[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-openai-jalapeno-chip-cuts-inference-costs-zh":3,"article-related-openai-jalapeno-chip-cuts-inference-costs-zh":33,"series-industry-41ddf4ec-a408-4095-b25b-b48c1d104a75":82},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":25,"views":29,"created_at":30,"published_at":31,"topic_cluster_id":32},"41ddf4ec-a408-4095-b25b-b48c1d104a75","openai-jalapeno-chip-cuts-inference-costs-zh","OpenAI 自研 Jalapeño，先砍推理成本","\u003Cp data-speakable=\"summary\">\u003Ca href=\"\u002Ftag\u002Fopenai\">OpenAI\u003C\u002Fa> 的首款自研晶片 Jalapeño 主要用來讓推理更快，也更省電更省錢。\u003C\u002Fp>\u003Cp>看完這 5 個重點，你可以判斷 OpenAI 是在補哪一段成本缺口、Broadcom 在這條供應鏈裡扮演什麼角色，以及這顆晶片會不會動到 \u003Ca href=\"\u002Ftag\u002Fnvidia\">Nvidia\u003C\u002Fa> 的生意。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>項目\u003C\u002Fth>\u003Cth>主要角色\u003C\u002Fth>\u003Cth>關鍵判斷\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>Jalapeño\u003C\u002Ftd>\u003Ctd>推理處理器\u003C\u002Ftd>\u003Ctd>早期測試顯示每瓦效能更好\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002F\">Nvidia\u003C\u002Fa> GPU\u003C\u002Ftd>\u003Ctd>通用 AI 算力\u003C\u002Ftd>\u003Ctd>仍可能主攻預訓練\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.broadcom.com\u002F\">Broadcom\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>設計與製造夥伴\u003C\u002Ftd>\u003Ctd>協助 OpenAI 做客製化晶片\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Google 自研晶片\u003C\u002Ftd>\u003Ctd>AI 加速器\u003C\u002Ftd>\u003Ctd>同樣是降低對外部 GPU 的依賴\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>1. Jalapeño：先打推理，不碰全能算力\u003C\u002Fh2>\u003Cp>Jalapeño 是 OpenAI 第一顆自研晶片，重點不是通吃所有 AI 工作，而是專門處理推理，也就是模型訓練完成後，回應使用者提示、產生\u003Ca href=\"\u002Fnews\u002Fai-code-review-tools-catch-issues-earlier-zh\">程式碼\u003C\u002Fa>或執行代理動作的那一步。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782515862249-pgls.png\" alt=\"OpenAI 自研 Jalapeño，先砍推理成本\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這種切法很務實，因為推理是產品端最常發生的成本。OpenAI 目前對外說法是，早期測試已經看到比現有主流方案更好的每瓦效能，代表單位工作消耗的電力有下降空間。\u003C\u002Fp>\u003Cul>\u003Cli>面向推理，不是預訓練\u003C\u002Fli>\u003Cli>鎖定即時程式碼模型\u003C\u002Fli>\u003Cli>仍在測試階段\u003C\u002Fli>\u003Cli>目標是壓低運行成本\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>2. Broadcom：把晶片做成 OpenAI 的工作流\u003C\u002Fh2>\u003Cp>\u003Ca href=\"https:\u002F\u002Fwww.broadcom.com\u002F\">Broadcom\u003C\u002Fa> 是這顆晶片背後的設計與製造夥伴，這讓 OpenAI 不只是買硬體，而是開始參與硬體定義。雙方合作先前已在 10 月對外宣布，這次則是成果首次浮出檯面。\u003C\u002Fp>\u003Cp>對 OpenAI 來說，這種合作的價值在於可控性。當晶片是圍繞自家模型行為設計，而不是拿通用 \u003Ca href=\"\u002Ftag\u002Fgpu\">GPU\u003C\u002Fa> 去適配模型時，效能、延遲與成本都更容易被一起優化。\u003C\u002Fp>\u003Cul>\u003Cli>合作已在 10 月宣布\u003C\u002Fli>\u003Cli>晶片依 OpenAI 工作負載定制\u003C\u002Fli>\u003Cli>屬於自研 AI 硬體布局的一部分\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>3. 