Cerebras
Cerebras is an American artificial-intelligence computing company that develops wafer-scale processors, AI supercomputers, and cloud services.
Last updated August 31, 2026
AI crawler reads
2requests
1companies- Anthropic2
- Profile
- 2
- Controversies
- 0
- JSON · MD
- 0
- First read
- Nov 4, 2026
- Latest read
- Nov 4, 2026
From this site's server log
Overview
Cerebras Systems is a Sunnyvale, California-based technology company focused on high-performance computing for artificial intelligence and deep-learning workloads. Founded in 2015 by a group of former SeaMicro executives and engineers, the company pursued an unconventional alternative to the dominant model of assembling large numbers of separate graphics processors. Its central idea is wafer-scale integration: instead of cutting a silicon wafer into many conventional chips, Cerebras uses most or all of the wafer as a single very large processor, linking its processing elements with an on-chip communication fabric. The company’s principal hardware product is the Wafer Scale Engine, or WSE. The architecture is designed to reduce the communication delays, memory movement, and software complexity associated with distributing large AI models across clusters of discrete accelerators. Cerebras processors use large arrays of AI-optimized cores and substantial on-chip memory, allowing some models to be trained or served within a single system rather than partitioned across many conventional processors. The approach requires specialized manufacturing, packaging, cooling, power delivery, defect management, and software technologies. Taiwan Semiconductor Manufacturing Company manufactures the company’s processors. Cerebras announced its first Wafer Scale Engine and CS-1 system in 2019. The CS-1 was followed by the CS-2, based on the WSE-2, and then the WSE-3 and CS-3 generation introduced in 2024. The company has marketed these systems to research institutions, national laboratories, pharmaceutical and life-sciences organizations, energy companies, and AI developers. Applications have included large-language-model training, scientific simulation, drug and peptide discovery, genomics, nuclear-stockpile analysis, and computational fluid dynamics. Alongside equipment sales, Cerebras has developed clustered installations and cloud-based access. Its systems can be connected into wafer-scale clusters, while its AI training and inference cloud services allow customers to use Cerebras computing without purchasing or operating the hardware themselves. The company has also built or supported dedicated data-center capacity and has worked with cloud, model, and infrastructure partners including G42, the Mohamed bin Zayed University of Artificial Intelligence, Dell Technologies, Mistral AI, Perplexity AI, and other organizations identified in company and reference materials. Cerebras positions its technology around throughput, low latency, simplified scaling, and predictable performance for large AI models. Its hardware competes with accelerator products and platforms from Nvidia, AMD, Intel, and Broadcom, while its cloud services compete with hyperscale and specialized providers such as Amazon Web Services, Microsoft Azure, Google Cloud, Oracle Cloud, and CoreWeave. The wafer-scale approach also has trade-offs: systems are physically large, power-intensive, expensive, and dependent on complex packaging and cooling. Cerebras therefore operates in a specialized segment of the AI infrastructure market rather than as a general-purpose consumer electronics brand. The company’s public profile grew through scientific-computing achievements, collaborations with national laboratories, and the deployment of large AI clusters. It has also used its cloud platform to emphasize fast inference for models from organizations such as DeepSeek, Mistral, Perplexity, and Alibaba. Cerebras remains an active private technology company in the supplied reference material, with operations in the United States, Canada, India, and other international markets.
History
Cerebras Systems was founded in 2015 by Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie, and Jean-Philippe Fricker, who had previously worked together at SeaMicro. The founders believed that conventional GPU clusters created substantial communication and memory bottlenecks for advanced AI workloads. Their proposed solution was to make a processor at wafer scale, retaining a very large connected array of compute cores and memory on one silicon substrate. This strategy created engineering problems beyond ordinary chip design. Cerebras had to develop methods for cooling and powering a very large processor, package the wafer onto a system board, route data around manufacturing defects, and construct equipment capable of handling and securing the wafer without damaging it. The company spent several years addressing these challenges and reported producing a working product in 2019. In August 2019, Cerebras announced the WSE-1 and CS-1. The rack-mounted CS-1 used a single wafer-scale processor with hundreds of thousands of processing cores and was aimed at AI and scientific workloads. Early users and collaborators included pharmaceutical companies, universities, research centers, and national laboratories. Their applications included drug discovery, genomics, computational fluid dynamics, and research related to COVID-19. Cerebras expanded internationally during the following years, including an office in Japan and later an office in Bangalore, India. In 2021 it introduced the CS-2, powered by the WSE-2 and manufactured using TSMC’s 7-nanometer process. The system increased the number of cores, memory capacity, and bandwidth and was promoted for training very large models. Customers and research users included TotalEnergies, nference, the National Center for Supercomputing Applications, and the Leibniz Supercomputing Centre. In 2022, the company demonstrated the use of a single CS-2 for training models with billions of parameters and described systems for connecting multiple CS-2 units into wafer-scale clusters. It also introduced Andromeda, a 16-system AI supercomputer. National-laboratory users applied Cerebras systems to nuclear-stockpile stewardship, scientific simulation, and infectious-disease modeling. A Computer History Museum exhibition that year highlighted the WSE-2 as an example of unusually large-scale chip design. Cerebras then broadened its business from individual systems to integrated infrastructure and cloud access. Its relationship with G42 led to the Condor Galaxy series of AI supercomputers, while collaborations with cloud and enterprise infrastructure providers supported wider access to its technology. In 2023, Cerebras worked with the Mohamed bin Zayed University of Artificial Intelligence and G42 subsidiary Inception on the Jais language model. In March 2024, Cerebras announced the WSE-3 and CS-3. The WSE-3 used a 5-nanometer process and was described as containing four trillion transistors and 900,000 AI-optimized cores. The company also expanded its inference offering, seeking to compete not only as a hardware supplier but as a provider of directly accessible AI compute. Subsequent collaborations involved Dell Technologies, Mayo Clinic, model developers, and AI search and conversational-service companies. Cerebras’s current strategy combines specialized processors, complete AI supercomputers, clustered installations, and training and inference clouds. Its differentiation is based on keeping more computation and data movement within a single wafer-scale system. The model can simplify some distributed-training workloads and deliver high throughput, but it also entails substantial power, cooling, manufacturing, and system-cost requirements. The company therefore serves demanding enterprise, scientific, government, and AI-platform customers rather than the mass consumer market.
