'AI Doesn't Care Where It Goes To School'

'No one manufactures intelligence at the moment.'
'This is a concept that your IT industry understands.'
'What you need is infrastructure. Everything else can then be taken care of.'

Jensen Huang, founder and chief executive officer, Nvidia, says India has a big role in exporting intelligence.

In a media briefing he speaks of the partnership with Reliance, diversifying its relationship beyond TSMC, and making artificial intelligence (AI) accessible to the masses.

Can you comment on the partnership with Reliance?

We will do three things. We're going to build the infrastructure for AI compute. It is going to be on a fairly large scale.

Two, we're going to start an innovation centre so that Nvidia's technology can be used by Reliance engineers to create an AI platform.

The third is to create applications that Reliance could offer to consumers in India.

There are talks that ties between Nvidia and TSMC are getting strained and Nvidia may work with others.

TSMC and Nvidia are busy building the Blackwell System right now and it is in full production.

Without TSMC's agility and scale and its management team working so closely with us and without two and a half decades of friendship, building these Blackwell Systems, which includes seven different chips, would have been impossible.

So TSMC is an extraordinary company and we're working together. We're scaling up Blackwell Systems hard, and we'll ship in Q4.

It is important for us to have as much diversity and redundancy ... we have options with Samsung and Intel as well and we will continue to evaluate.

How will Nvidia make sure that AI is accessible to the masses, especially in a country like India, which is price-sensitive?

The price of tokens, which is AI, has decreased 100 times in a year. The reason for that is because we've invented so many new algorithms for speeding up the processing of inference.

Two, there are many technologies that have been invented for distillation or small-language models.

Third, Nvidia's road map is fast, from Ampere to Hopper to Blackwell ... we're introducing new technologies and every year we double the performance, which is the same as reducing the cost by half or by a third.

On manufacturing in India.

Our best way to partner India is to partner Infosys, Tata Consultancy Services, Wipro and other information-technology companies here, so that we can revolutionise everything else.

Our best way to help India build manufacturing capabilities is to help it first be a manufacturer of AI ... remember other countries have been manufacturing chips and systems for a long time. This is a new time.

No one manufactures intelligence at the moment.

This is a concept that your IT industry understands. What you need is infrastructure. Everything else can then be taken care of.

Concern about data privacy and security with AI?

Many companies move data outside and it is dangerous. We have too much intellectual property encoded inside our data. It's important to build AI on-prem.

Energy situation and advances of AI systems

The reason we introduce new generations of architecture is that we could reduce the energy used. Blackwell uses half or one-third energy.

The reason for that is it does the job two or three times faster. Two, AI doesn't care where it goes to school.

It doesn't have to go to school right here in the middle of Mumbai.

Mumbai does not have excess power. But India has excess energy; it's just not in Mumbai. The near-term growth of AI is so fast that it is putting pressure on data centres.

Regulation and AI

There are no simple answers. The best way to make AI safe is to develop new technologies.

A year ago everybody was talking about hallucination. Now AI hallucinates a lot less.

Feature Presentation: Aslam Hunani/Rediff.com