Convequity on X
We think the gross margin gap between Cerebras ($CBRS) and NVIDIA ($NVDA) says little about which company delivers compute more cheaply. Cerebras reported a 14% gross margin in Q2 2026, against roughly 75% for NVIDIA. That is not a clean comparison, because Cerebras's commercial agreements include customer warrants. A customer warrant gives a buyer of Cerebras compute the right to buy Cerebras shares at an agreed price, adding potential equity upside to the compute it purchases. Cerebras records the value of these incentives over time as a reduction in revenue. Its cost of delivering the compute stays the same, so the full amount comes out of gross profit and its reported gross margin falls too. Add the warrants and a few other items back and Cerebras's Q2 margin was 41%, so the reported gap tells us little about what each company spends to deliver equivalent compute. Cerebras could challenge NVIDIA with attractive economics on selected workloads without matching every capability of NVIDIA's platform. In our view NVIDIA's stronger answer is Feynman reaching customers on schedule with the 3D-stacked SRAM we expect. Stacking more fast memory close to the compute would reduce the need to fetch data from slower external memory, helping NVIDIA narrow the architectural advantage behind Cerebras's fast inference. NVIDIA's adoption of Groq technology could ease some of this pressure, but we believe Groq's current architecture remains weaker than Cerebras's for low-latency inference. The longer Feynman takes to arrive at scale, the more time Cerebras has to win customers.
