Initiate at Buy: A Specialty Approach to an Inflecting Market

研报英文原文证据摘录

Initiate at Buy: A Specialty Approach to an Inflecting Market

TD Cowen Cerebras Systems, Class A

Global Research June 8, 2026

Executive Summary of Our Initiation

We initiate coverage of Cerebras Systems with a Buy rating and $275 PT. Cerebras'

purpose-built Wafer-Scale Engines position the company to address what we see as

the critical use-case in modern AI deployments: inference speed at scale. By

pioneering wafer-scale computing, Cerebras believes it can offer materially lower

latency inference compared to GPU-based systems. With key validating wins, most

notably a >$20B cloud services deal with OpenAI, we believe the size of the

opportunity in front of Cerebras makes risk/reward attractive, acknowledging the risk

associated with scaling to the ambitions Cerebras has.

Differentiated Wafer-Scale Technology: Cerebras’ core innovation is the Wafer-

Scale Engine (WSE), a single 300mm wafer integrated into one massive AI processor.

Compute and memory (SRAM) are co-located on this wafer, enabling the spatial

architecture that requires less back and forth between compute cores and memory,

thereby reducing latency in efficiently compiled processors when compared to the

traditional register-based architecture. The WSE's spatial dataflow allows (requires)

each operation to occur in a known order, limiting flexibility but eliminating

significant overhead like hardware scheduling. A spatial architecture "flows"

calculations across thousands of processing cores, eliminating the off-chip data

transfers and memory bottlenecks that limit GPU clusters.

Cerebras’ latest WSE-3 (~900,000 cores, 44GB on-chip SRAM, ~21 petabytes/s

memory bandwidth) can store large model weights entirely on-chip or across a few

wafers, achieving ultra-high memory bandwidth and near-linear scaling for big