OpenAI's New Chip Strategy

By Krystal Hu, Fanny Potkin and Stephen Nellis

(Reuters) – OpenAI is collaborating with Broadcom (NASDAQ:AVGO) and TSMC to develop its first in-house chip for AI systems, while also integrating AMD (NASDAQ:AMD) chips alongside Nvidia (NASDAQ:NVDA) to address growing infrastructure demands, sources reported.

The company, known for ChatGPT, has explored various options to diversify its chip supply and cut costs. Initially, OpenAI considered building its own factories (known as "foundries") but has now paused those plans due to high costs and the lengthy time required for development. Instead, it will concentrate on in-house chip design efforts, according to anonymous sources.

This strategy reveals OpenAI's approach to leveraging partnerships and a combination of internal and external solutions to secure chip supplies and manage costs, akin to larger competitors like Amazon (NASDAQ:AMZN), Meta (NASDAQ:META), Google, and Microsoft (NASDAQ:MSFT). OpenAI's decision to collaborate with various chipmakers while creating a custom chip is expected to influence the broader tech sector.

Following this announcement, Broadcom's stock rose by more than 4.5%, while AMD's shares increased by 3.7% within the same trading day.

OpenAI, AMD, and TSMC refrained from commenting, and Broadcom did not respond quickly to a request for comment.

As a major consumer of chips, OpenAI requires significant computing resources to train and operate its AI systems, particularly by utilizing Nvidia's graphics processing units (GPUs). These chips are vital for both training models and inference, helping to apply AI to new data for predictions or decisions.

Sources indicate that OpenAI has been working with Broadcom on an AI chip focusing on inference. Currently, the demand leans more toward training chips, but experts anticipate the need for inference chips will grow as more AI applications emerge.

Broadcom assists companies like Alphabet (NASDAQ:GOOGL)'s Google in refining chip designs and supplies essential parts enhancing data transfer on these chips, which is critical for AI systems involving numerous interconnected chips.

OpenAI is still evaluating whether to create or acquire other components for its chip design and may seek additional partners, according to two sources. The organization has formed a chip development team of about 20 engineers, led by veterans from Google who created Tensor Processing Units (TPUs).

Through Broadcom, OpenAI has secured manufacturing capacities with TSMC to produce its first custom chip by 2026, though the timeline is subject to change.

Currently, Nvidia commands over 80% market share in GPUs, but increased costs and shortages have prompted major clients such as Microsoft, Meta, and OpenAI to explore alternative in-house or external solutions.

OpenAI's intention to utilize AMD chips via Microsoft's Azure highlights AMD's effort to capture market share from Nvidia, with projected AI chip sales of $4.5 billion in 2024 following the MI300X chip's launch scheduled for Q4 2023.

With the costs of training AI models and operating services like ChatGPT soaring, OpenAI projects a $5 billion loss this year against $3.7 billion in revenue. Compute expenses constitute the bulk of these costs, pushing the company to optimize usage and diversify suppliers.

In maintaining good relations with Nvidia, OpenAI is careful about recruiting talent from the company, which it aims to continue collaborating with, especially for access to new Blackwell chips, as per sources. Nvidia declined to comment.