Data Centre Titans: Why the Rack Is Now the Unit of Compute – Vik Malyala, Supermicro
Thu 30 Apr 2026
As AI workloads move from pilot environments into sustained production, data centre design is becoming more tightly integrated. For Vik Malyala, Managing Director & President, EMEA and SVP Technology and AI at Supermicro, the rack is emerging as the primary unit of compute, bringing power, cooling, and networking into a single architectural decision.
In this Data Centre Titans interview, he discusses how rack-scale systems are reshaping deployment models, energy planning and infrastructure strategy for AI-driven environments.
– – – – – –
How would you describe your current role at Supermicro, and how has your experience across semiconductor and infrastructure sectors shaped your perspective on system architecture?
My current role is Managing Director & President, EMEA / SVP Technology and AI at Supermicro. I’m responsible for working with customers to bring the latest technology to their data centres and beyond. Historically, customers purchased full systems; however, now we are seeing that the rack is the lowest level of compute, providing more versatility and customisation around data centre infrastructure.
Supermicro itself creates a wide range of servers, but the real emphasis is on delivering to customers optimised clusters, ready for their workloads from day one.
Looking back across your career, what structural shifts in compute and infrastructure have most influenced how you think about scaling for emerging workloads?
Technology in general has evolved rapidly, especially in the last handful of decades. It‘s fascinating to experience this shift from MFLOPS to GFLOPS to TFLOPS based on CPU architectures, to systems now tailored to specific AI workloads. Single systems are not sufficient anymore, and the overall solution, complete with networking and storage, must be designed concurrently.
This includes the infrastructure to allow for scaling, not just within a server, but to thousands of servers. Careful attention to the networking must be designed early in the process. It is not just connecting a single Ethernet cable anymore, but multiple networks must be present for different types of data transfer within (and across) data centres.
What have been the most instructive challenges in delivering infrastructure for next-generation compute, and how did those experiences change your approach?
I have seen a tremendous explosion in the available compute performance in the last 20-plus years. What has changed so much is that the systems need to be networked for scaling across hundreds to thousands of compute nodes, mostly for the AI workloads. By working closely with our suppliers, we can create massive AI workload-tuned clusters and entire data centres.
As AI workloads move from pilot to sustained production environments, how is data centre architecture evolving at a system level?
Data centre designers need to consider all of the aspects of deployed infrastructure from the start. This means that aspects including power delivery and water usage are part of the planning process, before any construction begins.
Supermicro works closely with customers to design an optimised solution, which includes servers, storage, networking, complete cooling systems, and data centre management software. Complete solutions can then be assembled in a manufacturing facility, which ensures a smoother installation at the customer site.
With rack densities increasing significantly in some environments, how should operators be rethinking power distribution, thermal strategy and physical layout to maintain resilience?
As a first step, operators need to focus on the system design up front before even the first servers arrive. Additionally, they need to have a long-term mindset, planning for future scalability. While power is becoming a critical factor in planning, liquid cooling is an essential part of this.
Sophisticated software can be used to interface with and control the power use of the system itself and the associated IT infrastructure, in almost real time. By using liquid-cooling, an additional amount of computing power can be installed with the same amount of power draw, or even at a reduced level. In many cases, a data centre may have different “halls” with a mix of cooling solutions; some with liquid and some with air.
Where do liquid cooling and direct-to-chip approaches fit within the broader infrastructure mix, and under what conditions do they become essential?
By using liquid cooling, more power can be delivered to the systems, increasing the computing levels. While the densest racks today require liquid cooling, moving forward, with the increase in the thermal design power (TDP) of every CPU or GPU, this will be the preferred solution. Furthermore, some high-performance configurations may need to lean towards immersion cooling during intense workloads.
In general, liquid cooling becomes a necessity above around 30+kW. At this threshold, traditional fan, air and HVAC systems will not be sufficient. Liquid cooling in these configurations is far more efficient and economical, with the potential to reduce the overall power consumption of a data centre by up to 40%.
We are seeing growing interest in rack-scale and pre-integrated systems. What factors are driving this shift, and how does it change operational complexity?
Clusters are complex and require expertise to set up and test, with software installed, in advance of delivery to the customer. This preparation is important as it reduces the time needed to install and deploy at the customer site.
For example, systems as intricate as cabling between multiple clusters are often overlooked. Failure in designing and planning for this may result in inefficient or low-performing servers post-installation. At the rack level, server stacking, configurations with specifying units, power shelves, networking, coolant distribution, and storage require deep engineering knowledge of how different components work together. Only by testing at the rack level can engineers be sure that all the components work together as intended.
As enterprises and hyperscalers accelerate AI deployment, how are global build strategies adapting?
The architecture of both physical and IT infrastructure needs to be a factor at the start of the system design process, not an afterthought. This includes the power distribution, cooling, and layout. AI, compared to traditional enterprise workloads, requires specific and optimised hardware systems.
Global manufacturing enables these complex racks and clusters to be assembled fully in factories around the world. This reduces overall costs through lower shipping fees and enables customers to participate in the design and build process, making everything more collaborative to ensure end systems meet their needs. At the customer side for the facilities, advanced planning needs to be in place, so that the systems do not remain idle before they become revenue or token generators.
Where do you see structural bottlenecks emerging — whether in energy access, cooling innovation, supply chain resilience or system integration?
Energy availability has become the AI bottleneck. Current wait times are now measured in years, not months. In addition, the amount of energy that can actually be delivered is also being impacted by system capabilities and geopolitical uncertainty. Thus, any technology that reduces power needs is becoming very important and critical to building or running AI-focused sites. Additionally, many data centres are starting to build their own power sub-stations to guarantee power in the long run.
Looking ahead three to five years, what structural change do you expect will most significantly reshape AI-driven data centre infrastructure?
We are already seeing how a single rack is designed to work as a single “computer”. Looking forward, making racks larger allows for fewer restrictions on system efficiency and cost, allowing for the integration of more components to increase performance.
New technologies for power delivery, both to the overarching data centre and to specific racks and servers, will enable more critical computing within CPUs, GPUs, memory, and networking. AI-focused customers are buying multiple racks, complete with all the different configurations and components they need, directly, meaning they can strategically reduce deployment time and increase reliability.
