Sunil Gupta, co-founder and CEO of Yotta Data Services, is at the forefront of India's AI infrastructure revolution. Under him, Yotta became India's only NVIDIA Cloud Partner, and it's deploying thousands of GPUs at scale to cater to AI demand not just in India but globally.

Yotta under Gupta has big plans for India as it works towards democratising AI by making it accessible via GPU leasing, AI Lab services, and end-to-end AI solutions, enabling startups, students, and enterprises to build and deploy AI solutions, which are specific to India's unique problems.

In a freewheeling discussion with Fortune India's Manoj Sharma, Yotta's Sunil Gupta talks about his company's bold bets on GPUs, large-scale data centre capabilities, his views on digital sovereignty, vertical-specific AI solutions, and how India can become a leader in fostering AI innovation across industries.

Sharma: How is Yotta the only NVIDIA cloud partner, and how has that association driven AI infrastructure growth in India?

Gupta: This whole thing started about two and a half years ago. ChatGPT came on the world stage around November 23, 2022, and immediately the whole world started talking about AI, and AI, which was always in the backside, suddenly came into the hands of consumers. In India, too, a huge narrative got built, a huge number of discussions took place in government circles, and in industry circles. I used to participate in discussions on industry forums about how, while India is a software and services leader, and we export 13-14% of the IT from a world point of view, but that how India is not a leader in AI.

India has the potential to be both a supplier and consumer of AI, but it lacks large-scale GPU infrastructure. Without GPUs, building foundation models in local languages, datasets, and cultures isn’t feasible, nor is large-scale AI consumption. In August 2022, NVIDIA co-founder Jensen Huang visited India and met the Prime Minister to discuss India’s AI leadership and applications in agriculture, healthcare, and other sectors. Jensen emphasised that while India could be a large market, someone needed to invest in NVIDIA’s chips. At that time, there was a global shortage of GPUs, but he assured priority access for India.

Government channels reached out to Reliance, Tata, and Yotta. Initially, Yotta used A100s, A40s, and B40s for online gaming, content creation, and cloud services. NVIDIA India saw Yotta as agile and capable, given its data centre infrastructure and sovereign cloud setup. Overnight, I met Jensen in Pune and ordered 16,000 H100s with phased delivery.

Yotta already had the data centre capabilities to run GPUs. With India poised to become a major market, GPU scarcity was the main barrier. Initially, I offered GPUs to the global market, drawing training demand to India, while domestic demand gradually grew. Early adopters included IIT Bombay (BharatGen), Sarvam, Zoho, Fractal, Cure, and IIT Chennai.

As global demand tapered and domestic demand grew, GPU leasing alone wasn’t sustainable. We needed an orchestration layer, PaaS services, and a token-based model. A key milestone was becoming India’s only NVIDIA Cloud Partner (NCP) and one of six globally to follow NVIDIA’s reference architecture. Global partners include Lambda Lab, Nebius, Crusoe, and SoftBank. Yotta’s infrastructure fully implements NVIDIA’s reference architecture across hardware and software, including Lepton (powering a 4000-GPU cluster for Sarvam), NVCF (inferencing), NIMs (model containers), and Blueprints, all accessible to customers through Yotta.

Sharma:  I’m curious — at that time, did you anticipate that GPU demand in India would eventually grow this rapidly, especially when there was limited activity on the government side? And looking at the present scenario, how does the demand split between your global and Indian customers?

Gupta: I mean, I would never have taken a bet of 16,000 GPUs at that time, because 16,000 GPUs was more than Rs 8,000 crore. I would never have taken that bet if this conviction were not there. And the logic was simple: that globally, the ChatGPT moment has happened. The US market is very big. India typically follows the technology that has matured in the US.

Because of the various dynamics, the digital adoption rate in India is far higher than even in the US. So, technology will follow in India. And if we can get on that technology infrastructure upfront, whenever the wave comes in India, we will be the biggest beneficiary and the leader. India has all the right drivers to become a much larger market than the US in the time to come. So, this was obviously a conviction of the future market.

I ended up visiting NVIDIA headquarters twice. Obviously, there was a lot of discussion, but that conviction was there right up front. And second, more importantly, was that I also had all the underlying ingredients. I own my own data centres, sovereign cloud, so in terms of skill sets, so many of the ingredients are already there.

Talking about the current mix, when I started, it was almost 75-80% global, and the Indian demand was very less. In between, a phase came where the global demand started petering out, and India's demand started growing. And so, a time came when global demand was almost 25% and India's demand became 75%. It also happened just about six months back.

Now the Indian demand is going to go through the roof. India is acting as a convergence point for all the demands of all the startups and IITs. We are the biggest contributor to India's AI. So, the Indian demand is going through the roof. And if not millions, it's definitely getting into a few lakhs of GPUs in the next two years. India AI already has 503 proposals for GPU requirements, and I presume the sum total of all these GPUs adds up to more than 2 lakh GPUs. So India's demand will definitely be very big.

