Background
Sumit Mukhija, the Chief Executive Officer of DCI Data Centers will be speaking on how AI is impacting digital infrastructure at the Melbourne Cloud and Datacenter Convention on 3rd April. He speaks on the basis of more than three decades in the industry and experience of working for leading companies such as ST GDC India, Tata Communications, Microsoft and Cisco. The move to Sydney was prompted by being “hunted” for a leadership role by Brookfield and his subsequent move to DCI.
Across the number of markets in which he has worked, he considers there to be an ‘essence’ of ‘data center’ “in design as well as in operational aspects whether it is located in the US, Europe or the APAC market.” What changes, though, according to Mukhija “are the local nuances, such as land considerations, approval processes and the vendor and contractor ecosystem.”
Australia, he suggests, may be more stringent in terms of data center construction but he is surprised that “despite being such a large market, there is hardly any local manufacturing of key components, such as chillers, UPS and generator sets or even electrical panels. So almost everything is imported into the Australian market and the New Zealand market.” whereas in some of the other geographies where he has worked, those were manufactured locally by global companies, and that enabled faster delivery and was more economical.
Mukhija sees DCI’s forward momentum as being driven by “our pride in being able to operate in markets where we can really solve some customer problems and customer concerns and meet their requirements, which is what we will continue to follow.” The company operates data centers in Australia and New Zealand with further facilities in development in those two markets and in South Korea. Mukhija sees this approach as being effective in maintaining a strong customer focus: “We don’t want to spread ourselves thin by going into 20 geographies at the same time. We will focus on expansion into one or two more key markets over the next 12 to 18 months, where we can really make a difference with the customers there.”
AI and Data Centers
In terms of Artificial Intelligence, Mukhija starts with the observation that AI is not a necessarily a new technology:
“AI refers to the capability of any system or technology or an algorithm that has the capability of performing tasks that typically required human intelligence in the past. Yes, so is that new? Well, honestly, no. Automation, analytics and robotics have been around for decades. We’ve used them in multiple shapes and forms. We just perhaps didn’t call it AI or we maybe didn’t sort of appreciate it enough. The biggest difference is that AI in its modern form leverages deep learning and billions of parameters and spans across many more areas. All of a sudden the industry had a ChatGPT moment of sorts and everyone started talking about AI.”
In terms of the impact on data centers and digital infrastructure more broadly, Mukhija sees a parallel to the early years of cloud (before it became mainstream) when “I used to get a call from the IT heads and the Chief Information Officers of companies that were working with me, and they used to have a similar question about cloud to ask. It is my scorecard for the year. Can you tell me, how do I go about adopting the cloud? Can you tell me, how do I use it in my environment?” Now the questions have moved towards the scale on which to move to the cloud and the workloads to move there. He sees the adoption of AI as following a similar path and that discussions around how to support higher density equipment in data centers will be key to this.
“A couple of years ago when ChatGPT happened, generative AI picked up steam and I got similar calls as to what is AI and how can I use it in my environment?”
He sees the experience whereby the industry learned to adapt to cloud as providing a precedent for AI and that AI will become ‘mainstream’ in a far shorter period of time than did cloud. He believes also that while there has been a focus on the training side of AI, that mainstreaming will happen when inferencing commands a similar focus and when it goes to the edge and there are actual use cases and applications leveraging the models that have already been trained. He considers this will take only another two to three years.
Mukhija goes on to discuss how AI is changing the decisions that need to be made in relation to technology: “Now we are having to rethink and use or choose technology choices driven by AI. Where do we deploy the infrastructure that will support AI?”
Compute and Facility
He notes that this is changing the data center further since higher power densities have an impact, particularly on cooling. This is complicated, he suggests by the fact that a Direct to Chip liquid cooled setup, which is the predominant form of liquid cooling used in the Industry currently, is in reality, a hybrid cooling system with 60 to 70% of liquid and 30 to 40% is still being air cooled in terms of implementation.”
This situation introduces new complexities and responsibilities across the value chain and in terms of deployment, the choices are between relying on the edge and also between existing cloud availability zones versus creating new AI or ML-specific zones. He sees the evolution of AI as challenging the deployment topology, not just the choice of the components. He suggests that the answer “relies on the stated end use, whether the requirement is for training or inference and whether it’s for an actual end use application which needs to be closer to the source of data via edge or whether it is a training model which repeats its processing cycle and therefore is located in a central position.”
The next area of impact that Mukhija focuses on is the relationship between compute and the data center itself:
“Data center deployment and technology choices have always dovetailed into the changes in computing. This includes the actual compute, the storage, the network departments. Compute components have consistently worked to become faster and denser, mostly in terms of power.” He illustrates this through the evolution of compute “from centralized systems and mainframes to start with, and then we had distributed servers which are loosely comparable to the data centers of today and then there’s the move back to hyper-converged systems for private cloud deployments. And in more recent years, we have seen public cloud deployments at scale. So both of the deployments of private clouds and public clouds happen largely at the core.”
