Facings the growing demand for real-time computing, AI giant OpenAI is significantly expanding its hardware footprint. On Wednesday, OpenAI officially announced a multi-year deep partnership with chipmaker Cerebras, aiming to break through existing bottlenecks in AI response speed through dedicated chip technology.
Layout of Computing Resources at Hundred Billion Scale
According to insiders and Reuters, the total value of this deal exceeds 10 billion US dollars. Under the agreement, from this year to 2028, Cerebras will provide OpenAI up to 750 megawatts of computing power.
Cerebras was founded more than a decade ago and has emerged in the AI boom after ChatGPT with its non-GPU architecture system specifically designed for AI. Its core advantage lies in performance: Cerebras claims that its system is significantly faster than traditional GPU solutions represented by Nvidia when processing AI applications.
Core Objective: Achieving "Real-Time Inference"
The core of this collaboration lies in improving user experience continuity. OpenAI stated that these systems will greatly shorten the previously time-consuming response process. Sachin Katti, a representative from OpenAI, pointed out that the company's strategy is to build an "elastic system portfolio," with Cerebras serving as a specialized low-latency inference solution, providing users with a more natural and faster interaction foundation.
Andrew Feldman, CEO of Cerebras, gave a vivid evaluation: "Just as broadband changed the internet, real-time inference will change artificial intelligence."
Strong Capital Ties and Expansion Ambitions
This major collaboration is not accidental. OpenAI CEO Sam Altman was previously an investor in Cerebras, and OpenAI once considered acquiring the company.
Currently, although Cerebras has repeatedly postponed its 2024 IPO plan, it has shown strong performance in the primary market. On Tuesday, it was reported that the company is negotiating a new round of financing worth 1 billion US dollars at a valuation of 22 billion US dollars
