OpenAI Ra Mat Chip AI Rieng "Jalapeno": 9 Thang, Pha Vo The Doc Quyen Nvidia
2026-06-25BlockTempo / OpenAI / Broadcom
AI-chipOpenAINvidiaASICinference
Tom Tat: OpenAI va Broadcom vua cong bo chip AI ASIC dau tien mang ten "Jalapeno" vao 24/06/2026. Chip duoc thiet ke tu con so 0, toi uu rieng cho LLM inference, chi mat 9 thang tu thiet ke den tape-out. Hieu suat/watt vuot xa trang thai nghe thuat hien tai. Du kien trien khai quy mo gigawatt vao cuoi 2026, giam phu thuoc vao Nvidia.
Su Kien: Ngay 24/06/2026, OpenAI va Broadcom cung cong bo chip "Jalapeno" - mot ASIC (Application-Specific Integrated Circuit) duoc thiet ke tu dau danh rieng cho suy luan LLM. CEO Broadcom Hock Tan da trao tan tay mau chip cho Sam Altman va Greg Brockman. Chip da chay thanh cong GPT-5.3-Codex-Spark trong phong lab voi tan so va cong suat tieu thu muc tieu.
事件：2026 年 6 月 24 日，OpenAI 與 Broadcom 共同發表「Jalapeño」晶片——一款從零設計、專為 LLM 推論打造的 ASIC。Broadcom CEO Hock Tan 親自將樣品交付 Sam Altman 與 Greg Brockman。晶片已在實驗室以目標頻率與功耗成功運行 GPT-5.3-Codex-Spark。
2. Tai Sao Quan Trong? / Why It Matters
Y Nghia: Day la buoc di quan trong nhat cua OpenAI tu "cong ty phan mem" sang "nen tang toan dien" (full stack). Thay vi chi phu thuoc vao GPU Nvidia, OpenAI gio co the tu thiet ke phan cung toi uu cho chinh model cua minh. Dieu nay giong nhu Apple thiet ke chip M-series cho iPhone thay vi dung chip Snapdragon - kiem soat toan bo stack de dat hieu suat toi da.
意義：這是 OpenAI 從「純軟體公司」跨足「全端平台」的最關鍵一步。不再只依賴 NVIDIA GPU，而是為自己的模型量身打造硬體。就像 Apple 自研 M 系列晶片取代 Intel——掌控全端堆疊以實現最佳效能。九個月從設計到 tape-out 的速度也打破業界紀錄（通常 ASIC 需 18-24 個月）。
9 thang
Tu thiet ke den tape-out
ASIC
Khong phai GPU da nang
Cuoi 2026
Trien khai gigawatt
GPT-5.3
Da chay trong lab
3. He Qua Thi Truong / Market Impact
Thi Truong: Nvidia hien nam 80-90% thi phan chip AI. Su xuat hien cua Jalapeno - cung voi chip Trainium cua AWS, MI400 cua AMD, va chip Cerebras - danh dau su da dang hoa nguon cung cap phan cung AI. OpenAI cung dang hop tac voi AWS (Trainium), AMD, va Cerebras de giam phu thuoc vao Nvidia. Microsoft se la doi tac datacenter chinh cho viec trien khai quy mo gigawatt.
市場：NVIDIA 目前掌握 80-90% AI 晶片市佔。Jalapeño 的出現——加上 AWS Trainium、AMD MI400、Cerebras——標誌著 AI 硬體供應鏈的多元化。OpenAI 同時與 AWS（Trainium）、AMD、Cerebras 合作降低單一供應商依賴。Microsoft 將是吉瓦級部署的主要資料中心合作夥伴。
4. Goc Nhin DKY / Our Take
Phan Tich: Dung luu y den yeu to "9 thang". Day khong phai la ky thuat ban dan thong thuong - OpenAI da dung chinh model AI cua minh de thiet ke chip AI. AI thiet ke AI. Vong lap nay se chi con tang toc. Thu hai, chip ASIC chuyen dung luon co rui ro: no chi toi uu cho inference, khong cho training. Nhung GPT-5.3-Codex-Spark la mot agent model, khong phai chatbot thong thuong - inference la workload chinh cua agent. OpenAI dang dat cuoc rang tuong lai la inference-heavy, khong phai training-heavy.
Thuc Tien: Khi OpenAI tu thiet ke chip cho agent workload, ho dang xac nhan rang agent inference la mot loai workload moi - khac voi chatbot, khac voi training. Dieu nay cung co nghia: cac agent framework nhu Hermes can duoc thiet ke de chay hieu qua tren phan cung inference (khong GPU da nang). Oracle ARM (noi DKY dang chay) la mot vi du tot: ARM-based inference cho agent nho, chi phi thap. Khi chip inference tro nen pho bien va re hon, "personal AI agent" chay locally se tro thanh hien thuc som hon du kien.