As part of the development of AI-ready infrastructure, Mukhija notes that processor capabilities have consistently increased exponentially, and so has power consumption. He identifies that CPUs with a thermal design power of 200 to 300 watts are becoming very common and the top-selling processors are even touching 350 watts per processor. And these increases are magnified as multiple processors go into a system, and then multiple systems go into a rack. In terms of processors, GPUs primarily support AI applications, and they have designed power of up to 1000 watts per GPU. He observes: “these power densities were previously unheard of and now they’re becoming mainstream. Therefore the typical data center rack which back in the 2010s used to be 3 to 5 kilowatts for any enterprise application contrasts with today where even the simplest of enterprise applications using CPUs, not even GPUs, would get designed at 10 kilowatts. Some of the cloud deployments which are happening with AI are being designed at 20 to 30 kilowatts. AI infrastructure which primarily uses GPUs gets to considerably higher densities of 80 to 120 kilowatts.”
‘A Strategic Transformation’
Increasing load capacity means also changes to the commercial and transactional structure of the industry. Mukhija states that operators and customers need to work more closely than ever before while customers must be willing to pay a higher set of upfront costs because inherently, the AI driven data center infrastructure is costlier. His opinion based on his experience of working with such setups suggests higher capex of 30 to 40% and higher costs per month compared to what customers used to pay. Customers can be assured in terms of overall costs since they will be lower costs of components and operating costs of the chillers and lower energy consumption because of lower fan use. These will contribute to lower PUEs as well. So, the total net cost of ownership will reduce.
There will be the need to hire specialized resources across various trades, and especially for liquid cooling, and newer components such as CDUs, pipes and manifolds and newer plumbing requirements with components such as sensors and Flow Meters.
Mukhija believes that “This is not just a technical transformation, it’s a strategic one, and needs to be handled that way.” He uses the implications of transformation for security as an example of this:
“The operational SOPs [Standard Operating Procedure] need to be easier because you have newer components, you have water coming all the way into the data hall. Now you will have more people. You’ll have more visitors. You will have more technicians having to enter the data halls to take care of the plumbing or to take care of the mechanical components that are now placed inside the data hall. So this raises new concerns on security and safety, and some of the contractual obligations that we have with our customers need to be reviewed jointly with them and agreed on as well.”
AI and data center operation
Mukhija also considers how AI might help the data center in order to balance the debate: “Everyone talks about the impact of AI on data centers in terms of infrastructure and technology and deployment changes but very few people talk about what AI can do for the data center itself, yes, and the way we manage it.”
Again, the contribution of AI to the data center is not new: “Fan Wall units which can take the return air temperature and adjust the amount of cooling, or even adjust the amount of water that comes in on the basis of those input temperatures. We’ve had automation. We’ve had logic.”
Mukhija indicates that AI takes the principles of data center management further through its processing and analytic capabilities: “There is also this whole new set of correlations between alarms and other signals. BMS was just about monitoring. Now we can actually use all of the data, correlate, and get trends, get some sort of optimizations achieved on the fly. And so there’s so much more that we can do with all those sensors and data and alarms that we used to get earlier. We can even detect security breaches based on its ability to correlate all of that data using trained models of AI and using inferencing and an inference based application to really make sense of the data, and, you know, prevent a future event, or do a root cause analysis, or enable optimization, it is so much more now.”
He suggests that the new generation of data centers are actually adopting this much more but that they just don’t talk about it enough.
In Conclusion
In conclusion Mukhija asks: “how can we really prepare for some of these changes?”
He suggests that preparing for the advent of AI will always be an ecosystem play which brings in the whole industry: “it will never be one specific company, never just a cloud services provider play, or just a data central services provider play or an OEM play.”
So, he presents a case for AI representing a challenge not just to legacy technologies and operation but to the whole system of data center ownership and management, the whole traditional data center way of doing things:
“AIl that data center transformation spans across design, construction, choice of technical competence, especially around holding people, processes, risks, safety, security and even the contractual aspects. Whether it is enterprise or cloud service providers all must now adapt to new power and cooling paradigms as we need to plan ahead for AI driven workloads. So we need to now start designing for AI and not really trying to retrofit an existing facility for AI, which obviously is time consuming and may be ineffective.”
Sumit Mukhija will be speaking on the subject of ‘The Impact of AI on Digital Infrastructure’ at the Melbourne Cloud & Datacenter Convention on 3rd April 2025 at CENTREPIECE, Melbourne Park. Further information is available here:- https://clouddatacenter.events/events/melbourne-convention-2025/
