Hyperscale – where to next?

What’s next for data centers? As with every technological trend, it is difficult to predict with any certainty. However, one only needs to look at the current demand drivers and the technologies making their way into data centers to glimpse what the future might hold. 

My prediction? Data centers will continue to expand and become far more resource-efficient.

Multi-campus AI training

Like Pandora’s box, once opened, there’s no going back to a time before ChatGPT transformed the world. Generative AI is here to stay, and even the most sceptical among us would agree that the future will inevitably include AI in some form.

Even if AI advancements were to come to a complete halt right this moment, the changes already set in motion would continue to ripple through industries and society – ensuring AI’s ongoing use into the foreseeable future.

As global AI usage grows, AI inference will increasingly happen on client devices like smartphones and personal computers. This trend is already underway as device makers incorporate AI capabilities into their latest gadgets. Expect inference to also grow within data centers, driven by more advanced AI systems like OpenAI’s new o1 “reasoning” model.

What about AI training? Today, a fierce competition is underway as numerous tech firms vie to develop increasingly sophisticated AIs. For those at the forefront of this race, AI training infrastructure is expanding at a breathtaking rate. Cutting-edge AI models now require training clusters with up to 100,000 GPUs, which underscores the immense computational power required.

Multiple data center campuses with up to 300,000 GPUs, directly interconnected with dense networks of dark fiber, are being designed and built. This makes million-GPU clusters no longer improbable.

Enterprises are just as eager to embrace AI. While they are unlikely to train AI models from scratch, they are deploying private AI models tailored to their specific business needs and data sets.

More sustainable data centers

Data centers will become more efficient too. The largest tech giants are already implementing direct-to-chip cooling for their AI workloads to reduce energy consumption and lower their energy bills. Expect this trend to accelerate as enterprises seek sustainable solutions for their growing computational and AI demands.

The market is ripe for liquid cooling. Over the last few months, I’ve seen and learned about liquid cooling solutions from various vendors. Liquid cooling isn’t new, but what has changed are the refinements in design and how they are productionalised for versatile, large-scale deployments.

To be clear, data center deployments with lower power demands won’t disappear, so air cooling will remain prevalent. However, a vibrant ecosystem of hybrid cooling solutions will offer more options for discerning users to move the needle on sustainability. Finally, the integration of renewable energy sources into data center operations will become standard practice, though this may take many years to fully implement. 

The future of data centers is shaping up to be more than just technological advancement, but a significant shift in digital infrastructure. We are entering an era defined by unprecedented scalability, enhanced efficiency, and – I hope – a steadfast commitment to sustainability.

***

This piece was originally published in Issue 6 of w.media Cloud & Data Center Magazine