推理優先：省下的是日常流量的錢\u003C\u002Fh2>\u003Cp>推理和預訓練的差別很大。預訓練像是把模型「教會」，需要大量算力與記憶體頻寬；推理則是模型「上線回答問題」，更接近真實產品流量，也更容易長期累積成本。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782515859821-ayf8.png\" alt=\"OpenAI 自研 Jalapeño，先砍推理成本\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>OpenAI 特別提到這顆晶片是為了即時程式碼模型降低營運成本，這暗示它想先解最常見、最持續、最吃效能的場景，而不是先挑最昂貴的訓練大戰。\u003C\u002Fp>\u003Ccode>推理 = 已完成訓練的模型開始回答問題\u003Cbr>預訓練 = 用資料把模型教出來\u003Cbr>即時程式碼 = 高頻、低延遲的工作負載\u003C\u002Fcode>\u003Ch2>4. 全棧控制：從模型一路管到部署\u003C\u002Fh2>\u003Cp>OpenAI 的訊號不只是在做晶片，而是在做整套基礎設施。它提到的範圍包含晶片架構、核心函式、記憶體系統、網路、排程、部署系統，甚至產品體驗。\u003C\u002Fp>\u003Cp>這種全棧做法的好處是，模型、軟體與硬體可以一起調。當每一層都知道對方在做什麼時，速度、穩定性與成本都比較有機會一起改善，而不是\u003Ca href=\"\u002Fnews\u002Fproduct-hunt-best-prompt-engineering-tools-2026-zh\">各自最\u003C\u002Fa>佳化卻互相拖累。\u003C\u002Fp>\u003Cul>\u003Cli>晶片架構\u003C\u002Fli>\u003Cli>記憶體系統\u003C\u002Fli>\u003Cli>網路與排程\u003C\u002Fli>\u003Cli>部署系統\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>5. 對 Nvidia 的影響：不會立刻取代，但會慢慢分流\u003C\u002Fh2>\u003Cp>OpenAI 長期被視為高度依賴 \u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002F\">Nvidia\u003C\u002Fa> GPU，而 Jalapeño 的出現，就是在降低這種依賴。它不太可能全面取代 Nvidia，特別是在預訓練這種重度算力場景，但它有機會先吃掉\u003Ca href=\"\u002Fnews\u002Fdoubao-pro-turns-agent-into-office-tool-zh\">日常\u003C\u002Fa>推理流量。\u003C\u002Fp>\u003Cp>真正重要的是商業效果。只要推理成本下降一點點，像 \u003Ca href=\"\u002Ftag\u002Fcodex\">Codex\u003C\u002Fa> 或各種代理工具這類持續運作的產品，毛利改善就可能很明顯，因為使用量會隨著產品成熟快速放大。\u003C\u002Fp>\u003Ch2>怎麼挑：看你在意成本、供應鏈還是市場\u003C\u002Fh2>\u003Cp>如果你最在意 AI 產品的成本結構，Jalapeño 是最值得看的部分，因為它直接對準推理這個最常見的支出點。如果你關心晶片供應鏈與客製化硬體，Broadcom 的角色更關鍵，因為它代表 OpenAI 已經從買家走向共同設計者。\u003C\u002Fp>\u003Cp>如果你在追 AI 晶片市場的變化，這篇最重要的訊號是：競爭不再只看誰能訓練最大模型，也開始看誰能把模型跑得更便宜、更穩、而且更省電。\u003C\u002Fp>","OpenAI 首款自研晶片 Jalapeño 主攻推理，早期測試顯示有望提升每瓦效能並壓低即時 AI 工作負載成本。","techcrunch.com","https:\u002F\u002Ftechcrunch.com\u002F2026\u002F06\u002F24\u002Fopenai-unveils-its-first-custom-chip-built-by-broadcom\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782515862249-pgls.png","industry","zh","408080f9-631c-4d3c-87dc-be77bdd909b0",[17,18,19,20,21,22,23,24],"OpenAI","Jalapeño","推理晶片","Broadcom","Nvidia","自研晶片","AI 基礎設施","每瓦效能",[26,27,28],"Jalapeño 先攻推理，不是全面取代通用 GPU。","Broadcom 代表 OpenAI 正把硬體納入全棧優化。","這顆晶片的商業價值在於壓低日常推理成本。",0,"2026-06-26T23:17:18.106598+00:00","2026-06-26T23:17:18.077+00:00","caa87b65-9bbc-46fe-bba8-4f4158dd2d8b",{"tags":34,"relatedLang":41,"relatedPosts":45},[35,37,39],{"name":17,"slug":36},"openai",{"name":21,"slug":38},"nvidia",{"name":20,"slug":40},"broadcom",{"id":15,"slug":42,"title":43,"language":44},"openai-jalapeno-chip-cuts-inference-costs-en","OpenAI’s Jalapeño chip cuts inference