Exploring the AI Factory Farce

Here’s hoping I don’t get banned from every AI-related trade show and company in the future, which tends to be just about every electronics-themed trade show these days.

What I’m harping on now is the term “AI factory” that’s being bandied about by the biggest AI and computer vendors. The problem is these aren’t what we think of as a factory that creates physical items from physical resources delivered to the factory.

What is an AI Factory?

So, what is this AI factory that everyone is talking about?

It’s just a data center that focuses on supporting AI models such as chatbots and large language models (LLMs). They include most cloud services from the likes of Amazon Web Services (AWS) and Microsoft Azure, which are already being used for AI training and delivery.

The big difference between a real factory and an AI factory is the latter is just bits in and bits out. All changes are done in the virtual compute world, creating and training AI models or manipulating data using the trained models.

One reason many use the term AI factory goes hand in hand with wanting to build more factories or bringing manufacturing back to whatever country you happen to be in. AI factories don’t do this, but they’re useful nonetheless. More on that when we talk about digital twins.

Much of the AI factory discussion is about creating larger and more powerful dark data centers. Unlike a conventional factory that would have a few or many workers, these data centers have a minimal cast of characters to oversee their operation, occasionally turning on the lights to fix or replace components. All of the bits move through networking connections moving massive amounts of data around.

Digital Twins and Real Factories

The AI realm is already being incorporated into real factories, often in conjunction with digital twins. A digital twin is a software incarnation of a physical device like a motor or assembly line. It can be used in simulations as well as present real-time data from the physical device, assuming sufficient sensors are incorporated into the system.

The digital twin can be observational only. Data is presented to a user or utilized by an AI agent to monitor and analyze the system to handle things like preventive maintenance. Many systems provide remote control so that the user of a digital twin is able to invoke changes in the physical device. Agentic AI can be employed for real-time control without having a human in the loop.

Digital-twin support is possible with a conventional PC that’s networked with the physical device. Typically, though, the digital twin support is much more involved than a single device that perhaps incorporates an entire factory or assembly line, which requires significantly more compute power especially as AI comes into play.

Where are AI Data Centers Headed?

Assuming the AI chatbot bubble doesn’t burst, the focus on AI in the data center will continue. AI factories aren’t like real factories, as already noted, but they will play a major part in supporting digital twins and those working in physical factories.

Part of the challenge with big AI data centers and real factories is how much compute power should be in the cloud and what should be inside or closer to the physical factory. The AI data centers will have large amounts of storage and compute available. However, it usually will not be dedicated to a particular physical factory. Likewise, there are issues of support, security, reliability, etc., that become more complex with remote incarnations of digital twins.

While I wish everyone would stop using the term AI factories, I suspect that its use will continue to grow as confusion abounds. Just keep in mind that an AI factory is just an AI data center and you will be fine.