OpenAI and Broadcom announced Jalapeño as a custom inference accelerator. OpenAI later published results for selected language-model workloads, which need to be read with the comparison method and deployment plans.
Jalapeño inference chip announcement
The June announcement describes a platform designed around language-model inference, developed with Broadcom and other infrastructure partners. It reports engineering samples running workloads and describes a multi-generation deployment plan. An announced design and early samples are different from general retail availability or a UK-specific purchasing offer.
https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
What the later measurements concern
OpenAI’s August results use InferenceX to compare throughput, latency and power efficiency on selected model-serving workloads. The appendix identifies comparison systems including GB200 and GB300 for different models. The results are OpenAI’s published measurements, rather than an independent finding that one chip wins every workload or a direct estimate of a UK business’s operating costs.
https://openai.com/index/jalapeno-first-results/
Keep deployment plans and broader claims separate
The results describe continuing qualification and software work, with planned deployment in OpenAI’s infrastructure by year-end. They also say OpenAI will continue using NVIDIA and other partners for training and inference. A first-party accelerator therefore should not be presented as confirmation that all NVIDIA hardware has been replaced or that a deployment plan has already been completed.
https://openai.com/index/jalapeno-first-results/
Sources and further reading
https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
https://openai.com/index/jalapeno-first-results/