Indian AI firms take up super hard stuff

Indian artificial intelligence (AI) startups are undergoing a transformative shift, moving beyond merely applying existing technologies to pioneering innovative solutions in advanced fields such as science and engineering, physics, and neuroscience. This evolution, as highlighted by Swathi Moorthy in The Times of India, signifies a substantial change in the landscape of AI development in the country, with startups now focusing on creating foundational technologies rather than just building on top of existing models, often referred to as AI wrappers.

Among the notable players in this emerging trend is ZenetiQ, which is developing a scientific large language model (LLM) aimed at engineering applications. Another example is HumanTronik, which has created a personalized LLM that mimics human brain functions for enterprise applications. Additionally, Oru’el is making strides by predicting graphics processing unit (GPU) failures through physics-based architectures, showcasing the diverse applications of AI in addressing complex challenges.

The ascent of Indian AI startups in this frontier technology space is further illustrated by Sarvam, which has gained attention for its new AI models launched earlier this year, spanning vision, language, and voice capabilities. Sarvam is reportedly in the process of raising $300 million at a valuation of $1.5 billion, indicating strong investor interest in the sector. Murf AI, initially a platform focused on voice generation and dubbing, has recently introduced its foundational text-to-speech model, Falcon, and is also developing a speech-to-text model. Similarly, Maya Research is building a foundational model for speech from the ground up, marking a significant departure from the previous year’s focus on applications.

Priyanshu Ghosh, cofounder of Oru’el, reflects on this shift by noting that the past few years were dedicated to understanding AI’s capabilities, while the current focus is on delivering real value through innovation. This includes employing scientific principles alongside AI to tackle real-world problems, which is a notable evolution in the approach taken by these startups.

Pioneering Frontier Technologies

Oru’el, co-founded by Ghosh, Ajith Sai Chekka, and Nihith Tallapalli, specializes in GPU reliability solutions using physics-informed models. Chekka explains that the company initially developed expertise in GPU physics, drawing parallels from the lithium-ion battery industry, where established methods exist for predicting component degradation. Oru’el’s proprietary models integrate fundamental laws of physics, such as thermodynamics, with AI, trained on real telemetry data from GPUs, including performance metrics and operational parameters collected from live data center environments.

ZenetiQ’s founder, Sashikumaar Ganesan, is also venturing into the scientific domain by developing a scientific foundation model tailored for engineering use cases, particularly in product design for the automotive and aerospace sectors. Unlike current LLMs that focus primarily on linguistic applications, Ganesan emphasizes the need for predictive capabilities in scientific models. His approach involves training the model on specialized scientific datasets and employing advanced tools for data generation, supported by funding from the IndiaAI mission and utilizing Tensor Processing Units for training.

HumanTronik’s cofounder, Monish Darda, envisions a future where companies operate autonomously, integrated with AI as a core component. He emphasizes the importance of human creativity and the need for hyper-personalized language models that are customized to the cognitive styles of individual leaders and experts, aiming to deploy these models effectively in enterprise settings.

Navigating Challenges in AI Development

Despite the promising advancements, Indian AI startups face several challenges. Ganesan highlights that while the issues related to compute memory have improved, the scarcity of human capital remains a significant obstacle. The expertise required for managing the parallel training of AI models across distributed systems is notably rare in India. This includes optimizing data, model, and pipeline parallelism, which are essential for handling massive workloads efficiently.

Tallapalli from Oru’el points out that a lack of adequate testbed environments poses challenges for startups like theirs, particularly when building trust with clients in data centers. Furthermore, while Indian startups are making strides in innovation, they still have a considerable distance to cover to catch up with established players in countries like China and the United States.

As the landscape of AI development continues to evolve, the focus on creating foundational technologies in India represents a significant step forward. The commitment to addressing complex problems through innovative solutions could position Indian AI startups as key players in the global technology arena, provided they can overcome the challenges they face in talent acquisition and infrastructure development.