- 2025
Healthcare and model-platform collaborations expanded
[1]Cerebras announced work with Mayo Clinic and supported model and AI-service providers including DeepSeek, Mistral, and Perplexity.
- 2024
WSE-3 and CS-3 introduced
[1]Cerebras announced its 5-nanometer WSE-3 architecture and the CS-3 system.
- 2024
AI inference service launched
[1]The company expanded into hosted AI inference, giving customers cloud access to Cerebras processors.
- 2023
Jais language model collaboration
[1]Cerebras collaborated with the Mohamed bin Zayed University of Artificial Intelligence and Inception on Jais.
- 2022
Andromeda unveiled
[1]The company announced a 16-system AI supercomputer based on WSE-2 processors.
- 2021
CS-2 launched
[1]The CS-2 brought the WSE-2 architecture, with substantially more cores, memory, and bandwidth than the first generation.
- 2019
WSE-1 and CS-1 announced
[1]Cerebras introduced its first wafer-scale processor and CS-1 AI computing system.
- 2015
Company founded
[1]Cerebras Systems was founded by Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie, and Jean-Philippe Fricker.
Products and positioning
Specialized AI infrastructure provider differentiated by wafer-scale integration, very large on-chip processing capacity, high memory bandwidth, and direct cloud access to its processors.
Wafer-scale AI processor · 2019
WSE-1
Cerebras’s first announced Wafer Scale Engine, introduced in 2019 as the processor at the center of the CS-1 system. It established the company’s approach of using a wafer-scale array of interconnected processing elements rather than a conventional collection of separate accelerator chips.
AI supercomputer · 2021
CS-2
The CS-2 is a dedicated AI computing system based on the WSE-2. Introduced in 2021, it increased core count, on-chip SRAM, memory bandwidth, and fabric bandwidth and was designed to train and run very large neural-network models with less distributed-system complexity.
Wafer-scale AI processor · 2024
WSE-3
The WSE-3 is Cerebras’s third-generation wafer-scale processor architecture. Announced in 2024, it uses a 5-nanometer manufacturing process and is designed with a very large array of AI-optimized cores, extensive on-chip memory, and high-bandwidth internal communication.
AI supercomputer · 2024
CS-3
The CS-3 is the system built around the WSE-3. It targets large-model training, inference, and scientific workloads and packages Cerebras’s wafer-scale processor, memory, networking, cooling, and software into a data-center appliance.
Cloud AI service · 2024
Cerebras AI Inference Cloud
A hosted inference service that allows developers and enterprises to run supported AI models on Cerebras hardware without purchasing a CS-series system. Its market proposition emphasizes high token-generation throughput and low-latency serving.
Cloud AI service
Cerebras AI Training Cloud
A cloud-access model for using Cerebras computing resources to train AI systems. It extends the company’s business beyond direct hardware sales and supports customers that need wafer-scale compute without operating their own installation.
Marketing campaigns
- 2022The Biggest Chip in the World
United States
The Computer History Museum in Mountain View presented a WSE-2 display that highlighted Cerebras’s wafer-scale approach and the unusual physical size of its AI processor.
Outcome. The exhibition increased public visibility for wafer-scale computing as a distinct semiconductor design approach.
Brand decisions
2024
Expansion from hardware into cloud inferenceStrategyBackground
AI customers increasingly sought access to accelerators as a service rather than purchasing complete systems.
Action taken
Cerebras launched hosted inference and training offerings using its own data-center and partner infrastructure.
Aftermath
The company broadened its addressable market to developers and organizations that could consume AI compute through APIs.
