Former Intel czar’s AI plans for India include data centres within every 100 km radius

Raja Koduri, founder of Mihira AI and former Intel chief architect

Semiconductor chips and artificial intelligence have hogged the limelight in the tech world of late. And it is not every day you can find someone who sits at the intersection of these two technologies. Raja Koduri is one of the few who do.

After flourishing as a career chip designer which peaked when he became Intel’s chief architect, Koduri left the chip major earlier this year. Now, he has founded an artificial intelligence startup – Mihira AI – that seeks to lower the cost of creating visual spectacles on the silver screen, among other things.

Koduri was connected to the movie world even before his startup. He has been an investor and advisor in Makuta, a company that won the national award for visual effects in 2012 for Telugu fantasy film Eega. Its director, SS Rajamouli, who received worldwide acclaim for RRR last year, is Koduri’s cousin.

But movies are not where the money is, Koduri told Moneycontrol in an interview. So his ambitions stretch beyond movies. He wants to help India’s remote areas benefit from AI with a data centre set up within every 100-km radius. He wants to figure out a way to develop AI without relying on Nvidia’s chips. Edited excerpts:

You started an AI company after exiting Intel. What is Mihira AI all about?

We want to harness the power of AI to make every artist a super-artist. If we can train thousands of artists across India on the new tools, we can create visual spectacles that are 10 times bigger than Game of Thrones. And do it within reasonable budgets. Avatar cost around $400 million. The challenge is to make it in $10 million with the help of AI.

How much of the tech development will be in India?

The majority of our workforce will be in India, but we have offices in America and in Singapore as well. Over 80 percent of our work workforce will be in India. We want to establish facilities even in remote small towns of India. But to do that, you need computing infrastructure, data centre infrastructure. So, we're also investing into data centres in India.

At every less than 100 km radius, we need a huge supercomputer that services all kinds of AI needs. And that's what I've been pushing with the government of India. If India aspires to be the number one country in technology, we need to have the best infrastructure for computing in the entire world.

Won’t setting up that many data centres cost hundreds of billions of dollars? Is it really feasible for India?

We can make it happen. With all the manufacturing in India, we can bring the cost down to a few billion dollars. If you don’t have a goal in mind, you don’t create technology. Today, it might cost, say, $300 billion. You need to then ask what needs to be done to bring it down to a few tens of billions. The computing power inside an iPhone is what supercomputers had in 2005 and cost millions of dollars. Now, it sits inside my pocket and costs $1,000. Humans can be very creative if the right targets and goals are set.

India doesn’t have a lot of money to offer. How do we attract the best chip and AI engineers to come back home?

I don't think money gets the talent. I think the collective passion to build products from India is the key. And there has to be an ecosystem which is the reason why the US is where it is. You might be a chip designer in Silicon Valley and you can walk across the street to talk to your customer. The good thing is that it’s happening in India. The momentum is there. It has taken everyone decades to build out an ecosystem — be it the US, China, Taiwan or Japan. I think India’s time has come.

Most chip design startups in India end up as service companies rather than building intellectual property-based products. How can this be corrected?

We need to aspire to do fundamental IP development in India, not just services. I think it will certainly happen. How do you create a new IP? You do it because you are solving somebody’s problem. Now, to be able to know what your problem is and build a product for that, I need access to you. So, that end-user perspective is needed in India.

A lot of people will come and say, ‘I will help you build a graphics processing unit (GPU).’ They need to say, ‘Hey, I will build the GPU for you.’ That’s the key difference. We need startups to be building for the end user, be it in gaming, AI, cell phones or other domains. I have no doubt it will happen in India.

Are you doing angel investments in chip startups?

I have been talking to several and already advising half a dozen. There’s one called Infinipack that has recently started and I am encouraging them to build a world-class packaging design, and not packaging services, company from India.

Since India has got a big commitment for a chip packaging plant, there’s this idea floating around that advanced packaging will deliver more efficiency in chips after the death of Moore’s Law. Do you agree?

Have you watched Game of Thrones? It has a famous dialogue — ‘What is dead may never die’. Moore’s Law has been dead for a while. The essence of Moore’s Law was that the performance of the chip will double at the same price in every two years. We kept shrinking the size of the chip and doubling the performance.

There’s nothing today whose performance will double next year. If you are lucky, its cost won't increase. But the cost of the iPhone has been increasing every year.

Will advanced packaging double the performance? It’s yes and no. It’s not as straightforward as some PowerPoint presentations want you to believe. All of the advanced packaging is really expensive today. This whole idea of chiplets and packaging will reduce cost and continue Moore's Law — all of this propaganda — can be traced back to a set of slides I made in 2018. That caused all this confusion and chaos. You can blame it on me.

Are you going to create a foundational large-language model at Mihira AI?

We will look at whether we should build a foundational model for India. Once we have enough data centre infrastructure here, that is an easy decision to train a model. Creating a foundational model today is about a lot of computing power and money. There is no fundamental innovation. If we can have a 1000-GPU data centre in India, we can build a foundational model here. The infrastructure is important.

There’s a long wait time for Nvidia GPUs which power OpenAI’s ChatGPT. Is there a workaround available?

We are focused on building an infrastructure with non-Nvidia GPUs — AMD, Intel, Tenstorrent and other GPUs. We want to work on figuring out and building IP and optimising those GPUs. So that's one of our ambitions. We aren't there yet. We are working on it. But when we meet you next time, we'll show you progress.

How much investment will all of it take? Are you raising funds?

Our goal for this platform is to invest several hundreds of millions of dollars. About half of that amount will be invested in India. We have raised enough money for now. If there’s a strategic investor who can help us with networks, not financing, then we can engage.

As Mihira AI will work with movie studios, is there a chance you might raise money from an HBO, Netflix or even a studio in India? Can we expect you to be working with your cousin Rajamouli any time soon?

We are working with top creators, including Rajamouli. But our goal is not just movies. We love movies, but that is not where the money is. Movies are where the fun is. The money is in manufacturing, the money is in healthcare and the money is in sports.

Do you know how much Apple spends on each of their iPhone launches? That money is much more than a lot of movie budgets. India has a huge workforce that needs to be trained, but doing so inside a chip fab or even a factory is very expensive. This can be done with AI-based virtual reality.

Where are you on the debate about AI killing jobs?

In the future, you will have to do your own electricity repairs and plumbing. They will be sitting in a nice resort and guiding you virtually how to do it. The soft-knowledge people like us will be at the bottom of the food chain in the post-AI era. The hard physical labour knowledge people will be at the top. That's my prediction of what's going to happen. It's controversial, but that's what I feel.

ChatGPT is smarter than any of us. So just learning a whole bunch of books by heart is not valuable anymore.