Cerebras CEO Andrew Feldman describes the company's dinner-plate sized wafer-scale chip, which he claims delivers roughly 50× faster inference than NVIDIA GPUs on AI workloads. The company built a PyTorch compiler to avoid NVIDIA's software lock-in and has won customers including Meta, IBM, and Mistral, though 80% of revenue currently comes from G42 in the UAE.
Chapters
Sure, we chose a radically different approach. We built a chip the size of a dinner plate while everybody else is building chips the size of a postage stamp. This allowed us to keep information on chip, to move it less. As a result, use less power. As a result, do work more quickly. And as a result, we're order 50 times faster than NVIDIA GPUs at the key portion of the AI work inference. And this is how we deliver AI to customers through inference.
And with models getting increasingly bigger, with companies talking about agentic AI and more and more demand, how are these chips optimized to handle those workloads?
These chips are designed exactly for those workloads. It's in agentic workloads and reasoning workloads. In which lots of compute is required in order to generate better answers. And that's sort of the, the beauty of agentic work and reasoning work is we get really good answers out of the AI. But the only problem is in traditional work, they're slow. And so with GPUs, when you're doing agentic work or reasoning work, you have to wait for answers. And we don't like to wait. And so what you've seen is that a massive move to faster AI, and that's what we're at the heart of and we're the leader of.
I want to talk about, obviously, your positioning versus NVIDIA, because NVIDIA has grown to a very strong position, not just from its GPUs and the rack scale solutions, but incredibly important to that story has been software and the lock-in we've seen from NVIDIA customers into NVIDIA's ecosystem. How do you begin to eat away at that dominance?
Sure. I think there are a number of ways. The first is that the ML practitioner likes to write in a language called PyTorch, and this is above any vendor-specific language. And with our 400 software engineers, we were able to build a compiler that takes directly this, this PyTorch language and compiles it for our hardware. And so we can bypass the this lock-in completely. And so it hasn't been as big a challenge for us as maybe others, but we've been able to work around it and to deliver to customers a seamless experience where they can go directly from what they want to write in to the high performance delivered by our machines.
I want to talk about the scaling plans here because I've spoken to a lot of the neo-cloud vendors here. And I've asked them, what chips are you buying right now? What are you building on? And it's NVIDIA. It's NVIDIA. It's NVIDIA. There's no other name in town. So when it comes to scaling, if you've got some of the big neo clouds right now building for NVIDIA, how do you manage to get to a scale again where you're offering an alternative?
Well, recently we've announced wins at Meta, at IBM. We are a major supplier for Mistral, and here in France, the European leader in, in AI. We have hundreds of customers around the world. And so we've, we've seen extraordinary growth and exponential demand for, for fast inference.
Andrew, you're a company that's hurtling towards an IPO at some point. And I think there's a lot of people watching who are going to just want to know a little bit more about your company's trajectory at this point, but also a bit about, you know, your business. And one of the big things points that's been, been brought up is the fact that more than 80% of your revenue is reliant on a customer, G42. How are you planning to diversify your revenue stream at this point?
It's a good question. I think we do have a very large customer in the Middle East, and G42, as you know, is the national champion for the UAE. We build for them part of a cloud, and on that cloud are dozens of of different companies from the ecosystem in the UAE. They're leading universities, they're leading industrial enterprises. G42 itself is 9 different companies. And so it is a cloud for many customers at once. Now, we are also building businesses in Europe, building businesses in the USA, in parts of Asia. We don't sell in China, but in Singapore, in Korea, in Japan. And so over time, I think we will have a much more diversified customer base.
And in terms of the timing of the IPO, again, you know, I know that's still up in the air, but you're going to market at a time where you're still unprofitable. Understandably, it's a CapEx-intensive business, but there's a lot in the public market now who are very discerning about the type of company that they're willing to invest in. Unprofitable companies haven't done so well. For a while. Why do you think now is the right time to be going public?
I think there's a yawning appetite for AI companies and particularly for pure-play AI companies and for alternatives to NVIDIA. I think we've seen with CoreWeave's IPO a tremendous reception for rapid growth and for participation in this revolution.
And in terms of timing, I mean, some of the difficulties with coming to market have been resolved now with CFIUS, etc. Do you have an idea of timeline at this point?
We do. Unfortunately, I can't share it with you. I'd love to, but that's— those aren't the rules. Those aren't the rules. And so I think there were some hurdles. Those are in the rearview mirror. The new administration was helpful in that. And so we're excited. The business is growing and it will happen when it happens.
