Hello Reader,

Earlier this week, I sat down for an informal chat with two gentlemen from the data centre industry in Singapore on a business trip. As with all good conversations, I shared freely but left with new perspectives and ideas.

And one thing that I'm increasingly convinced about is this: Future data centres will need to run a mix of traditional and AI workloads. In short, they will be "hybrid" facilities.

Why hybrid-AI is the future

With all the news about the mind-boggling amount of money poured into AI data centres, it's easy to lose track of the fact that these high-tech facilities, stuffed to the gills with GPUs, are actually terrible for general computing purposes.

This is because GPUs are unsuitable for standard IT use cases like powering cloud services or handling enterprise workloads such as database or ERP deployments. Notwithstanding the fact that GPU servers such as the 7U "ESC N8-E11V" Asus GPU server have an MSRP north of USD$300,000 each, the relatively small number of CPUs means they will struggle with conventional computing needs.

If you are wondering how an AI data centre differs from a non-AI facility or why we even need them, I wrote about this at length in the latest edition of the W.Media quarterly magazine for January 2025. The story will make it to an online feature at some point. In the meantime, you can access the story "Building the data centre for tomorrow's AI" on pages 11 to 13 here (free registration).

More AI or less?

I daresay no one really knows where we are headed with AI. Some popular predictions show demand for inference in the data centre will overtake AI training over the next few years. This could be realised more quickly given how much more processing power "reasoning" AI models require - and how popular they have become.

Yet inference is also shifting from the cloud to local devices such as laptops and smartphones. In China, the DeepSeek LLM is making its way into household appliances such as TVs, fridges and robot vacuum cleaners. Apart from reducing latency and dependency on Internet connectivity, on-device AI could significantly reduce the need for AI data centres.

As it is, expect predictions of demand for AI data centres to vary widely depending on who you ask. Is it a tech giant with a vested interest in raising more money to meet insatiable infrastructure costs? Or a certain software leader intent on pushing the message that perhaps we don't need quite so many AI data centres after all.

Two things we know for sure. AI isn't going to disappear overnight. And we will still need traditional servers to power all the digital services we now use. As AI adoption eventually picks up around the world, this would mean that we can expect to see both AI and traditional workloads within the same data centres.

An AI-ready data centre

What would a hybrid AI data centre look like? I didn't give it a second look when I first came across the press release for the new data centre by SM+ in Jakarta, which had its groundbreaking on Thursday. As I studied the plans, however, I become convinced that this is a model for hybrid AI data centres of the future.

In my quick call with Avnish Patankar of SM+, he explained how certain floors will be allocated for high-density, mid-density, and low-powered workloads. He also told me there will be adequate floor space for the requisite cooling hardware for up to 60MW of capacity - should demand come in. And oh, there will be a power substation built on the grounds of the data centre.

Overall, a very pragmatic and highly flexible design. Exactly like what a data centre would look like to meet the diverse needs of tomorrow's digital landscape for both AI and traditional workloads.

What are your thoughts on how data centres will evolve to accommodate both AI and traditional computing solutions?

As usual, you can just hit reply to this email to reach me.

Regards,
Paul Mah.

Spotlight