OpenAI Jalapeño(首款自研推理芯片)OpenAI Jalapeño (first in-house inference chip)
未公布单价No public price未公布单价No public price · 官方页未给公开单价;这类产品通常按整机、云实例或询价成交,不拿经销商价冒充官方价。No public unit price on the official page; these usually sell as systems, cloud instances or on request, so dealer quotes are not presented as official.
回答「OpenAI 有没有自研芯片、现在到哪一步」。Jalapeño 是 OpenAI 首款 Intelligence Processor,专为 LLM 推理设计,由 OpenAI 设计、Broadcom 做硅实现与网络、Celestica 做板卡/机架;工程样片已在实验室以量产目标频率与功耗跑 GPT-5.3-Codex-Spark,设计到流片 9 个月,计划 2026 年底初步部署,后续扩到吉瓦级。
不主张:官方没有公开 HBM 容量、算力、功耗、制程与单价;「每瓦性能大幅超越现有最先进水平」是 OpenAI 基于早期测试的口径,详细技术报告尚未发布;2026 年底是初步部署计划,不是已量产;Jalapeño 不对外销售或出租,先服务 OpenAI 自家模型与产品。Answers whether OpenAI has its own silicon and how far it has gone. Jalapeño is OpenAI's first Intelligence Processor, designed for LLM inference: OpenAI designs it, Broadcom handles silicon implementation and networking, and Celestica builds boards and racks. Engineering samples already run GPT-5.3-Codex-Spark in the lab at production target frequency and power; design-to-tape-out took nine months; initial deployment is planned for the end of 2026, expanding over following years.
It does not claim: OpenAI has not published HBM capacity, compute, power, process node or unit price; "substantially better performance per watt than today's state of the art" is OpenAI's early-test claim and the detailed technical report is still to come; end-2026 is an initial deployment plan, not volume production; Jalapeño is not sold or rented externally and first serves OpenAI's own models and products.
关键参数Key specs
- 定位 / role
- LLM 推理
- 设计 / design
- OpenAI
- 硅实现与网络 / silicon & network
- Broadcom(含 Tomahawk 网络芯片)
- 板卡与机架 / boards & racks
- Celestica
- 进度 / progress
- 工程样片以量产目标频率/功耗运行 GPT-5.3-Codex-Spark
- 开发周期 / cycle
- 设计到流片 9 个月
- 部署 / deployment
- 2026 年底初步部署,后续扩至吉瓦级
- 公开规格 / public specs
- 官方未披露 HBM / 算力 / 功耗 / 制程
- 状态 / status
- 2026-06-24 发布
官方原文怎么说What the source says
OpenAI 與 Broadcom(NASDAQ: AVGO)今日發佈 Jalapeño,這是 OpenAI 首款 Intelligence Processor(智能處理器):一款按 OpenAI 對 LLM 推理未來願景而設計的加速器。
Jalapeño 晶片的工程樣本已在實驗室以量產目標頻率與功耗運行機器學習(ML)工作負載,包括 GPT‑5.3‑Codex‑Spark。
Jalapeño 從初步設計到製造流片僅用九個月便共同開發完成
該平台計劃於 2026 年底前初步部署,並在其後數年擴展
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-20
- 许可 / 类型Licence / kind
- official-release
- 许可说明Licence evidence
- https://openai.com/zh-Hant-HK/index/openai-broadcom-jalapeno-inference-chip/
- 类目Category
- 加速卡Accelerators
- 状态Status
- 原型Prototype
- 价格口径Price basis
- 未公布单价No public price · not-published