Cerebras Unveils CS-4, OpenAI and AMD Partnerships to Accelerate AI Inference

Key Points

• Cerebras unveiled its CS-4 AI system, which is expected to deliver up to twice the token-generation speed of CS-3, six times higher system-level performance and up to 10 times more tokens per watt in select applications.
• The company highlighted partnerships with OpenAI and AMD to accelerate inference, including an AMD-GPU/Cerebras split in which GPUs handle model prefill and Cerebras handles token decoding.
• Cerebras is expanding its data-center footprint, with 600 megawatts of power online or under contract by the end of next year, while targeting AI-agent, design, coding and cybersecurity applications that benefit from lower latency.

Cerebras Systems NASDAQ: CBRS outlined a product roadmap centered on faster AI inference, expanded data-center capacity and new partnerships with OpenAI, Arista Networks and Advanced Micro Devices during a company event led by CEO and Co-Founder Andrew Feldman.

Feldman said the company recently went public and is seeking to position its wafer-scale computing technology as an infrastructure platform for real-time AI applications. He argued that inference speed has become a product-level consideration rather than solely a technical benchmark, particularly as AI shifts toward interactive tools and autonomous agents.

“In AI, speed is productivity,” Feldman said, asserting that faster response times allow users to run more workloads and address more complex tasks without the traditional trade-off between model intelligence and latency.

OpenAI Discusses Ultrafast Service Tier

Feldman highlighted OpenAI’s recently announced GPT-5.6 Sol Ultrafast mode, which he said runs OpenAI’s most intelligent models at up to 14 times the speed of its standard offering for select customers. He said Cerebras hardware powers the service and presented a comparison based on Humanity’s Last Exam, a graduate-level reasoning benchmark.

According to Feldman, the Cerebras-powered system completed the full 2,500-question benchmark in 11 hours, 11 minutes and 26 seconds, while the comparison system required more than three days.

Thibault Sottiaux, a member of the technical staff at OpenAI, said the company’s strategy has been to build leading models and serve them at scale, with increasing focus on agents. He said ultrafast inference reduces the need to choose between a smaller, lower-latency model and a larger model that takes longer to respond.

Sottiaux said OpenAI would like Ultrafast to become the default experience over time, though he stressed that the company is still early in deploying it. He said OpenAI reserves some Ultrafast capacity for incidents, security matters, major internal projects and research efforts, while also allocating capacity for customers through its API.

He added that ChatGPT has surpassed 1 billion users, while the company’s more sophisticated agentic workflows remain used by a smaller but rapidly growing segment. Feldman said OpenAI has 15 million weekly users of Codex and ChatGPT agents, a figure Sottiaux referenced as recently published by OpenAI.

Data-Center Buildout and Infrastructure Partners

Feldman said Cerebras has data centers operating or being developed across North America and Europe, naming locations including Santa Clara, Toronto, Dallas, Minneapolis, Montreal, Oklahoma City, Alabama, Lyon, France, Norway and Mikkeli, Finland. He said the company has brought on 600 megawatts of power that is online or under contract for delivery by the end of next year.

Jayshree Ullal, CEO of Arista, said AI infrastructure requires coordinated networking alongside compute capacity. She described the growing importance of networking across scale-up, scale-out and geographically distributed “scale-across” deployments.

Ullal said that AI workloads differ from traditional cloud traffic and require networks that can handle varying traffic patterns, model types and agent workloads. She also said bandwidth requirements could continue to increase rapidly, citing potential networking speeds of 3.2 terabits and 6.4 terabits in the coming years.

Feldman also discussed a disaggregated inference approach with Mark Papermaster, AMD’s executive vice president and chief technology officer. Under that approach, GPUs handle the “prefill” stage of inference, in which an AI model processes an input context, while Cerebras systems handle “decode,” the token-by-token generation phase.

Feldman said this combination could provide inference that is 10 times faster than GPUs and five times more throughput than Cerebras alone. Papermaster said the partnership combines GPU strengths in parallel computation with Cerebras’ low-latency decode capabilities.

CS-4 System Targets Higher Performance

Sean Lie, Cerebras’ CTO and co-founder, introduced the company’s next-generation CS-4 system, which is in early access and is expected to be generally available later this quarter.

Lie said CS-4 is designed around the company’s Nexus rack-scale platform and includes three wafer-scale engines in a single system. He said the platform will offer up to twice the token-generation speed of the current CS-3 generation, six times higher system-level performance, three times more density per rack and up to 10 times more tokens per watt in certain solutions.

According to Lie, the system uses the company’s WSE-3 Turbo chip, with each wafer providing 43 petabytes per second of memory bandwidth. He said the architecture is intended to reduce bottlenecks associated with moving model weights across multiple chips and to support larger models through lower-latency wafer-to-wafer connections.

• CS-4 is expected to provide up to 30 times faster token generation than GPU solutions, according to Cerebras.
• The system is designed with twice the I/O bandwidth per wafer and lower network latency than the prior generation.
• Cerebras said its modular rack architecture is intended to reduce components by 50% and shorten data-center deployment times from days to hours.

Lie said the Nexus architecture is intended to support future CS-5 and CS-6 systems. Cerebras’ roadmap calls for doubling speed annually and delivering solutions with up to 20 times higher throughput by 2027.

Focus on AI Agents and Time-Sensitive Applications

Jessica Liu, Cerebras’ senior vice president of product, said fast inference can improve the usefulness of AI agents by allowing more model calls, planning loops and tool use within the same time budget. She said slower systems can force users to choose smaller models or wait longer for more capable agents.

Speakers from Figma, Cognition and CrowdStrike described applications for faster model inference in design, software engineering and cybersecurity. Figma AI Research Product Lead Pavi Bhatter said the company’s Design Agent, currently in beta, uses Cerebras hardware as part of its effort to make design workflows more interactive. Cognition’s Silas Alberti said the company has deployed coding models on Cerebras and reported faster end-to-end task completion in certain use cases.

Keith Coley, CrowdStrike’s vice president of engineering, said cybersecurity decisions can be highly time-sensitive and described the company’s partnership with Cerebras as a way to allow more AI-based inspection within acceptable response windows.

About Cerebras Systems (NASDAQ:CBRS)

Cerebras Systems is a technology company focused on building artificial intelligence infrastructure, including hardware and software designed to accelerate deep learning and large-scale AI workloads. The company is best known for its wafer-scale processor architecture, which is intended to provide high-performance compute for training and inference applications.

In addition to its AI chips, Cerebras offers systems and related software tools that support researchers and enterprises working with machine learning models.

This instant news alert was generated by narrative science technology and financial data from MarketBeat in order to provide readers with the fastest reporting and unbiased coverage. Please send any questions or comments about this story to contact@marketbeat.com.

Continue following MarketBeat

Add MarketBeat as your preferred source on Google to see our latest stories in your feed.

Add As Preferred Source