OpenAI has developed its first in‑house inference chip, named Jalapeno, and its hardware team says the design reached production readiness in just nine months. The company plans to deploy these chips in its data centers before the end of the year. The move was driven in part by growing compute needs that expansions by major chipmakers such as TSMC, Samsung and SK Hynix could not fully satisfy.
Why OpenAI started building its own hardware
Two years ago Sam Altman, co‑founder of OpenAI, and Richard Ho, head of the company's hardware division, traveled to Asia to meet leaders of the largest chip manufacturers. Altman urged them to build many more fabs because OpenAI's demand for compute would continue to rise.
Although those manufacturers expanded capacity, it did not fully meet OpenAI's needs. That shortfall, together with the accelerating demand for compute, led OpenAI to begin developing its own hardware.
The Jalapeno chip and the "Hot Peppers" team
Richard Ho previously worked at Google, where he participated in the design of the first TPU (Tensor Processing Unit), a processor optimized for machine learning. At OpenAI he built the semiconductor team from scratch.
The company's initial in‑house chip was named Jalapeno. The internal development group is called "Hot Peppers," a name inspired by the team's shared fondness for spicy food. The Jalapeno is designed for inference workloads — running already‑trained models.
Ho told Forbes that the cycle from design to production‑ready silicon took only nine months.
Deployment timeline
OpenAI indicated that Jalapeno chips could be integrated into its data centers before the end of the year. The company has not provided a detailed rollout schedule or specified which sites will receive the chips based on the available information.
Implications for suppliers such as Nvidia
OpenAI's move shows that large AI companies are investing in proprietary hardware to alleviate reliance on external suppliers' capacity limits. This announcement raises questions about how the role of third parties, including Nvidia, may evolve within OpenAI's infrastructure. However, developing internal chips does not necessarily imply a complete move away from industry partners; it primarily increases OpenAI's independent compute capacity.
Conclusion
The rapid nine‑month development of an inference chip and plans for deployment before year‑end indicate OpenAI's intent to reduce dependency on external suppliers for compute capacity. In the short term, Jalapeno may ease capacity constraints; in the longer term, it could influence supplier relationships in the AI hardware ecosystem. The full impact will depend on further deployment and operational details.
