A report says OpenAI is moving toward a custom AI chip with Broadcom, raising questions about Nvidia’s pricing power

An investing-focused analysis published by Yahoo Finance on June 26 raised a scenario that investors in Nvidia (NVDA) may want to watch closely. The piece argues that OpenAI is building its own custom AI chip in partnership with Broadcom, potentially targeting the same cost and performance levers that have helped Nvidia cement pricing power in the AI data-center market.

The article frames OpenAI’s chip program as a direct attempt to reduce dependency on third-party accelerators and to tailor hardware to its own model and inference needs. In that view, a customer with scale and customization capabilities could negotiate differently, especially if it can shift parts of its compute stack away from vendor-supplied GPUs.

Crucially, the Yahoo Finance write-up is presented as an investor question rather than a confirmed hardware rollout timeline. The analysis focuses on the strategic implication: if large AI operators diversify away from relying solely on Nvidia’s chips, Nvidia’s ability to sustain premium pricing could face more pressure, even if Nvidia remains the leading supplier for training and deployment in the near term.

Broadcom is mentioned in the piece as the chip-making partner. While the article’s framing suggests the effort could involve advanced silicon design and manufacturing expertise, no additional technical details, contract terms, or delivery milestones are provided in the information available for this review. As a result, it is not possible to determine from the published analysis how much compute, what model workloads, or what portion of OpenAI’s roadmap would be served by the custom chip.

For Nvidia, the key question is less about whether custom chips exist, and more about how they compete inside the full stack of AI infrastructure. Nvidia’s advantage has historically been tied not just to raw chip performance, but also to the software ecosystem around GPUs, including acceleration libraries and tooling that make it easier for developers and data centers to train and run models at scale. If a custom accelerator requires significant changes to software workflows, adoption may be slower and the economic impact may be smaller than a simple comparison of chip prices would suggest.

The article’s prompt to investors is therefore cautious: even if OpenAI builds custom accelerators, that does not automatically displace Nvidia across all workloads. A mixed approach is common in data centers, where operators may use multiple chip families depending on model type, latency needs, cost per token, power constraints, and supply availability. Without disclosed procurement volumes or performance benchmarks, the degree of competitive threat remains uncertain.

Another factor is that Nvidia’s customer base typically includes both hyperscalers and enterprise users that may not have the same scale or in-house engineering capacity as the largest AI labs. If custom chip development stays concentrated among a small number of top-tier model owners, the overall market dynamics could still favor Nvidia, even if individual customer negotiations become more complex.

What to watch next is whether OpenAI or Broadcom provides additional disclosures about the chip’s purpose and deployment, such as performance claims, manufacturing timelines, or how it will be integrated into production inference. Investors will also want to see whether Nvidia’s guidance and customer commentary begin to reflect pricing or volume pressure, particularly for the highest-demand accelerator configurations tied to training and inference workloads.