OpenAI is reportedly working with Broadcom Inc. and Taiwan Semiconductor Manufacturing Co. Ltd. to build a new artificial intelligence chip specifically designed to run AI models after they’ve been trained.
The decision by OpenAI comes after the company has examined a range of options to diversify chip supply and reduce costs. According to Reuters, OpenAI considered building everything in-house and raising capital for an expensive plan to build its own foundries but has since dropped the plan given the huge cost and time needed to do so, instead now pursuing plans to focus on in-house chip design efforts.
Bloomberg reported that the decision to pursue in-house chip design is part of OpenAI not looking to replace GPUs such as those provided by Nvidia Corp. Instead, it’s looking to design a specialized chip that will undertake inference, the process of applying trained models to make predictions or decisions on new data in real-time applications. The same report notes that investors and analysts expect the need for chips to support inference will only grow as more tech companies use AI models to undertake more complex tasks.
Sources told Bloomberg that the decision by OpenAI to work with partners on custom chips was made because it’s a quicker and more attainable path for now. However, the company may continue research on setting up its own network of foundries in the future.
OpenAI’s move into in-house chip design is part of a broader trend among major tech firms and AI companies to develop specialized hardware that can meet the unique demands of AI workloads more efficiently than general-purpose GPUs.
Currently, OpenAI relies heavily on Nvidia’s GPUs for training its models, a process that requires immense computational power to refine algorithms using vast datasets. However, inference requires different chip capabilities that are optimized for speed and energy efficiency rather than raw computational power.
With chip costs continuing to rise relative to surging demand, such as Elon Musk’s xAI Corp. seeking to double its Colossus data center to 200,000 Nvidia H100 graphics cards and an overall increased demand for AI-driven services, creating custom chips will allow OpenAI to tailor its infrastructure to meet both technical needs and budget constraints.
By collaborating with Broadcom and TSMC, OpenAI can leverage established manufacturing expertise while moving faster than it could with a fully in-house production approach.