No Surprises in Cerebras's First Steps as A Public Company

Research evidence excerpt

No Surprises in Cerebras's First Steps as A Public Company

handles decode (sequential). Management expects the AWS relationship to contribute in

2027 (not 2026) and framed disaggregated decode-as-a-service to GPU owners as a

broader, repeatable opportunity beyond AWS.

• Model momentum: Cerebras launched enterprise trials of Kimi K2.6 - the first trillion-

parameter model served on Cerebras, ~1,000 tokens/sec per Artificial Analysis - and

Gemma 4 31B, which it says runs an order of magnitude faster on its hardware. A live

demo showed a trillion-parameter Kimi model finishing a prompt in 21 seconds versus 4

min 37 sec on a leading GPU endpoint (management's understanding: a B300), ~13x

faster. Cerebras’s ability to handle this model should allay concerns that cropped up

around the WSE’s ability to deal with newer larger models due to its more modest

memory capacities (SRAM) vs other solutions using HBM.

• Supply chain: Management pointed out that Cerebras sidesteps the industry's three

binding constraints - it uses on-wafer SRAM rather than HBM (in short supply); does not

use TSMC's CoWoS packaging; and runs at the less-contended 5nm node rather than 3nm

(where wafer supply is extremely tight). Also, management highlighted that CS-3 systems

are manufactured exclusively in the U.S (mitigating some geopolitical risk) and that

Cerebras expanded its partnership with Flex and added Sanmina as a 2nd global contract

manufacturer (in addition to Cerebras having historically worked with Rocket).

• Data center capacity is the gating factor: Management was explicit in noting that neither

demand nor wafer supply is their primary constraint. Rather data centers space is

currently CBRS’s primary bottleneck. New facilities are coming online across the U.S.,