costs","en",[46,52,58,64,70,76],{"id":47,"slug":48,"title":49,"cover_image":50,"image_url":50,"created_at":51,"category":13},"21f2afc6-8551-47be-8d10-639f01864016","microsoft-bare-metal-aks-ai-training-zh","微軟把 AKS 推向 AI 訓練核心","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782540169706-uqoi.png","2026-06-27T06:02:26.373585+00:00",{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":13},"2827a68b-3dd5-41cf-bbac-4e0d3779732c","deepseek-low-cost-chatbot-changed-ai-pricing-zh","DeepSeek 低價策略改寫 AI 定價","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782537467872-9zvb.png","2026-06-27T05:17:24.123784+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":13},"d61c0387-0b92-4df1-992d-b78395a6bc57","openai-latest-model-us-user-vetting-zh","5 個訊號看懂 OpenAI 新模型審查","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782527563380-nvrg.png","2026-06-27T02:32:18.096357+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":13},"b83ef891-1bb8-4772-99c0-5e23e273b26a","mythos-access-runs-through-washington-zh","Mythos 讓模型先過華府再上線","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782526689575-5qwa.png","2026-06-27T02:17:45.806919+00:00",{"id":71,"slug":72,"title":73,"cover_image":74,"image_url":74,"created_at":75,"category":13},"4f89db70-f0be-445d-887c-ddc3a91b1cc4","kalshi-turns-openai-ipo-timing-into-a-wager-zh","Kalshi把 OpenAI IPO 變成押注","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782525788869-to4z.png","2026-06-27T02:02:40.208569+00:00",{"id":77,"slug":78,"title":79,"cover_image":80,"image_url":80,"created_at":81,"category":13},"f8deff96-3085-4e62-9410-e37d688c2367","openai-gpt-56-controlled-preview-release-zh","GPT 5.6 受控預覽的 5 個重點","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782524867967-0jwd.png","2026-06-27T01:47:19.797614+00:00",[83,88,93,98,103,108,113,118,123,128],{"id":84,"slug":85,"title":86,"created_at":87},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":89,"slug":90,"title":91,"created_at":92},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":94,"slug":95,"title":96,"created_at":97},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":99,"slug":100,"title":101,"created_at":102},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":104,"slug":105,"title":106,"created_at":107},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":109,"slug":110,"title":111,"created_at":112},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":114,"slug":115,"title":116,"created_at":117},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":119,"slug":120,"title":121,"created_at":122},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":124,"slug":125,"title":126,"created_at":127},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":129,"slug":130,"title":131,"created_at":132},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]