跳到主要內容

Google TPU評測:Broadcom出貨、Meta洽談,Nvidia股價下跌 | Google TPU Review: Broadcom Ships, Meta Talks, Nvidia Dips

By Kit 小克 | AI Tool Observer | 2026-09-05

🇹🇼 Google TPU評測:Broadcom出貨、Meta洽談,Nvidia股價下跌

Google TPU這幾天成為AI圈最熱話題:Broadcom在最新財報電話會議證實,新一代TPU v8i已進入量產出貨階段,加上傳出Meta正在洽談租用Google TPU的合作案,消息一出Nvidia股價應聲下跌逾2%,Alphabet股價則大漲超過4%。這場晶片戰不只是股價數字的遊戲,而是牽動未來AI服務成本與運算資源分配的關鍵轉折。

Google TPU這次到底發生了什麼事

Broadcom執行長Hock Tan在財報會議上證實,Anthropic「有望成為我們最大的XPU客戶」,同時新一代TPU v8i已開始量產出貨,接續先前Ironwood(TPU v7)對Google與Anthropic的大量交付。幾個重點整理:

  • Anthropic已與Google、Broadcom簽署協議,將TPU算力擴大到每年最高5GW,2027年起陸續上線
  • Anthropic目前混合使用AWS Trainium、Google TPU與Nvidia GPU三種硬體,依任務類型分配算力
  • 分析師估計Broadcom今年從Anthropic獲得約210億美元營收,2027年可能翻倍到420億美元
  • Google目前已占Broadcom AI運算業務營收超過八成

Meta也要跟進?Nvidia股價為何應聲下跌

更讓市場緊張的是,據傳Meta正在與Google洽談租用TPU的合作案,預計2026年先以租賃Google Cloud TPU的方式起步,2027年再轉為直接採購自建。消息傳出後,Nvidia股價單日下跌超過2%,Alphabet(Google母公司)則上漲逾4%——市場把這解讀為Google晶片策略獲得業界巨頭驗證的信號。

Nvidia真的要被超車了嗎?誠實看待

先別急著下結論。目前Nvidia仍握有約七到九成的AI晶片市場份額,市值逼近5兆美元,短期內不會被撼動。TPU與其他客製化ASIC晶片(如Meta的MTIA、Amazon的Trainium)成長快速,預估2026年ASIC類AI伺服器出貨占比會衝上27.8%、年增44.6%,遠高於通用GPU的16.1%成長——但這代表的是「多元化」而非「取代」。對大多數開發者來說,短期內不會直接感受到差異,Nvidia的CUDA生態系與工具鏈優勢依然難以撼動。

真正該關注的,是長期而言雲端AI服務成本是否會因競爭而下降。Anthropic官方公告提到混合硬體策略能提升「效能與韌性」,Broadcom財報會議紀錄證實TPU出貨已進入實質放量階段,而Meta與Google洽談TPU的報導則顯示這場晶片多元化的趨勢正加速擴散到更多AI巨頭身上。

好不好用,試了才知道


🇺🇸 Google TPU Review: Broadcom Ships, Meta Talks, Nvidia Dips

Google TPU is suddenly the hottest topic in AI hardware: Broadcom confirmed on its latest earnings call that the next-generation TPU v8i has entered volume production shipment, and reports surfaced that Meta is in talks to rent Google TPU capacity. The news sent Nvidia shares down more than 2% in a single day, while Alphabet climbed over 4%. This isn't just a stock-price story — it's a real shift in how AI compute gets sourced and priced.

What Actually Happened with Google TPU This Week

Broadcom CEO Hock Tan confirmed on the earnings call that Anthropic is "on track to become our largest XPU customer," while the new TPU v8i has begun volume shipments, following high-volume deliveries of the previous-gen Ironwood (TPU v7) to both Google and Anthropic. Key facts:

  • Anthropic has expanded its agreement with Google and Broadcom to up to 5GW of annual TPU capacity, rolling out starting 2027
  • Anthropic currently mixes AWS Trainium, Google TPU, and Nvidia GPU hardware, matching workloads to whichever chip fits best
  • Analysts estimate Broadcom will pull in roughly $21B in revenue from Anthropic this year, potentially doubling to $42B in 2027
  • Google now accounts for over 80% of Broadcom's AI compute revenue

Is Meta Joining Too? Why Nvidia Stock Dipped

What really spooked the market: Meta is reportedly negotiating to rent Google TPU capacity, starting with Google Cloud rentals in 2026 before shifting to direct purchases in 2027. When the talks leaked, Nvidia shares fell over 2% in a day while Alphabet (Google's parent) rose more than 4% — investors read it as validation of Google's chip strategy by another Big Tech giant.

Is Nvidia Actually Getting Overtaken? An Honest Take

Don't jump to conclusions. Nvidia still holds roughly 70-90% of the AI chip market and sits near a $5 trillion market cap — it's not getting displaced anytime soon. TPU and other custom ASICs (Meta's MTIA, Amazon's Trainium) are growing fast — ASIC-based AI server shipments are projected to hit 27.8% market share in 2026, up 44.6% year-over-year, far outpacing the 16.1% growth for general-purpose GPUs — but that's diversification, not replacement. Most developers won't feel any direct difference in the near term; Nvidia's CUDA ecosystem and tooling lead is still hard to beat.

What's actually worth watching is whether cloud AI service costs drop over time as competition intensifies. Anthropic's official announcement notes that a mixed-hardware strategy improves "performance and resilience," Broadcom's earnings call confirms TPU shipments have hit real volume, and reports on Meta-Google TPU talks show this diversification trend is spreading fast to more AI giants.

好不好用,試了才知道

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

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

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