Most organizations scaling AI are pouring investment into GPUs and computational power — but overlooking the layer that determines whether those GPUs actually deliver.
At ATxSG 2026, TuringData's VP of International Sales & Market Expansion Nikhil Madan explains why the data infrastructure layer is the hidden bottleneck in enterprise AI — and how TuringData is solving it.In this interview, Nikhil covers:
- Why GPUs underperform without data fed at the right speed and volume？
- How TuringData's software-defined AI infrastructure lowers the cost barrier to deploy AI at scale?
- Integrating legacy data (files, images, videos) scattered across cloud, edge, and core into the AI ecosystem？
- Why Singapore is TuringData's strategic headquarters？
TuringData delivers full-stack, software-defined AI storage infrastructure — purpose-built to make GPU investments work harder, and accelerate AI model training and inference.
Transcript
00:13 About TuringData and Nikhil Madan
My name is Nikil Madan and I am responsible for the global sales of TuringData. We are headquartered in Singapore and we're looking forward to the event and engaging with people, customers, partners and see how we can work together and solve the problems of bringing AI to the community.
00:30 Could you share what inspired TuringData to make Singapore its strategic hub for global and especially Southeast Asia, and what unique capabilities do you bring to our local Al ecosystem?
Nikhil Madan: Singapore was a deliberate intent because Singapore is the gateway to Southeast Asia. We clearly see the maturity. We see most of the organizations building their AI strategy beginning through Singapore. So it's clearly the gateway for the digital highway into ASEAN into Southeast Asia.
We also see a level of maturity of AI adoption in the region driven by the government. The government clearly thinks about strong AI policy, strong AI strategy and that's why we're seriously interested in building our headquarters in this forward-looking country in Singapore.
As far as AI goes, clearly we see most of the customers really thinking about building AI and the focus so far has been about building more computational power, more GPU.
But one of the biggest things that we solve today is how do we solve their data problems? How do we get them to use their GPUs more efficiently? It's a huge investment in GPU. We help them build a data infrastructure layer which helps them use their GPUs more efficiently. So that's been our focus so far.
01:42 In your experience, why is the 'data layer' often the hidden bottleneck in achieving peak Al performance, and how does your technology ensure that these expensive computing resources are fully utilized?
Nikhil Madan: That's been a common myth in the market.
Clearly, people have been thinking when they think about AI, they think about GPUs, they think about computational power. One thing that's kind of suddenly missing is the data and the information piece. It's like building a Ferrari without good fuel, good power to power the Ferrari.
So that's where we are. We're trying to solve this problem by building the data infrastructure layer for organizations who want to build a strong AI. While you invest a lot of money in building the computational power and the GPUs, I think the good quality and intelligent AI comes in if you have all the information and you have the GPUs being fed the data at the right speed.
Only when the GPUs are getting the information at the right speed they are able to generate an intelligent AI and that's been where our focus is. So not only just the volume of data, all the information, but also at the speed at which the data is being fed to the GPU both put together gives you a good mature AI that can help you get the use case that you want out of AI.
02:57 What is the most critical hurdle these organizations can overcome by partnering with TuringData, and how do you help them accelerate their journey from raw data to a fully trained model?
Nikhil Madan: When somebody wants to think about AI, you really need to think about what's your data AI strategy. Right? When you start thinking about data, you realize that the only way for you to get a good quality, good intelligent AI that you can really use is having as much information, and having the right amount of data. And it's not just the volume of data, but it's also at the speed at which you can pump the data to the GPUs.
So you need information, you need the speed at which the data can be pumped.
But while you're doing this, one of the biggest problem that most of the organizations I've seen in the region is everybody's got lots of legacy of data.
How do you turn your legacy data which is lying across the organization in the cloud, on the edge, in the core? How do you integrate that into AI?
While you're doing all this, one of the biggest barriers for AI adoption has been the cost. It needs a high CAPEX. So we've been focused on building a low entry barrier to deploy AI at scale.
So what we've done is build a software-defined AI infrastructure layer which runs on commodity x86 architecture keeping the cost of building AI from a data infrastructure standpoint at the right minimum. So you can really start small, grow as you grow big but still make your GPUs work at really high efficient utilization that everybody wants.
While doing all this, we take care of the legacy data that's there in the enterprise. We integrate hundreds and billions of files that are lying there — those are tokens, those are images, those are video files. We integrate them into the AI ecosystem.
So clearly we're solving the problem of the enterprises today when they want to build AI. We give them an efficient AI. We give them a high performance AI and we give them a way to efficiently use all the data that they already have. So clearly, we're really aligned and wanting to bring AI to everybody in the market.