Sharma:  Now, let's talk about your data centre business. How do you see the overall demand for data centres in India? Where's your presence, and how do you plan to expand?

Gupta: When I started this whole blueprint of Yotta Infrastructure six years back, GPU was not a big story. The two stories were Yotta Co-lo and Yotta Tech. Co-lo was about building large campuses, thanks to my Hiranandani Group background, where each campus could host multiple data centre buildings connected through fibre, with our own power substations and fibre. I planned six or seven such locations across the country to host hyperscalers, banks, and governments.

The second was creating our own sovereign cloud in India for niche segments that don’t want to go to foreign cloud operators but still need a good-quality cloud. Geopolitical developments have validated this strategy. Governments always bought into this logic, which led to contracts from the National Informatics Centre (NIC) and the Software Technology Parks of India (STPI). NIC gave us contracts for managing data centres in Delhi and Pune; STPI for Bhubaneswar and Chennai. We invested heavily in these, set up our own clouds, and large government contracts are now coming onto those clouds.

Today I’m focusing mainly on two campuses—Mumbai and Delhi (Greater Noida). The Mumbai campus is 70 acres, part of a 600-acre Hiranandani township. It can scale to 1 GW with multiple buildings. We host co-lo customers, private and public clouds, and have all our GPUs installed there. Since the data centre and power are owned by us, we can scale massively.

The Delhi campus sits on 20 acres and can scale to 250 MW with six data centre buildings. We’ve built our own power substation and fibre network. There’s also a small boutique data centre at Gift City, serving local customers. Earlier, the plan was to have a data centre in six locations. That’s changed. I’m no longer chasing pure co-lo. The business mix has shifted toward cloud, GPUs, AI services, and sovereign cloud. These can be delivered from Mumbai and Delhi. If needed, we can expand to other cities later, but that’s not the core vision.

Currently, we have 8,000 NVIDIA H100 GPUs live and are implementing another 8,000 NVIDIA Blackwell B200 GPUs by November–December. India's AI demand is high, and the next lot may be NVIDIA Blackwell B300. India will require lakhs of GPUs, and I intend to be a big leader in this space. I got an early lead in a very blank space.

Sharma:  How is your company democratising AI adoption to AI lab as a service?

Gupta: We have a huge focus on democratising AI. First, GPUs are expensive, ₹2-3 crore a box, and not practical to host in offices due to power and cooling needs. By investing in massive GPU infrastructure, we’ve created an abundant supply in India. Two years ago, people asked, “Where are GPUs?” That question is irrelevant now. With subsidies, GPU access in India costs under $1 per hour versus $4-5 in the US. Any startup can rent GPUs for a short period without capital investment.

Second, we launched programs with NASSCOM and NVIDIA. Through NASSCOM’s DeepTech and AI Foundry programs, hundreds of startups receive $15,000-$25,000 worth of GPU credits to build and test products. Through NVIDIA’s Inception Partner network, startups across India, the Middle East, and Africa get $10,000-$15,000 in credits, plus access to NVIDIA engineering support.

Third, with our AI Lab model, startups don’t need a full GPU; they can take a slice. A GPU with 80 GB can be divided into smaller virtual desktops, accessible remotely via a regular laptop. This brings AI computing to anyone with an internet connection. We’re in talks with All India Council for Technical Education (AICTE), IITs, colleges, schools, and tech companies to offer this at ₹25,000-₹50,000 per month, enabling students and developers to build small models and apps. The goal is to grow the entire ecosystem, from students and startups to enterprises, so AI isn’t limited to a few large players.

Sharma:  How do Yotta's end-to-end AI solutions differentiate themselves from the pure compute power it provides through its data centres?

Gupta: On compute, what we’ve built is very different. Just buying a GPU box and connecting it to the internet is one thing. But implementing a full reference architecture with thousands of nodes, InfiniBand, BlueField, high-speed storage, and the NVIDIA software stack benchmarked by their engineers is another. This is also why IndiaAI allocated models like Sarvam and Bhashini to us, while others didn’t get more than a handful of GPUs. Many setups aren’t suitable for training because they lack a single GPU fabric. That’s our biggest advantage, and why NVIDIA routes global demand to us. If you see Jensen’s speeches at Computex Taiwan and Paris, he mentions Yotta among their top partners.

But GPUs will become a commodity. Others and hyperscalers will scale up. What will set us apart is the platform layer we’re building on top of GPUs. We already provide full Kubernetes and Slurm clusters, VM layers for startups, and unique AI Lab setups for institutions. A college can invest in just 10 cloud workstations and allocate them to hundreds of students in time slots.

We’ve also built a large inferencing layer. If someone wants to use LLaMA, they don’t need to buy GPUs; they can simply pay per token and use it. We have access to open-source models and NVIDIA Studio models through our partnership. These are the services the market needs today: training, fine-tuning, RAG, inferencing, and AI labs. We keep evolving to stay aligned with demand. India lost the cloud opportunity to hyperscalers because local players didn’t build at the right scale. With AI, we’re at a similar moment. By building the services layer early, we can compete strongly with hyperscalers.

