OpenAI has announced its first in-house accelerator chip, named Jalapeño. The company says the chip is a 700-watt inference accelerator developed in partnership with Broadcom and designed in nine months. Jalapeño is presented as the first of three planned generations.
Performance and cost
According to benchmarks published by SemiAnalysis, Jalapeño delivered up to 104× speedups on certain tasks compared with Nvidia’s GB200 and GB300 GPUs. The same independent analysis estimates the cost at roughly $1.56 per chip-hour.
Implications for the CUDA ecosystem
Following the announcement, some observers declared that "CUDA is dead," arguing that Jalapeño threatens the dominance of the CUDA-based GPU ecosystem. The coverage recalls earlier predictions that alternatives would supplant CUDA—Google’s TPU and Groq were both previously cited as potential disruptors; the article notes an assertion that Nvidia purchased Groq’s technique for $20 billion.
In practice, however, previous declarations of CUDA’s demise have not eliminated Nvidia’s market position: Google has kept TPU largely for internal use, and several alternative projects have ended up either acquired by Nvidia or confined to a single company’s data centers. The Jalapeño announcement raises a similar question: whether OpenAI will make the chip broadly available or keep it for internal use only. The initial disclosure does not promise general commercial availability.
What to watch next
- SemiAnalysis’s benchmarks and cost estimates are significant but warrant independent confirmation.
- If Jalapeño remains a closed, internal resource for OpenAI, its technical edge may not translate into broader weakening of the CUDA ecosystem for outside users.
- As the first generation of a planned series, Jalapeño suggests OpenAI intends to continue investing in custom hardware, which could shape competition for serving large language models over time.
Overall, Jalapeño represents a notable hardware development and potential competitive pressure on current GPUs, but its ultimate impact depends on OpenAI’s distribution choices and how independent, real-world testing validates the reported performance and cost figures.
