What Ola founder Bhavish Aggarwal’s AI chip ‘moonshot’ means

New Delhi: If Bhavish Aggarwal, founder of now-listed Ola Electric, is known to dream big, his newest ambition is audacious by even his standards: to build “India’s first AI (artificial intelligence) chips” in four years. Such a plan would thrust Aggarwal, who is likened to Elon Musk for his public persona, into an industry with only half a dozen global companies, which have developed competence over decades after investing hundreds of billions of dollars.

At a launch event—‘Sankalp 2024’—in Bengaluru on Thursday, Ola showcased what it called its “own” chip design that will power its foundational model development and its own data centres, and will be used in its own vehicles. The said chip will “challenge state-of-the-art AI chip performance by 2028” the company said on stage—showing an upward-curve graph that offered no details or metrics.

Developing the capability to build chips is not impossible but requires the world’s most advanced technology. Even India's government has been making a concerted effort to bring semiconductor manufacturing into the country, but the effort failed the first time, and the progress has been slow. The $10 billion-plus semiconductor production-linked incentives have led to the announcement of only one greenfield chip fabrication plant, or ‘fab’, so far—by Tata Electronics and Taiwan’s Powerchip Semiconductor Manufacturing Co. (PSMC). The $10 billion project is expected to take at least three years to become operational and will make chips of ‘older’ nodes.

Read more | Bhavish Aggarwal announces Ola’s first AI chip for India, set for 2026 launch: Everything announced at Sankalp event

For designing chips, something even more challenging, reference designs are built to ensure maximum compatibility with a wide base of developers around the world—who ensure compatibility across global software platforms. Typically, this takes extensive cross-development with global firms and is not done in silos.

“Chipmaking is a complex process and requires dedicated fabs in the country—as well as adopting ASML’s photolithography machines,” said Jaspreet Bindra, founder of tech consultancy firm The Tech Whisperer, referring to the Dutch firm that is the only company in the world with the technology. “This typically takes a very long time, even a decade, to establish,” he said.

Ola, a mobility company that is yet to show any profit, will have to outperform the likes of the US’ Nvidia, Intel, Qualcomm and Advanced Micro Devices (AMD)—which have three to five decades of experience in making chips that power not just AI, but the world’s greatest supercomputers.

Ola announced a partnership with Canada-headquartered Untether AI to build its chips. Untether has no proprietary chip technology and uses Nvidia’s reference design. 

Also read | Ola to roll out portable dark stores for quick commerce, launches first electric motorcycle

Kashyap Kompella, a tech analyst and founder of consultancy firm RPA2AI, said Ola needs to provide details. “The company has made a lot of claims, some of which include taking on the world’s greatest technology companies with the deepest of pockets, to build something that is not in their domain of expertise,” he said. “It would be very important for Ola to furnish proof of their claims on public benchmarks such as MLCommons, if they are to be taken seriously.”

Ola founder Aggarwal termed his AI ambitions a “moonshot project.” The first of Ola’s chips, called ‘Bodhi 1’ with “best-in-class power efficiency”, will be ready by 2026, said the company. By 2028, the company aims to introduce ‘Bodhi 2’, offering “best-in-class performance for AI training, inference and fine-tuning”, it said.

According to Ola, ‘Bodhi 2’ will support 10 trillion data-parameter AI models and will be “scalable to exascale computing”. Such powerful chips are capable of supporting the best supercomputer is in the world today. Nvidia and AMD design such chips and has taken them decades of research and development (R&D).

“It’s great to see ambition, but it’s important to realize that to make chips like Nvidia, Intel or even AMD, it will take billions of dollars and decades just to get the capability—let alone four years,” said Bindra. "It’s important to get key technical details on what these claims are—and how such development will play out. To create something of the class of a TSMC chip, from absolute zero, is next to impossible. We’ll be curious to get more details.”

Ola linked its ambition to the country’s interests. “India has 1.25 lakh chip designers, but not one meaningful chip comes out of India. We’re going to change that,” claimed Sambit Sahu, vice-president of silicon design at Ola. There are “only one or two companies in the world that are doing well in AI, and they are not accepting India’s needs—whether it is security, data privacy, or our power needs,” he said.

Kompella of RPA2AI said Ola should shun hyperbole as its shareholder base extends well beyond its venture partners, and into the public sphere.