2021
Move to WSE-2 and CS-2Generation changeBackground
The first generation demonstrated the wafer-scale concept but required greater capacity for increasingly large AI models.
Action taken
Cerebras introduced the WSE-2 and CS-2 with expanded cores, memory, and internal bandwidth.
Aftermath
The platform became suitable for broader research, scientific, and enterprise deployments.
2019
Commitment to wafer-scale AI computingProduct launchBackground
Cerebras concluded that conventional accelerator clusters created communication and memory bottlenecks for some large AI workloads.
Action taken
The company commercialized the WSE-1 and CS-1, making a wafer-scale processor the foundation of its product strategy.
Aftermath
The launch established Cerebras as a specialist competitor in AI accelerators and high-performance computing.
Recent events
Reported
Report: OpenAI to power GPT-6.1 Sol "Ultrafast" mode with Nvidia GPUs instead of Cerebras hardware[3]
Research firm SemiAnalysis posted on X on Wednesday (2026-09-30) that OpenAI will power its "Ultrafast" mode for the GPT-6.1 Sol AI model with Nvidia graphics processing units instead of Cerebras hardware. The report was cited as a key driver of Cerebras shares falling nearly 20% during the week to their lowest level since the company's May IPO. On Friday (2026-10-02), OpenAI CEO Sam Altman responded to speculation about the partnership, writing that "Cerebras is a close partner," and Cerebras shares rose almost 3% in extended trading.
Weekdays and relative dates follow the original report; their exact calendar dates have not been verified.
Reported
Cerebras launches CS-4 rack-scale AI inference system[4]
Cerebras Systems unveiled the Cerebras CS-4, a rack-scale AI system built on three Wafer Scale Engine 3 Turbo processors, at its SUPERNOVA event in San Francisco. The company claims it is the industry's fastest AI accelerator, with 750 PFLOPs of AI compute and over 4,400 tokens per second per user on GPT-OSS-120B, up to 30 times faster than GPU alternatives. Shipments are expected to begin in the third quarter of 2026, with a small number of customers currently sampling the system.
- Unconfirmed
Cerebras launches CS-4 rack-scale AI system[5]
Cerebras Systems launched its new rack-scale AI system CS-4, which the company says can output 30 times more tokens per second than traditional GPU servers in single-user scenarios.
Reported
Cerebras Systems launches CS-4 rack-scale AI system[6]
Cerebras Systems launched its next-generation rack-scale AI system CS-4 on August 18, claiming it is the industry's fastest AI accelerator with inference speeds up to 30 times that of GPU solutions. CS-4 offers 750 PFLOPs of AI compute, 129.6 PB/s memory bandwidth and 7.2 Tbps I/O throughput, built on TSMC 5nm process with system-on-wafer packaging and SRAM instead of HBM.
Reported
Cerebras launches CS-4 server rack with WSE-3 Turbo chip[7]
Cerebras Systems announced the CS-4, a new server rack powered by three of its large chips and based on its Nexus server architecture, which the company says will speed AI chatbot queries. The system is available in the third quarter and the chips are fabricated with TSMC's 5-nanometer process.
- Unconfirmed
OpenAI launches GPT-5.6 Sol UltraFast preview for enterprise users[8]
OpenAI launched a preview of GPT-5.6 Sol UltraFast for enterprise users, a mode that runs up to 14 times faster than the standard mode and generates up to 750 tokens per second, with compute support from Cerebras. Access is currently application-reviewed, and OpenAI plans to expand availability via API based on enterprise workload feedback.
Sources13 sources
- 1Cerebras Systemsen.wikipedia.org
- 2Cerebras official websitecerebras.ai
- 3据报道:OpenAI 改用 Nvidia GPU 而非 Cerebras 硬件驱动 GPT-6.1 Sol「Ultrafast」模式cnbc.com
- 4Cerebras发布CS-4机架级AI推理系统163.com
- 5Cerebras发布CS-4机架式AI系统emsnow.com
- 6Cerebras Systems发布CS-4机架级AI系统163.com
- 7Cerebras发布搭载WSE-3 Turbo芯片的CS-4服务器机架economictimes.indiatimes.com
- 8OpenAI面向企业用户推出GPT-5.6 Sol UltraFast超快模式预览chinaz.com
- 9比GPU方案快30倍！Cerebras发布CS-4：AI算力达750 PFLOPS！163.com
- 10Cerebras launches new server chip and system designed to speed AI chatbotsthestar.com.my
- 11Cerebras launches new server chip, system designed to speed AI chatbotsbusiness-standard.com
- 12Cerebras推出CS-4系统强势对标英伟达 称其推理速度可达GPU服务器的30倍 _ 东方财富网finance.eastmoney.com
- 13Cerebras推出CS-4系统强势对标英伟达，称其推理速度可达GPU服务器的30倍163.com
Cite this profile: Cite the canonical profile. /brand-wiki/cerebras · Editorial policy · How profiles are compiled