OpenAI, the company behind generative artificial intelligence (AI) systems including ChatGPT, is collaborating with Broadcom and Taiwan Semiconductor Manufacturing Company (TSMC) to develop its first custom AI chip, reports Reuters, citing sources.
The company, facing surging demands for its AI systems, is also expanding its chip sources by incorporating AMD chips alongside NVIDIA’s to support its operations.
Initially, OpenAI explored building a network of in-house “foundries” for chip manufacturing.
However, due to high costs and extended timelines, the company has shelved these plans, opting instead to focus on in-house chip design, according to sources.
OpenAI’s exploration of various strategies to diversify its chip supply and manage costs reflects a broader industry trend, with tech giants such as Amazon, Meta, Google, and Microsoft adopting similar approaches.
Its decision to design own chip, while still sourcing from a range of chipmakers, could influence the wider technology sector.
NVIDIA currently dominates the market with majority share in GPU technology.
However, chip shortages and escalating costs have prompted companies to seek alternatives.
AMD’s entry into the market with its new MI300X chips, facilitated by OpenAI’s planned use through Microsoft’s Azure, indicates a potential shift in the competitive landscape.
OpenAI, AMD, and TSMC have declined to comment on these developments. Broadcom has not responded to requests for comment.
OpenAI’s reliance on substantial computing power for training and running its AI models has made it one of the largest purchasers of NVIDIA GPUs.
These chips are essential for both training AI models and for inference tasks, where AI applies learned data to make new predictions or decisions.
The collaboration with Broadcom focuses on developing an AI chip tailored for inference tasks.
Industry analysts predict that the demand for inference chips may soon outpace that for training chips as AI applications become more widespread.
Broadcom brings expertise in refining chip designs for manufacturing and supplies design components that facilitate rapid data transfer are crucial for AI systems that require numerous interconnected chips.
OpenAI is still considering whether to develop additional elements of its chip design in-house or to seek further partnerships.
The company has formed a dedicated chip team, including engineers such as Thomas Norrie and Richard Ho, who have experience in creating Tensor Processing Units at Google.
Through its partnership with Broadcom, OpenAI has secured a manufacturing slot with TSMC to produce its inaugural custom-designed chip, with a tentative launch planned for 2026.
However, sources indicate that this timeline could be subject to change.
OpenAI’s valuation recently soared to $150bn following a new multi-billion funding round
