Cerebras earnings and AI chip market dynamics (with Ed Ludlow)

Andrew Feldman explains Cerebras' margin contraction as a strategic decision to rent back gear to customers to meet extraordinary demand. He highlights Cerebras' architectural advantages: no HBM dependency, no CoWoS, and use of 5nm node instead of 3nm, insulating them from key supply chain constraints. He sees a heterogeneous AI chip market with room for GPUs, ASICs, and pioneering architectures.

Yields

implicit

Oil

Metals

USD

Bloomberg 7.0
Financial Media
Ed Ludlow 2.5

Cerebras earnings and AI chip market dynamics (with Ed Ludlow)

ndx

Feldman is bullish on AI compute demand but Cerebras stock dropped 17%, indicating market skepticism about margins and competition from ASICs.

rut

No specific mention of small-cap stocks. Cerebras is a mid-cap AI company.

The stock is down 17% despite sales outlook beating estimates. Is the sequential margin decline about ramping output for two big customers?

Ed Ludlow
Andrew Feldman

We put forward a plan and we're ahead of it. Record revenues of $191M, up 92% YoY. Cloud business up 167% YoY. We beat margin consensus substantially. The margin impact in Q2 and Q3 is from renting back gear to customers to keep up with demand - about 10-15 points.

We did that to keep our customers close and ensure we could keep up with extraordinary demand for fast inference.
Have IPO proceeds allowed you to move more quickly in ramping capacity?

Ed Ludlow
Andrew Feldman

Capacity is the largest constraint - data centers. We've significantly increased our pipeline. Announced a partnership with Bell Canada for 120 megawatts delivered in 2027. We're pursuing data centers across the US, Canada, Europe, and the Middle East.
How difficult is it to get planning approval, concrete, labor, and build data centers?

Ed Ludlow
Andrew Feldman

The AI market moves at blistering speed but we're constrained by data centers which move at the speed of real estate. This problem confronts everyone - hyperscalers, neo clouds, new generation clouds.
Cerebras doesn't use traditional off-chip HBM. How insulated are you from the memory bottleneck?

Ed Ludlow
Andrew Feldman

Because of our wafer-scale architecture, we don't use HBM. There's a global shortage, it's extremely expensive with long lead times. We also don't use CoWoS or 3nm capacity. We're at 5nm. Our architecture delivers the fastest inference in the world while avoiding main supply chain constraints.
Can you actually deliver compute faster than others?

Ed Ludlow
Andrew Feldman

We signed our contract with OpenAI on December 24th and had them in full production on February 1st. That's unheard of.
How do you interpret frontier labs going to custom ASICs alongside other compute options?

Ed Ludlow
Andrew Feldman

This market is enormous and will be met with a heterogeneous collection of architectures. It will not consolidate around GPUs. There will be ASICs from hyperscalers, labs, and pioneering architectures like Cerebras - all taking big bites of this enormous market.

AI makes tractable for compute much of the world around us, which wasn't the case prior to AI.
Can we see you come back to capital markets to move quicker?

Ed Ludlow
Andrew Feldman

We have more than $9 billion on the balance sheet. We're always scanning both equity and debt markets for ways to accelerate growth.
How are you judging success? What milestone should the market watch?

Ed Ludlow
Andrew Feldman

When you lay out an aggressive plan and crush it - that's how you feel good. We gave full year guidance 10 gross margin points above consensus. We need to continue executing and beating high bars.

SignalTube

markets at a glance