OpenAI outlines Jalapeño inference chip plans, with large-scale internal deployment targeted for 2027
Редакция ЦОД:Главное
OpenAI has disclosed further details of Jalapeño, its first in-house AI inference chip. The company plans a limited production-environment trial before targeting large-scale deployment in 2027, initially for internal use only.
- Jalapeño is OpenAI’s first disclosed in-house AI inference chip.
- OpenAI plans a limited production-environment trial before a targeted large-scale deployment in 2027.
- Initial Jalapeño capacity is intended solely for OpenAI’s internal ecosystem.
- The chip is described as programmable, with support for Transformer architectures and hardware-assisted speculative decoding.
- OpenAI’s projected 3x to 5x internal-model performance gain relates to planned multi-token prediction and is not a confirmed current result.
- No power, chip-volume, production, CAPEX or data-center-capacity figures were disclosed.
OpenAI has provided additional details on Jalapeño, its first internally developed chip for AI inference, positioning it as part of a broader effort to improve inference efficiency under data-center compute and power constraints.
Richard Ho, OpenAI’s vice president of hardware, described a co-design approach spanning models, software and hardware. The stated objective is to reduce compute cost and service latency while improving the efficiency of inference workloads constrained by available data-center resources and electrical supply.
OpenAI says Jalapeño is programmable and designed to support not only its own models but also a range of open models and Transformer architectures. The chip includes hardware support for speculative decoding. The company also plans to introduce multi-token prediction over time; it has indicated that this could provide a 3x to 5x performance improvement for internal models, though this is a projected outcome rather than a demonstrated current result.
According to the source, the team completed testing and a public demonstration in about two months after receiving physical chips. The next step is a limited trial in a live production environment. Large-scale deployment is planned for 2027, rather than representing commissioned capacity today.
In the near term, all Jalapeño volume is intended for OpenAI’s internal ecosystem, with no external customer supply. OpenAI has also indicated that it will retain a diversified hardware strategy and continue assessing external chips where they offer a better cost-performance balance.
For data-center operators and infrastructure suppliers, the significance is less about a disclosed chip specification—none has been provided—than about the optimization target. If OpenAI’s plans are realized, inference efficiency per unit of available power and infrastructure cost could become a more important procurement criterion than accelerator peak performance alone. The initiative could also reduce part of OpenAI’s internal requirement for external inference accelerators, although the company’s stated multi-platform approach suggests external hardware will remain under evaluation.
Jalapeño reflects an effort by a major AI infrastructure buyer to address inference cost, compute availability and power constraints through a custom ASIC and full-stack co-design. If executed, the approach could shift part of demand away from general-purpose external accelerators toward programmable platforms tuned for inference efficiency per unit of available power and infrastructure cost.
- TechNews 科技新報 / КитайскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.