Jalapeño評測：OpenAI自研AI晶片效能超車輝達 | Jalapeño Review: OpenAI's Nvidia-Beating AI Chip

8月 30, 2026

By Kit 小克 | AI Tool Observer | 2026-08-31

🇹🇼 Jalapeño評測：OpenAI自研AI晶片效能超車輝達

Jalapeño 是 OpenAI 與 Broadcom 合作打造的第一顆自研 AI 晶片，專門處理大型語言模型的推論（inference）工作。SemiAnalysis 上週公布的獨立測試顯示，Jalapeño 在單位功耗算力上贏過輝達（Nvidia）現行的 GB200／GB300 系統，讓外界開始討論輝達的晶片霸主地位是否出現裂縫。

規格重點：13.4 PFLOPS、700W、HBM4 記憶體

Jalapeño 採用台積電 N3P（3 奈米等級）製程，主要規格如下：

算力：13.4 PFLOPS（MXFP4 精度），功耗僅 700W，相較輝達 Rubin 平台的 900～1,150W 更省電
記憶體：搭配 6 顆 HBM4，總容量 216GiB，頻寬達 15.4TB/s
吞吐量：在 DeepSeek R1 上可達每秒 700+ token／使用者，GPT-OSS 上更達約 1,400 token／秒

效能比拚：贏輝達，但贏多少？

根據 OpenAI 與 SemiAnalysis 的測試數據，Jalapeño 在 GPT-OSS、DeepSeek R1、Kimi K2.5 等開源模型上，單位功耗的工作量比輝達高出 1.5～1.9 倍，延遲最多降低 3.6 倍。Broadcom 執行長 Hock Tan 也表示，Jalapeño 能把每個 token 的推論成本壓低約 50%。這些數字如果屬實，代表大型 AI 公司自研晶片的策略正式進入「能打」的階段，不再只是省成本的備案。

開發時程：9 個月流片，號稱史上最快

OpenAI 表示 Jalapeño 從設計到流片（tape-out）只花了 9 個月，是「高效能先進半導體有史以來最快的 ASIC 開發週期」。晶片預計今年底前開始部署在 OpenAI 自家的運算基礎設施中。

誠實看：這不是要賣你的晶片，也不會馬上打趴輝達

先說重點：Jalapeño 不對外銷售，一般開發者無法購買或租用這顆晶片，它只會用在 OpenAI 自己的資料中心，跑自家模型的推論。所以短期內輝達的營收不會因此掉一塊肉——OpenAI 依然是輝達最大的客戶之一，訓練用的晶片也還是輝達的天下。真正的意義在於：Google（TPU）、Amazon（Trainium）、Meta（MTIA）之後，OpenAI 也加入自研晶片陣營，代表「用輝達的晶片跑推論」不再是唯一選項。對一般使用者來說，這可能在未來讓 OpenAI API 的推論成本下降，間接反映在價格上，但目前還看不到直接影響。

好不好用，試了才知道。

🇺🇸 Jalapeño Review: OpenAI's Nvidia-Beating AI Chip

Jalapeño is OpenAI's first custom-built AI chip, co-developed with Broadcom specifically for large language model inference. An independent benchmark published last week by SemiAnalysis shows Jalapeño beating Nvidia's current GB200/GB300 systems on performance-per-watt — reigniting the debate over whether Nvidia's chip dominance is starting to crack.

The Specs: 13.4 PFLOPS, 700W, HBM4 Memory

Built on TSMC's N3P (3nm-class) process, Jalapeño's headline numbers:

Compute: 13.4 PFLOPS at MXFP4 precision, drawing just 700W — well under Nvidia's Rubin platform at 900-1,150W
Memory: 6 HBM4 stacks totaling 216GiB, with 15.4TB/s of bandwidth
Throughput: 700+ tokens/sec/user on DeepSeek R1, and roughly 1,400 tokens/sec/user on GPT-OSS

Benchmarks: It Beats Nvidia — But By How Much?

According to OpenAI and SemiAnalysis's testing across open-weight models like GPT-OSS, DeepSeek R1, and Kimi K2.5, Jalapeño delivers 1.5-1.9x more work per watt than Nvidia, with latency cut by up to 3.6x. Broadcom CEO Hock Tan claims it also cuts inference cost per token by about 50%. If these numbers hold up in production, it signals that custom silicon from major AI labs has moved past "nice cost-saving side project" into genuinely competitive territory.

Nine Months to Tape-Out — Reportedly the Fastest Ever

OpenAI says Jalapeño went from design to tape-out in just nine months, which it calls the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors. The chip is expected to start deploying in OpenAI's own infrastructure by the end of the year.

The Honest Take: You Can't Buy It, and Nvidia Isn't Going Anywhere Yet

Here's the catch: Jalapeño isn't for sale. Developers can't buy or rent this chip — it's exclusively for OpenAI's own data centers, running its own models. So Nvidia's revenue isn't taking a hit anytime soon; OpenAI remains one of Nvidia's biggest customers, and training workloads are still firmly Nvidia territory. The real significance is that OpenAI now joins Google (TPU), Amazon (Trainium), and Meta (MTIA) in the custom-silicon club — proof that running inference on Nvidia GPUs is no longer the only option for frontier labs. For everyday users, this could eventually translate into cheaper OpenAI API pricing, but don't expect any direct impact just yet.

You won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士