OpenAI and the US AI chip unicorn Cerebras jointly announced the deployment of a total of 750 megawatts of Cerebras wafer-scale systems, building the world's largest high-speed AI inference platform. This collaboration will be implemented in stages starting from 2026, with full operation expected by 2028, and the transaction value exceeds 10 billion US dollars (about 69.7 billion RMB), marking that large model manufacturers are accelerating to move away from traditional GPU architectures.

Cerebras' chips are known as "giants" - each chip integrates 4 trillion transistors, with an area equivalent to hundreds of conventional GPUs. Its core advantage lies in integrating computing, memory, and bandwidth on a single wafer-level silicon chip, completely bypassing the latency and energy consumption bottlenecks caused by multi-chip interconnection. According to OpenAI's calculations, when running large models, the response speed of the Cerebras system can reach 15 times that of GPU-based solutions. For AI applications that pursue millisecond-level interactive experiences, this is not just a performance improvement, but a qualitative change in experience.

Notably, Sam Altman, CEO of OpenAI, is also an early individual investor in Cerebras. Their relationship dates back to 2017 - at that time, OpenAI had just been established, and it had already discussed the possibility of cooperating with this chip company founded the same year. Court documents show that OpenAI has long sought more efficient and cost-effective alternatives to NVIDIA. In the past year, it has jointly developed custom chips with Broadcom and purchased AMD's new MI450 accelerators. Now adding Cerebras, it shows its determination to build a diverse computing infrastructure base.

Andrew Feldman, CEO of Cerebras, revealed that both sides started formal negotiations in the fall of 2023 and finalized the cooperation intention before Thanksgiving. The driving force behind this decision is the unprecedented demand for "extremely fast computing." Sachin Katti, OpenAI's infrastructure director, said: "Computing power directly determines our revenue potential. In the past two years, computing power has doubled every year, and revenue growth has risen in tandem." Engineer feedback shows that existing hardware still struggles with high-load tasks such as programming assistance, prompting the company to accelerate the introduction of Cerebras solutions.

In terms of capital, Cerebras has also experienced a surge in valuation. According to The Wall Street Journal, the company is negotiating to raise $1 billion at a valuation of $22 billion, which is nearly three times the previous valuation of $8.1 billion. Although it had submitted an IPO application in 2024 and then withdrew it, with now having major clients such as OpenAI, Meta, IBM, and Abu Dhabi G42, its commercialization path has become clear. So far, Cerebras has raised a total of $1.8 billion in funding, not including the new funds in this round.

This collaboration is not only about the fate of the two companies, but also reflects a deep transformation in AI infrastructure: when large models enter the stage of large-scale commercialization, inference efficiency has become the core lever of user experience and commercial monetization. Although NVIDIA still dominates the ecosystem, heterogeneous approaches such as wafer-level integration and custom ASICs are being heavily invested in by top players. Future AI competition may no longer be a comparison of model parameters, but rather who can give an answer the moment it is spoken out.