DeepMindSecure is a Noida-based voice AI company working on a problem most voice AI still gets wrong: Indian languages as people actually speak them.
They build small, focused speech models instead of chasing a giant LLM. Their 270-million-parameter speech-to-text model was trained on under 1,000 hours of recorded speech, covering Hindi and six of its dialects. Today it runs production voice agents for NSE-listed companies: pre-insurance screening for some of India’s top insurers, loan recovery calls for a large fintech firm, and lead calls in Malayalam for a major edtech company.
They’ve filed six patents, with one granted, including work on speech inference at 20 watts for cars, kiosks and phones.
One GPU, then two, then three
We’ve been building for DeepMindSecure since July 2025, one GPU at a time. We sat down with their founder to talk about how that turned into a four-GPU liquid-cooled workstation.
SAURABH VAJPAI
Founder and CTO, DeepMindSecure
Saurabh spent 23 years in corporate roles, most recently at Microsoft, after studying electronics at IIT BHU. The company started with something he noticed on the road. Drive a few hours out of Delhi and the dialect changes. Drive a few more and it changes again.
“Every 250 kilometres in India, the food and the language changes very drastically… There is nothing like Hindi. There’s nothing like Telugu. There’s nothing like Malayalam,” said Saurabh.
That meant training their own model. And training their own model meant three things had to fall into place.
What a model like this demands
1. Hardware, fast
Speech models live and die by GPU memory, and Saurabh knew it going in. “The amount of memory on the GPU is the key. The more you can have it, the better for you,” he said.
He asked ChatGPT which vendors in India could build a machine to train his model. It pointed him to us. In July 2025 he picked up his first RTX PRO 6000 Blackwell. One card became two, and two became three.
2. Cloud is a no-go
Saurabh found that Indian cloud providers price GPU rental to earn back a card’s full cost in about a year, while the card itself lasts a lot more years. By his numbers, one RTX PRO 6000 rents for ₹3 lakh a month on AWS at the best negotiated rate, and ₹1.5 lakh a month from Indian providers. Four cards puts that at ₹6 to 12 lakh, every month, for hardware you never own.
“When the rental in Bangalore starts to go to one lakh for a 2BHK, you take a loan and you buy that house,” he said.
3. Heat
Inference was the easy part. Training is where it gets ugly. Every GPU runs at 100% for hours, and every card keeps talking to every other card to share weights. All of that power turns into heat.
Saurabh describes it with something anyone from north India will recognise: the Loo, the hot, dry wind that blows across Rajasthan and the northern plains in May and June.
“If you have ever experienced a Loo… you stand behind one of the AI servers who is doing a training run. You will immediately experience the Loo. For any company who wants to train models, my sincere advice is go for liquid-cooled servers.”
Creating a modern-day supercomputer
So we went liquid. Four RTX PRO 6000 Blackwell cards, 384GB of VRAM between them, in one cabinet.
You can’t order parts needed to liquid cool this card in India. We imported the blocks, then took each card apart to fit them ourselves. Somewhere on that board sits a small HDMI connector. If it snaps during the fit, there is no replacement part, and a card that cost over ₹10 lakh is gone, so we had to be most careful while doing it, FOUR TIMES.
More than half the loop came from overseas: fittings, radiators, quick disconnects.
Each card can pull up to 600 watts, and the CPU up to 350 more. In our stress test, under Linux with every component pinned at full load, the CPU held at 80 to 85°C and the GPUs stayed under 70°C. For four 600-watt cards in a single cabinet, those are numbers we were happy to see!
We’re proud to have played a part in DeepMindSecure’s work, and Saurabh’s take on the finished machine is the one we’ll remember.
“In India, building a liquid cooled system that I asked you to build is an impossible task. So I was very impressed with what you guys came up with. How do you assemble it? How do you package it? How do you ship it? I mean, it’s like out of the world.”
Under the glass
DeepMindSecure relies on this machine to train and run their speech models in-house:
What’s to come
With 384GB of VRAM under their own roof, DeepMindSecure’s training runs on their schedule. No rental queue, no monthly meter.
It’s already paying off in places nobody planned for. When a friend needed practice papers for students, Saurabh ran google’s AI model on an RTX PRO 6000 and generated nine lakh NEET questions and nine lakh JEE questions in one week, in English and Hindi, at three difficulty levels.
Next up is every Indian language, and the edge: speech models small enough to run inside a car, a kiosk or a phone, with nothing sent anywhere. That’s the 20-watt work, and it gets trained here.
“In five years, we are definitely going to cover all the languages and offer these services at a price point which nobody can offer and in an accuracy which nobody can beat,” said Saurabh.
His advice to other AI founders is the same call he made in July 2025.
“If you are an AI company, please, please, please own compute… A worker of the future is basically a set of electrons which are getting tokenized, running inside a compute in a GPU, and getting tokens out. So you have to own the infrastructure.”
A worker of the future is basically a set of electrons which are getting tokenized, running inside a compute in a GPU…
Saurabh Vajpai, Founder and CTO at DeepMindSecure
Check out DeepMindSecure to see their voice AI work across Indian languages and dialects.
For more on what this workstation can do, check out our video on it.
Planning an AI build? Configure yours at themvp.in, or book a 30-minute spec call with our team: [spec call link]. Every system ships with a 3-year doorstep warranty across 220+ cities. Toll-free: 1800 309 2944.
