Cerebras Systems

Full briefing

Cerebras Systems, Inc. is a private semiconductor startup headquartered in Sunnyvale, California, founded in 2016 by Andrew Feldman and a team of veterans from SeaMicro (acquired by AMD) and other semiconductor companies. Cerebras pioneered the concept of wafer-scale integration for AI compute — using an entire silicon wafer as a single monolithic chip rather than dicing it into individual dies — to eliminate the chip-to-chip communication overhead that constrains GPU cluster performance on large neural network models. The WSE-3 (Wafer Scale Engine 3), announced in March 2024, is the third generation of Cerebras's wafer-scale chip and the largest semiconductor device ever manufactured for commercial sale. Fabricated by TSMC on a 5nm process, the WSE-3 occupies a full 300mm wafer — 46,225 mm² — and integrates 900,000 AI-optimized processing cores, 44 GB of on-chip SRAM (with 21 PB/s of aggregate memory bandwidth), and 7 trillion transistors. For comparison, NVIDIA's H100 die is approximately 814 mm², about 57 times smaller. The key advantage of the WSE-3 is that all 900,000 cores communicate through on-chip interconnects at SRAM speeds, eliminating the HBM memory bandwidth bottleneck and NVLink/InfiniBand latency that limits how quickly a distributed GPU cluster can process each training step. The practical performance implication is dramatic for specific workload types. For large language model training and inference on models that fit within the WSE-3's 44 GB SRAM, Cerebras has demonstrated training speeds measured in tokens per second that are 10–100× faster than comparable H100 clusters, with fully deterministic latency — a property valuable for scientific computing applications that require reproducible timings. The CS-3 system (the rack unit that houses one WSE-3) delivers approximately 125 petaFLOPS of sparse compute. Multiple CS-3 units can be scaled using MemoryX (which off-loads weight storage to external DRAM) and SwarmX (which provides a cluster interconnect fabric), enabling training of models with hundreds of billions of parameters on WSE-3 cluster configurations. Cerebras's target market segments are deliberately differentiated from mainstream AI cloud providers. The company has focused on pharmaceutical and biotechnology customers (where protein folding, molecular dynamics, and drug discovery models benefit from its architecture), national laboratory and government research customers (NCAR, Argonne National Laboratory, and several Department of Energy sites have deployed Cerebras systems), and defense and intelligence community customers through its Cerebras Government division. In September 2024, Cerebras partnered with the UAE's G42 to provide AI inference capacity, and in late 2024 filed confidentially for a Nasdaq IPO. The TSMC dependency is Cerebras's single most critical supply chain constraint. Each WSE-3 requires an entire 300mm TSMC 5nm wafer — an unusually large allocation compared to standard chip orders — and the manufacturing yield management for a wafer-scale device is significantly more complex than for standard chips, since defects that would destroy a small die merely become non-functional cores on a wafer-scale design (addressed through redundant core design). TSMC's 5nm capacity is shared with Apple, NVIDIA, AMD, and other high-priority customers, making Cerebras's wafer allocation dependent on its standing in TSMC's customer priority queue. Export control regulations do not directly affect Cerebras's current product line (it sells to US and allied-country customers), but any future TSMC capacity constraints driven by geopolitical factors would directly impact its production.