OpenAI is undergoing a significant hardware strategy revision to optimize computing resources and reduce costs. The company plans to incorporate AMD's MI300 series chips while continuing to utilize NVIDIA's (NVDA, Financial) GPUs. OpenAI is also collaborating with Broadcom and TSMC to produce custom AI chips by 2026.
A team of approximately 20 engineers, including experts from Google's Tensor processor project, has been established for chip development. Despite the heavy reliance on NVIDIA’s GPUs, which hold over 80% of the market share, chip shortages and rising costs have led OpenAI to seek alternatives.
By introducing AMD's MI300 chips, OpenAI aims to ensure high-performance computing and mitigate supply risks. In addition to AMD collaboration, OpenAI is working with Broadcom on custom chips for AI inference tasks and partnering with TSMC to secure manufacturing capabilities.
Initially considering building a network of chip manufacturing plants, OpenAI has temporarily shelved this plan due to cost and time constraints. The focus is now on internal chip design and partnerships with industry leaders like Broadcom and TSMC to ensure stable chip supply.
Training AI models like ChatGPT is expensive, with OpenAI expecting a $5 billion loss this year against a $3.7 billion revenue. Computing costs, including hardware, power, and cloud services, are the largest expenses for the company.
To address these challenges, OpenAI is exploring internal hardware development and external alternatives, a strategy similar to tech giants like Amazon, Meta, Google, and Microsoft. These companies aim to reduce costs and secure AI hardware access through custom chip development, although OpenAI might require significant investment to become truly competitive.
In October, OpenAI completed a new funding round of $6.6 billion, led by Thrive Capital, with participation from Microsoft, NVIDIA, Altimeter Capital, Fidelity, SoftBank, and Abu Dhabi's MGX Investment Company. The post-investment valuation reached $157 billion, with plans to strengthen computational resources and expand AI research capabilities.
