跳到主要內容

Google Cloud Next 2026:75% 程式碼由 AI 生成、第八代 TPU 雙晶片、Gemini Agent 平台全面升級 | Google Cloud Next 2026: 75% AI-Generated Code, 8th-Gen TPU Dual Chips, Gemini Agent Platform

By Kit 小克 | AI Tool Observer | 2026-04-23

🇹🇼 Google Cloud Next 2026:75% 程式碼由 AI 生成、第八代 TPU 雙晶片、Gemini Agent 平台全面升級

Google Cloud Next 2026 在 4 月 22 日登場,Sundar Pichai 丟出一個震撼數字:Google 75% 的新程式碼現在由 AI 生成,工程師負責審核。同時發表第八代 TPU 雙晶片架構(TPU 8t + TPU 8i)、Gemini 企業級 Agent 平台,以及 Ironwood TPU 正式 GA。這場大會的訊號很明確——Google 正把 AI 從輔助工具變成開發主力。

Google 75% 程式碼由 AI 生成是什麼意思?

Google CEO Sundar Pichai 在 Cloud Next 2026 上宣布,Google 內部 75% 的新程式碼現在由 AI 撰寫,工程師負責審核和批准。這個數字去年秋天還只是 50%,半年內就跳升了 25 個百分點。

  • 工程師的角色正從「寫 code」轉向「審核和指揮 AI agent」
  • 一個複雜的程式碼遷移專案,由 AI agent 和工程師協作,速度比一年前快了 6 倍
  • 這不只是 autocomplete——Google 已經進入真正的 agentic workflow,AI 自主執行完整任務

TPU 8t 和 TPU 8i 有什麼不同?

Google 首次把 TPU 拆成訓練和推理兩條產品線,各自針對不同工作負載優化。

  • TPU 8t(Training):單一 superpod 可擴展到 9,600 顆晶片,共享 2PB 高頻寬記憶體,處理能力是 Ironwood 的 3 倍,能耗效率提升 2 倍
  • TPU 8i(Inference):採用 Boardfly 拓撲直連 1,152 顆 TPU,on-chip SRAM 增加 3 倍,推理性價比比上一代好 80%
  • TPU 8i 的設計目標:讓數百萬個 AI agent 同時運行,成本可控

Gemini 企業級 Agent 平台能做什麼?

Google 推出 Gemini Enterprise Agent Platform,定位是企業 AI agent 的「任務控制中心」。

  • 從管理單一助手,擴展到同時管理 數千個專業化 agent
  • 擴展 A2A(Agent-to-Agent)協議,支援多 agent 協作
  • Vertex AI Agent Builder 2.0 推出按成果計費的定價模式
  • 近 75% 的 Google Cloud 客戶已在使用 AI 產品,過去 12 個月有 330 家客戶各處理超過 1 兆個 token

Ironwood TPU 正式上線代表什麼?

Ironwood(第七代 TPU)正式 GA,是 Google 第一款專為推理時代設計的 TPU。它的定位是為企業提供高性價比的 AI 推理基礎設施,搭配 TPU 8i 形成完整的推理產品線。對企業用戶來說,這意味著跑 AI agent 的成本會持續下降。

Google 這次的訊息很清楚:AI 不再只是功能,而是基礎設施。當 75% 的程式碼都由 AI 寫出來時,「AI 會不會取代工程師」這個問題的答案已經不重要了——重要的是你會不會用 AI 來寫 code。好不好用,試了才知道。


🇺🇸 Google Cloud Next 2026: 75% AI-Generated Code, 8th-Gen TPU Dual Chips, Gemini Agent Platform

Google Cloud Next 2026 opened on April 22 with a headline stat from CEO Sundar Pichai: 75% of all new code at Google is now AI-generated, with engineers reviewing and approving the output. The event also introduced 8th-generation TPU chips split into training and inference variants, the Gemini Enterprise Agent Platform, and the general availability of Ironwood TPUs. The message is clear — Google is shifting AI from a developer tool to the primary development engine.

What Does "75% AI-Generated Code" Actually Mean?

At Cloud Next 2026, Pichai revealed that 75% of new code at Google is now written by AI and approved by human engineers — up from 50% just last fall.

  • Engineers are moving from writing code to orchestrating and reviewing AI agents
  • A complex code migration project completed 6x faster using AI-engineer collaboration versus engineers alone one year prior
  • This goes beyond autocomplete — Google is running full agentic workflows where AI autonomously executes complete tasks

TPU 8t vs TPU 8i: Why Google Split Its Chip Line

For the first time, Google is shipping separate TPU chips optimized for training and inference workloads.

  • TPU 8t (Training): Scales to 9,600 chips per superpod with 2PB shared high-bandwidth memory, 3x processing power over Ironwood, 2x better performance per watt
  • TPU 8i (Inference): Boardfly topology connecting 1,152 TPUs directly, 3x more on-chip SRAM, 80% better price-performance than the prior generation
  • TPU 8i is purpose-built to run millions of concurrent AI agents cost-effectively

What Is the Gemini Enterprise Agent Platform?

Google launched the Gemini Enterprise Agent Platform — essentially mission control for managing AI agents at enterprise scale.

  • Designed to manage thousands of specialized agents simultaneously, not just a single assistant
  • Expanded A2A (Agent-to-Agent) protocol for multi-agent orchestration
  • Vertex AI Agent Builder 2.0 with outcome-based pricing
  • Nearly 75% of Google Cloud customers already use AI products; 330 customers each processed over 1 trillion tokens in the past 12 months

Ironwood TPU Now Generally Available

Ironwood (7th-gen TPU), Google's first TPU designed specifically for the inference era, is now GA. Paired with TPU 8i, it forms a complete inference product line. For enterprises, this means the cost of running AI agents will continue to drop.

Google's message at Cloud Next 2026 is unmistakable: AI is no longer a feature — it is infrastructure. When 75% of code is AI-generated, the question is no longer whether AI will replace engineers. The question is whether you know how to use AI to write code. You never really know until you try it yourself.

Sources / 資料來源

常見問題 FAQ

Google 75% 程式碼由 AI 生成是真的嗎?

是的,Sundar Pichai 在 Cloud Next 2026 上確認,Google 內部 75% 的新程式碼由 AI 撰寫、工程師審核,去年秋天還只有 50%。

TPU 8t 和 TPU 8i 有什麼差別?

TPU 8t 專為訓練設計,可擴展到 9,600 顆晶片;TPU 8i 專為推理設計,推理性價比提升 80%,可同時運行數百萬個 AI agent。

Gemini Enterprise Agent Platform 是什麼?

企業級 AI agent 管理平台,可同時管理數千個專業化 agent,支援 A2A 協議和按成果計費的定價模式。

Google Cloud Next 2026 最重要的發表是什麼?

75% AI 生成程式碼、第八代 TPU 雙晶片架構、Gemini Agent 平台、Ironwood TPU 正式 GA,以及 Vertex AI Agent Builder 2.0。

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

留言

這個網誌中的熱門文章

Stanford 研究登上《Science》:11 個 AI 模型有 47% 機率說你對,即使你錯了 | Stanford Study in Science: AI Models Validate Harmful Behavior 47% of the Time — Sycophancy Is a Real Problem

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code

Snap 裁員 1000 人因 AI 寫 65% 程式碼:科技業 AI 取代人力的第一槍 | Snap Cuts 1000 Jobs as AI Writes 65% of Code: First Major AI Workforce Replacement in Tech