OpenAI is finalizing its first in-house AI chip design and plans to begin production by 2026 with Taiwan Semiconductor Manufacturing Co. (TSMC), leveraging TSMC’s cutting-edge 3-nanometer process, according to Reuters.
This move aims to lessen OpenAI’s reliance on Nvidia, which holds an estimated 80% share of the AI chip market, and represents a broader trend of big tech firms pursuing proprietary silicon to meet escalating demands for AI workloads.
This development follows industry-wide efforts to control the supply chain for AI infrastructure. Companies like Meta, Microsoft, and AWS have all turned to custom chip development to optimize performance while reducing dependency on Nvidia.
Meta, for instance, has invested heavily in its Meta Training and Inference Accelerator (MTIA) chips to power its Llama AI models, which form a key part of its $60 billion AI infrastructure budget. Similarly, AWS has introduced its Trainium processors to accelerate large-scale workloads like generative AI models, while Apple is developing its Baltra server chips to improve its position in AI.
OpenAI’s efforts to build its own chips reflect a growing recognition that AI infrastructure needs cannot always be met by off-the-shelf hardware.
OpenAI’s Vision for Custom Silicon
Richard Ho, a veteran engineer who previously worked on Google’s Tensor Processing Units (TPUs), leads OpenAI’s chip development team. The group has grown to 40 engineers, focusing on creating chips designed to meet OpenAI’s unique requirements.
The initial chip iteration will target inference tasks, which involve running trained AI models efficiently. Over time, OpenAI aims to expand its chip capabilities to support both inference and training workloads, which require significantly higher computational power.
The planned chip will integrate a systolic array architecture—a highly specialized design for efficiently handling matrix operations—paired with high-bandwidth memory (HBM) to manage the vast data flows associated with advanced AI systems.
According to internal estimates, a single iteration of such a chip could cost $500 million to develop, making the process both a technical and financial challenge.
The Risks of In-House Chip Development
Custom chip manufacturing is fraught with risks, particularly during the tape-out phase, when the finalized chip design is sent for fabrication. Mistakes at this stage can cost tens of millions of dollars and result in months-long delays.
OpenAI is expected to mitigate some of these risks by partnering with TSMC, which has a proven track record in high-performance chip production. TSMC’s 3-nanometer process is one of the most advanced in the world, enabling greater transistor density and energy efficiency.
If successful, OpenAI plans to test its first chips on a limited scale by late 2025, paving the way for full deployment in its data centers the following year. This shift is expected to reduce costs and provide OpenAI with greater control over the hardware that underpins its advanced language models.
Geopolitical Challenges in AI Chip Manufacturing
OpenAI’s reliance on TSMC ties its chip development to broader geopolitical factors. TSMC, headquartered in Taiwan, plays a leading role in global semiconductor manufacturing, producing chips for major players like Nvidia, Apple, and now OpenAI.
The U.S. government’s export restrictions on advanced chips, particularly targeting China, add further complexity to the semiconductor supply chain. Former United States Secretary of Commerce Gina Raimondo highlighted the rationale behind these restrictions, stating last year, “The semiconductors that power artificial intelligence can be used by adversaries to run nuclear simulations, develop bio weapons, and advance their militaries.”
By partnering with TSMC, OpenAI ensures access to the latest manufacturing technology while navigating around the restrictions placed on China. However, this dependency on Taiwan’s semiconductor industry underscores the geopolitical risks associated with a region that has become increasingly central to global tech production.
The Future of OpenAI’s Hardware Strategy
While the risks are high, the potential benefits of in-house silicon are substantial. OpenAI’s custom chips could offer it a critical edge by improving performance, reducing costs, and enabling greater flexibility in its AI development.
This control over its hardware infrastructure could also strengthen its negotiating position with external suppliers like Nvidia, which continues to dominate the GPU market.
OpenAI’s strategy could have broader implications for the global AI hardware ecosystem. Success in this initiative might inspire other companies to invest in similar custom silicon programs, further intensifying competition in the semiconductor industry.
At the same time, geopolitical tensions and the increasing complexity of chip development may continue to challenge the global AI supply chain.
