OpenAI is advancing its artificial intelligence capabilities by partnering with Broadcom (AVGO, Financial) to develop a new AI chip for model inference processes, as revealed by informed sources. They are also in discussions with TSMC (TSM), the world's largest contract chip manufacturer. These plans are still in the early stages, as OpenAI has been exploring custom chip development for about a year. Broadcom and OpenAI have not commented on these reports.
In June, The Information reported on Broadcom’s potential collaboration with OpenAI for AI chip manufacturing. OpenAI is exploring various strategies to diversify chip supplies and reduce costs. Originally, they considered establishing their own chip fabrication network, but high costs and time constraints led them to refocus on internal chip design.
The strategic partnerships OpenAI is forming highlight an industry trend of leveraging both internal and external resources to secure chip supplies, similar to larger tech companies such as Amazon (AMZN), Meta (META), Google (GOOGL), and Microsoft (MSFT). As a major chip purchaser, OpenAI's initiative to develop custom chips while sourcing from different manufacturers could significantly impact the tech industry.
The company’s focus is not on graphics processing units (GPUs) dominated by Nvidia (NVDA), which are used for training AI models, but on developing specialized chips for running software and responding to user requests—a process known as inference. With the increasing complexity of tasks handled by AI models, analysts expect a rising demand for inference-supporting chips.
On Tuesday, Broadcom's stock closed up by 4.2% at $179.24. This followed a year-to-date increase of 54%. TSMC's stock also closed over 1% higher. Broadcom is a leading designer of application-specific integrated circuits, serving major clients like Google, Meta, and ByteDance.
Broadcom aids companies like Google in fine-tuning chip designs for manufacturing, crucial for AI systems that require thousands of interconnected chips working together. OpenAI has reportedly been collaborating with Broadcom for several months on its first inference-focused AI chip. Despite a greater current demand for training chips, the need for inference chips is expected to surpass them as AI applications expand.
OpenAI has assembled a team of approximately 20 engineers, led by experts who previously developed tensor processing units (TPUs) at Google. Through Broadcom, they have secured manufacturing capabilities with TSMC and plan to produce their first custom-designed chip by 2026, although timelines may change.
Currently, OpenAI relies heavily on Nvidia’s chips for computational power, commanding over 80% of the market. However, supply shortages and rising costs are prompting major clients like Microsoft, Meta, and OpenAI to explore alternative solutions.
Recently, OpenAI has considered using AMD's new MI300X chips through Microsoft’s Azure, marking AMD’s attempt to capture a share of the Nvidia-dominated market. AMD anticipates $4.5 billion in AI chip sales in 2024, following the release of new products by the end of 2023.
In addition to chip development, OpenAI is actively investing in and collaborating on data centers. The leadership has lobbied the U.S. government for large-scale data center expansion and is seeking funding from global investors, including those from the Middle East.
OpenAI's CFO, Sarah Friar, emphasized the significance of infrastructure in determining success in AI endeavors, highlighting both capital and personal learning expansions.
