Hello and welcome to a very special edition of Startup Street. I'm Shi Mishra and I have with me Arati Ramanan. Not only are we celebrating 1,000 episodes of Startup Street, we're getting you all the action from Web Summit Qatar in Doha. The conference sees 25,000 plus attendees, 1,500 plus startups, and over 600 investors from across the globe, all of them gathered here today at the Doha Exhibition and Convention Center.
Absolutely, Shi, and what a way to kickstart our 1,000th episode. We are here not just at any other conference, but we're looking at the Gateway to the Future. We're at the epicenter of groundbreaking ideas, of brilliant innovations, and big billion-dollar deals. So if you're a startup and you want to witness the next big thing in AI, fintech, or digital transformation, you're at the right place.
Absolutely, and in fact, I caught up with a host of people today who are at the epicenter of this AI innovation. Take a look.
We're here at the Web Summit in Qatar and it's all about AI and innovation. And today we have with us Andrew Feldman of Cerebras Systems. Andrew, thank you so much for taking time out and being here with us at the CNBC TV18 studio in Doha.
Thank you so much for having me. What a fun event.
Absolutely, and you know, before we get into the nitty-gritty of everything, if you could sort of explain to us what it is you do at Cerebras.
Sure, we build big, fast computers whose job is to accelerate AI. AI has two parts: it has training, which is how we make AI, and inference, which is how we use AI. And we built a chip, and then a system, and the entire computer that makes both training and inference faster and easier and use less power.
Right, and you know, sort of to explain that to us a little bit further, you know, when we talk about chips, the first thing that comes to mind is NVIDIA, right, when it's AI chips. But if you could tell us, how are you doing this differently?
Sure, there are lots of ways to make chips. We chose a way that focused on the exact problem posed by AI. And the hard part about AI for a computer is the amount of data that needs to be moved. And so we built a very big chip. In fact, we built the biggest chip in the history of the computer industry, and that allows us to move data less often, to keep it on chip, to work on it faster, and to use less power.
And you know, that was what I was going to come to because when I was reading about Cerebras, the one thing I saw is about the size of the chip that you make, right? So you know, how much faster is it, how much more efficient is it, like what makes it different?
Sure, so the chip is about 56 times larger than any other chip made. So compared to an NVIDIA chip, it's 56 times larger. It's over 70 times faster in inference, used as a fraction of the power. So this one chip is faster, it uses less power, and it makes your AI experience more enjoyable.
So it's faster, it uses less power, and it's more enjoyable. And as we know, you now host DeepSeek on your platform, right? And it's making waves, it's what people are talking about. And the thing about DeepSeek is that it's so much cheaper than, you know, what a ChatGPT would do or how you would do that. So why is that the case?
So we host a whole range of models. We host open-source models like Llama from Meta, DeepSeek from DeepSeek, and we also host closed-source models for particular customers like Mistral and Perplexity and many others. And all of those are served to customers via our cloud. What makes it faster? It's a good question. By being a bigger chip, we have to move information less far, right? And that's what takes time in compute. And so these big chips move information more quickly, using less power. And as a result, all these different models, whether it's DeepSeek or Llama or Perplexity or Chat for Mistral, all of them provide a better user experience.
Right, and you know, one of the things that comes up is also about accuracy, right? Like for instance, on DeepSeek, there are certain questions that you can't sort of ask or you won't get an answer to. So how does someone circumvent that, you know, when so much information is being passed through?
