Cerebras unveils CS-4 with three wafers per rack at SUPERNOVA event

Cerebras unveiled the CS-4 at its SUPERNOVA event, a rack with three WSE-3 Turbo wafers and 750 petaflops of sparse FP16 compute. The company says it runs frontier models up to 30 times faster than GPU systems, a claim measured as tokens per second per user on one model. First shipments are due before the end of the third quarter, which matters for hardware revenue in Q3.

• Three WSE-3 Turbo wafers per rack; 750 petaflops sparse FP16
• Memory bandwidth 129.6 petabytes per second
• Claimed up to 30 times faster than GPU systems on gpt-oss-120b
• Wafer-to-wafer latency cut to two microseconds from five
• First shipments due before end of Q3 2026