in the sense of the AI technology stack. MIT Sloan Administration Evaluation’s Thomas H. Davenport and Randy Bean outline it as a mixture of “expertise platforms, strategies, knowledge, and beforehand developed algorithms that make it quick and straightforward to construct AI programs.”
“Not having the ability to construct on a longtime basis makes it each dearer and extra time-consuming to construct AI at scale,” they wrote.
The European Commission also has a broad definition of the term; it envisions ecosystems that embody computing {hardware}, human expertise, and knowledge with the tip product being AI fashions and functions.
The EC additionally talks about “AI Gigafactories” — large datacenters devoted to coaching next-generation AI fashions. In 2025, the EU allotted €20 billion to create as much as 5 AI gigafactories.
Lastly, Deloitte defines an AI manufacturing unit as a mixture of {hardware}, software program, and providers that allow the AI lifecycle. “The first product is intelligence, measured by token throughput,” stated Nicholas Merizzi, a principal at Deloitte Consullting and the agency’s AI infrastructure US providing lead.
Simply as a conventional manufacturing unit takes in uncooked supplies and generates a completed product, an AI manufacturing unit takes in energy, knowledge, and enormous language fashions (LLMs) and outputs insights, he stated.
Given how ubiquitous the time period has abruptly turn into, it’s vital to be clear on what’s being described when a vendor is promoting an AI manufacturing unit. Doe they imply the metaphorical form? An precise constructing? A sequence of server racks? It’s the type of misunderstanding that would price an organization billions if it invests within the mistaken idea.
Editor’s observe: Lenovo paid for Maria Korolov’s transportation and resort prices for this 12 months’s CES, however had no editorial function within the creation of this story.
This text first appeared on Computerworld.
