tools
Cerebras CS-3
Data-center AI infrastructure built around a wafer-scale processor for large-scale training and inference.
In inglese
Cerebras CS-3 is a data-center system for organizations running large AI and HPC workloads. It combines the WSE-3 processor with power delivery, cooling, networking, and system management, and can connect with other CS-3 systems for larger deployments.
It is intended for enterprise, research, hyperscale, and sovereign-AI environments rather than individual users. Cerebras does not publish CS-3 purchase pricing on the pricing page; deployment requires contacting the company.
Features
- WSE-3 processor with 900,000 AI-optimized cores
- 44 GB of on-chip SRAM
- 21 petabytes per second of on-chip memory bandwidth
- Supports models up to 24 trillion parameters on one logical device
- Connects through 12 standard 100 Gigabit Ethernet links
- Scales from 1 to 2,048 CS-3 systems
- Supports large-scale AI training and inference
Use cases
- Deploy large language models in enterprise data centers
- Run high-speed inference for conversational AI
- Train and serve large AI models for research
- Build sovereign AI infrastructure
- Scale AI and HPC workloads across connected systems
Pros
Cons
Latest updates
- Flex and Cerebras Expand Partnership to Scale American Manufacturing of Cerebras AI Supercomputers
New manufacturing lines in Milpitas, California will support an anticipated 7x increase in production of Cerebras CS-3 systems.
Security
Pricing
- Prices checked
- 2026-09-24
Developer Tier
- $0.35/M tokens
- $0.75/M tokens
- $0.99/M tokens
- $1.49/M tokens
- Self-serve pay-as-you-go
- OPENAI GPT OSS 120B: ~3000 tokens/s
- QWEN Qwen 3.8 27B: ~1,850 tokens/s