OpenAI has inked another major computing power agreement, partnering with Cerebras to deploy 750 megawatts of wafer-scale chip capacity. On January 14th, OpenAI and Cerebras announced a three-year agreement, committing to procure up to 750 megawatts of computing power, all utilizing Cerebras' wafer-scale chips. Citing sources familiar with the matter, The Wall Street Journal reported the total value of the deal exceeds $100 billion. Cerebras' chips are designed to run AI models and generate responses faster than competitors, thereby meeting OpenAI's demand for rapid computational capabilities. According to the agreement, the deployment will be carried out in phases starting in 2026. Cerebras' AI chips focus on inference tasks, with their unique speed advantage stemming from integrating massive computing power, memory, and bandwidth onto a single, giant chip. This eliminates the bottlenecks in traditional hardware that cause inference speeds to slow down. OpenAI CEO Sam Altman is a personal investor in Cerebras, and the two companies had previously explored collaboration opportunities back in 2017. This deal could potentially bring a significant valuation jump for Cerebras. Reports citing informed sources indicate the company is in talks to raise $1 billion at a valuation of $22 billion, which would nearly triple its previous valuation.
Surge in Demand for Inference Chips
OpenAI is accelerating its efforts to acquire more data center capacity to prepare for its next phase of growth. According to OpenAI's Katti, the company began considering collaboration with Cerebras after receiving feedback from engineers who desired chips capable of running AI applications, particularly programming-related ones, more quickly. The two companies commenced negotiations last autumn and signed a letter of intent before Thanksgiving. Cerebras CEO Andrew Feldman demonstrated a series of presentations to media, showing that chatbots powered by Cerebras chips could provide responses to users faster than those using competitors' hardware. Feldman stated:
What is currently driving the market is immense demand for rapid computing power.
Chip startups focusing on inference—the process of running trained AI models to generate responses—are experiencing high demand, as AI companies race to acquire cutting-edge technology capable of providing fast, cost-effective computational power. In December, NVIDIA signed a $20 billion licensing agreement with Groq, granting it access to chips developed by that startup, which are also designed for handling such tasks.
Seeking Alternatives to NVIDIA
OpenAI is actively seeking cheaper and more efficient alternatives to NVIDIA's chips. Last year, the company announced a collaboration with Broadcom to develop custom chips and separately signed an agreement to use AMD's new MI450 chips. This partnership with Cerebras is a continuation of OpenAI's strategy to diversify its chip supply. The company is preparing to enter its next growth phase and urgently needs to secure more data center capacity. OpenAI generated approximately $13 billion in revenue last year, but its new cloud contracts with Oracle, Microsoft, and Amazon total nearly $600 billion, raising investor concerns about its ability to pay. Altman has stated that these contracts are executed in phases and that the company will pay for them through future revenue growth.
Cerebras Valuation Could Rise Significantly
Cerebras, founded about a decade ago, has not been a prominent player in the semiconductor market until now. When the company filed for an IPO in 2024, it disclosed that the majority of its revenue came from a single client: Abu Dhabi-based G42. It was mentioned that Cerebras withdrew its IPO plans in October last year, having previously completed a $1.1 billion private financing round at an $8.1 billion valuation. Feldman noted that Cerebras has since signed new contracts with IBM and Meta. According to PitchBook data, the company had previously raised a cumulative $1.8 billion from investors including Benchmark, the UAE company G42, Fidelity Investments, and Atreides Management.
Comments