OpenAI is collaborating with Broadcom and TSMC to create its first in-house chip tailored to support its artificial intelligence systems, while also adding AMD chips to its lineup alongside Nvidia’s. This approach aims to meet the company’s expanding infrastructure needs, sources told Reuters.
As the company behind ChatGPT, OpenAI has been exploring options to diversify its chip supply and cut costs. Initially, the company considered establishing its own network of chip manufacturing factories, or “foundries,” but later abandoned this plan due to the high cost and time involved.
Now, OpenAI’s focus has shifted to designing its own in-house chip through a mix of industry partnerships and internal efforts. This strategy mirrors similar moves by Amazon, Meta, Google, and Microsoft, which are diversifying their chip supplies and managing costs by leveraging in-house resources. Being one of the largest buyers of chips, OpenAI’s decision to work with multiple suppliers and develop its own custom chip could have significant effects on the tech industry.
Following this news, Broadcom’s stock rose over 4.5%, while AMD saw an increase of 3.7%. OpenAI, AMD, and TSMC declined to comment, while Broadcom did not respond to requests for a statement.
To support its AI models and operations, OpenAI relies on extensive computing power. Currently, the company primarily uses Nvidia’s GPUs to train AI models and for inference, applying AI models to make decisions based on new data. While demand is currently highest for training chips, experts predict that inference chips will see growing demand as AI applications continue to expand. Sources reported that OpenAI has been working with Broadcom for months on an AI chip specifically geared for inference tasks.
Broadcom’s involvement includes assisting in fine-tuning chip designs and enhancing data transfer capabilities, crucial for AI systems where tens of thousands of interconnected chips work together. OpenAI is also considering whether to develop or acquire additional chip components and may bring in more partners, according to sources.
A team of about 20 engineers at OpenAI, including former Google engineers experienced in building Tensor Processing Units (TPUs), is leading the chip design effort. Through Broadcom, OpenAI has also secured manufacturing capacity with TSMC, targeting the release of its custom chip by 2026, though the timeline may be subject to change.
With Nvidia’s GPUs currently holding over 80% of the market, shortages and rising costs have driven other major tech companies like Microsoft, Meta, and OpenAI to look for alternatives. OpenAI’s upcoming use of AMD’s new MI300X chips through Microsoft’s Azure platform further highlights AMD’s move to capture some of Nvidia’s market share. AMD expects AI chip sales to reach $4.5 billion in 2024 following the MI300X’s launch at the end of 2023.
Training AI models and running services like ChatGPT come with steep costs, and OpenAI expects a $5 billion loss this year on $3.7 billion in revenue. Expenses related to hardware, electricity, and cloud services for large-scale data processing make up OpenAI’s biggest costs, prompting efforts to streamline usage and diversify suppliers.
Despite expanding its chip sources, OpenAI has been cautious about hiring Nvidia talent to maintain a positive relationship with the company, which it relies on for the latest generation of Nvidia’s Blackwell chips, sources added.