Sharma:  Now let’s talk about digital sovereignty in AI, something the Prime Minister has also emphasised. What’s your perspective on India’s approach to building AI capabilities domestically rather than following someone else's playbook?

Gupta: We must create our own infrastructure, from chip layer to server layer to dataset layer to model layer to app layer. At the same time, we shouldn’t become inward-looking. But we must have the confidence that if someone tries to isolate us, we don’t get isolated. We must have self-sufficiency.

The government is fully aware of this. If you see the big picture, the IndiaAI mission clearly talks about building large language models. There was a debate a year back — Mr Nilekani said India should be the use-case capital of the world, not replicate what the Valley has done. But tomorrow, if someone says “close your weights,” your market disappears. So India must build its own LLMs, both for sovereignty and because foreign-trained models won’t work for Indian languages.

Only a small part of India is proficient in English. Most people driving digital adoption live in tier 2 and tier 3 towns and villages. They are digitally enabled but speak in their native languages. Social media activity is happening in these languages. If we give them AI that works voice-to-voice, in their language and slang, and it responds the same way, imagine the use cases, agriculture, education, healthcare, climate, everything. This is India’s priority, not necessarily a U.S. priority. So we must build our own models with our own datasets.

The good part is that there’s now a clear focus on building a GPU cloud in India. That’s why companies like us are getting attention, and others are also building. The semiconductor mission will be the final step. India is not yet doing chip fabrication. Chip design happens here, but the IP belongs to parent companies. India should design and soon manufacture its own chips. Micron Technology and the Tata Group are already working on this. Once we manufacture GPUs in India, we’ll have full independence, from chip to server to model to dataset to app layer. We need self-sufficiency while working productively with global tech companies.

Sharma:  Do you think the government is doing enough to build the infrastructure required for India to become self-reliant in AI? For example, through initiatives like the Semicon India Mission? Are you satisfied with the pace and results so far?

Gupta: Frankly speaking, as citizens, we’ll always have many expectations. But given the democratic framework in which our country works — unlike China — I feel a lot has happened. If you sit on the fence and look at what’s happening, things are moving in the right direction. There are always push and pull factors, which make progress slower, sometimes one step forward and two steps back.

Two years ago, there was nothing, and look at what the IndiaAI mission has achieved. Of course, speed can be faster, but they’ve done a great job. They also face skillset challenges; only a few people are driving a lot of work.

Two years ago, when I invested in GPUs, people said there was no market. Today, if you add up the 503-510 proposals IndiaAI has received, the requirement is over 2,00,000 GPUs, which the government wants to fund. They don’t want to stop at ₹10,000 crore. As much as I understand, there’s essentially an open budget if it helps grow the AI ecosystem in India.

The India Semiconductor Mission and the electronics component mission are also moving. Look at how India has captured mobile phone manufacturing; the value addition may be just 10%, but today, 44% of iPhone shipments to the U.S. are from India. This mission started about a year and a half ago, and the first chip production is expected by December 2025.

The government is not looking at things in isolation; they’re seeing the whole stack, starting from the chip layer. The Prime Minister visited several small countries in South America and Africa. Many wondered why. It’s because China has stopped exporting rare earths, which are critical for electronic components in EVs, mobiles, servers, and chips. The government is securing supplies from other countries like Japan and Australia to reduce dependency.

Sharma:  From your vantage point, where do you see India’s real edge in AI? Will it be in building large horizontal platforms like LLMs, or in developing vertical, problem-specific AI solutions?

Gupta: Yeah, one straightforward answer is — yes, vertical-specific solutions to solve problems. But that doesn’t mean I only focus on apps and ignore the underlying models, whether LLM or SLM. The idea is to build end-to-end vertical-specific solutions. As a country, we created the DPI framework and Unified Payments Interface. We’re expected to have multiple “UPI moments” in AI — in agriculture, healthcare, education, space research, and more. AI has use cases in every sector, especially B2C at a population scale. I’m not talking about B2B use cases for enterprise benefits but large-scale citizen-facing applications.

Vertical-specific solutions are key. But look at Bhashini. It’s not vertical-specific; it’s a horizontal platform for language translation and transliteration. It has now migrated fully to Yotta Infrastructure. From 2 million inferencing requests a week in April, it has crossed 16 million in four months. Any government website today can instantly be translated through its modular backend.

This shows the impact of horizontal language platforms, like a “ChatGPT in Bengali” or “ChatGPT in Malayalam”. There will be strong demand for such country-specific language platforms. But India’s real niche will be in vertical-specific solutions. We have massive data generation across sectors every day. If we build strong vertical use cases and scale them like UPI, exporting them globally becomes easy. UPI is already going to 39+ countries after its success in India. So yes, we will build horizontal LLMs mainly for language and voice but also focus strongly on vertical-specific solutions.