Synopsis
AI startups in India are now shifting their focus to move beyond applications and AI wrappers to build solutions in frontier deeptech areas, writes Swathi Moorthy.Cases in point include ZenetiQ, a scientific large language model (LLM); HumanTronik, a personalised LLM that can mimic human brain for enterprise use case; and Oru’el, a startup that predicts graphics processing unit (GPU) failures using physics-based architecture.
In addition, Indian AI startups are moving up the tech stack. Sarvam has been garnering attention since the launch of its new AI models early this year across vision, language and voice, and is reportedly raising $300 million at a $1.5 billion valuation.
Murf AI, which started as a creator-focused platform for voice generation and dubbing, launched its text-to-speech foundational model, Falcon, this year and is working on a speech-to-text model. Maya Research, a voice AI startup, is building a foundational model for speech from ground up.
This marks a significant shift from as recently as last year when the narrative was primarily around AI applications and Indian AI models had yet to take off.
Priyanshu Ghosh, cofounder, Oru’el, said the past couple of years were about understanding what could be done with AI, while now it is all about adding real value, with innovation kicking in. This includes using science and AI to solve real-world problems.
Playing in the frontier tech
Founded by Ghosh, Ajith Sai Chekka and Nihith Tallapalli, Oru’el offers GPU reliability solutions through physics-based models.
Chekka said that the startup first built expertise in GPU physics and drew parallels from the lithium-ion battery industry, where proven methods existed for predicting component degradation and downtime. Oru’el, he said, developed proprietary physics-informed models, which integrate laws of physics such as thermodynamics with AI, trained on real GPU telemetry data, including performance metrics, health indicators and operational parameters captured from live data centre environments.
ZeneteiQ founder Sashikumaar Ganesan is developing a scientific foundation model for engineering use cases such as product design for the automotive and aerospace sectors. But the current LLMs are primarily for linguistic applications, whereas the scientific models need to be predictive. “The key idea behind the scientific foundation model we are building is this: can we bridge the gap by fusing the reasoning of LLMs with predictive capability?” he said.
The company is achieving this by training the model on scientific datasets and utilising specialised tools for data generation. Backed by funding from the IndiaAI mission, it is leveraging Tensor Processing Units for the training process.
Monish Darda, cofounder, HumanTronik, said if in the next five to 10 years, companies are going to run fairly autonomously, with AI as part of their DNA, “the question was what do I need to do to leverage that and build a foundation today?”
“The answer I got was that human creativity is going to be scarce and more important. So we started exploring hyper personalised language models, models that are tuned to how individual brains of leaders and experts work, and how we can deploy these models in the enterprise set-up,” Darda said.
Investors are keeping an eye on this. Pratyush Choudhury, cofounder and general partner, Activate AI, said the venture capital firm is working with founders building specialist foundation models for domains such as healthcare, materials and physics.
Challenges
While the challenges of compute memory have eased, Ganesan pointed out that human capital remains a primary hurdle. “The real scarcity isn’t in understanding Transformer theory, but in the engineering required for the parallel training of these models,” he said.
“The expertise needed to manage massive workloads across distributed clusters, optimising data, model and pipeline parallelism, is extremely rare in India. Even with a dedicated research team in place, the lack of depth in distributed systems talent in India is a significant bottleneck. Oru’el’s Tallapalli said that for products like theirs, there is not enough of a testbed environment and that creates a challenge. In addition, given that their clients are data centres, building trust as young founders is also a challenge, he added. Others said that while Indian startups are rising up the ladder in terms of innovation, the country may have a long way to go in terms of catching up with China and the US.
