Cerebras CEO Andrew Feldman Talks Earnings

**SPEAKER_1** (0:02)
Bloomberg Audio Studios, podcasts, radio, news.

**Ed Ludlow** (0:10)
Cerebras reported quarterly earnings for the first time since going public last month. Its sales outlook beat Wall Street estimates, but still disappointed investors, hoping to see the company carve out a bigger slice of the AI Dennison Center market. Right now, shares down 17.5%.
Its biggest drop in its quite short history is the public company. What's behind that? CEO Andrew Feldman is with us. Welcome back to Bloomberg Tech, Andrew. You know, there was a time where we would talk about the merits of top-to-tail server ownership, how owning all of the content of... Now we're going to talk about margin contraction, and we're going to talk about the stock being down 17%.
That, to me, is kind of the mismatch. The outlook on the sales side beat Wall Street estimates. I think a lot of people are trying to understand the sequential margin decline. And for me, this is about ramping output for two big customers. Is that true?

**Andrew Feldman** (1:03)
Yeah, I think what we did is we put forward a plan in the start of 26 We shared it with investors as we went public. And we're ahead of plan. We delivered record revenues of 191 million, up 92% year over year. And for our cloud business, it was up 167% year over year.
We beat margin consensus substantially. And then we guided for full year, the gross margins would be 10% better than planned. We also shared that in Q2 and Q3, we would go back to some of our customers and we would rent back year that we'd sold them to try and keep up with demand. And that would have a margin impact on the order of 10 or 15 points. We did that to keep our customers close to be sure we could keep up with their extraordinary demand for our product, for fast inference. And so that was the story.
On every metric we put out, we're ahead of plan.

**Ed Ludlow** (2:14)
Have the proceeds from the IPO actually allowed you to move more quickly in ramping up capacity?

**Andrew Feldman** (2:20)
Yeah, I think capacity is the largest constraint right now for everyone. Data centers are. And we've significantly increased our ability and our pipeline for data centers, which is now very large. We announced a data center partnership with Bell Canada for 120 megawatts. That will be delivered in 2027 We are pursuing data centers across the US., in Canada, in Europe, in the Middle East.
The vast resources that we have now at our disposal give us tremendous advantage in the pursuit of this, the limiting factor data centers.

**Ed Ludlow** (3:11)
What you're talking about matter of factly is buildings, not necessarily the compute, right? It's not what you guys are offering. How difficult right now is it to get moving in America or other markets to get planning approval, get the concrete, get the labor, get the thing built?

**Andrew Feldman** (3:28)
That's the irony of this market, that the AI market is moving at blistering speed and we are being constrained by data centers which move with the speed of real estate. And so, that is a problem that is being confronted by everybody in the category, by the hyperscalers, by the Neo clouds, by the new generation clouds. Everybody is confronting this similar problem.

**Ed Ludlow** (3:57)
Andrew, Cerebras does not rely on traditional off-chip HBM. Would you just explain that, the basics of the technology, but how insulated are you from the memory bottleneck that others are experiencing?

**Andrew Feldman** (4:10)
Yeah, that's a really good point. Because of our innovative architecture, because of our way for scale approach, we don't use HBM.
HBM is a type of DRAM, and it's made by three companies, one of whom is reporting shortly, right? That's Micron, Hynix and Samsung. There's a global shortage. It's extremely expensive. Lead times are long and we don't use it. So we have a tremendous advantage there.
The other constraints in the supply chain for many are co-oss, which is a process inside of TSMC. Again, we don't use it. And the third is capacity at the three nanometer node, that space in TSMC's factory that makes three nanometer chips. Again, we don't use it. We're at the five nanometer node. So our architecture has allowed us to deliver the fastest inference in the world by order of magnitude while avoiding the main supply chain constraints faced by others in the field.

**Ed Ludlow** (5:16)
Can you say, hand on heart, not just winning business, but have you actually been able to go to a customer and say, we can get this compute online faster than others for those reasons you just outlined, and then actually gone and done it?