Etched AI晶片評測：估值一個月狂飆到210億美元硬剛Nvidia | Etched AI Chip Review: $21B Valuation Doubles, Takes on Nvidia
By Kit 小克 | AI Tool Observer | 2026-08-19
🇹🇼 Etched AI晶片評測：估值一個月狂飆到210億美元硬剛Nvidia
Etched AI晶片估值在短短一個月內從103億美元翻倍衝上210億美元，這家由哈佛輟學生創立的新創，靠一顆只做一件事的晶片，正在挑戰Nvidia在AI推理市場的霸主地位。這輪7億美元新募資由老班底Sequoia、a16z與交易巨頭Jane Street領投，而Jane Street同時也是Etched的第一個正式客戶——這才是真正讓市場興奮的地方：不是簡報上的跑分，是有人真金白銀把它接上生產環境。
Etched的Sohu晶片憑什麼打Nvidia？
Etched的答案很直接：放棄通用性，換取極致速度。他們的Sohu晶片是一顆ASIC（專用積體電路），從電路設計那一天起就只為Transformer架構的推理運算服務，不像Nvidia H100那樣要同時兼顧訓練、推理、圖形運算等各種負載。這種「單一用途」的取捨讓Etched敢對外宣稱Sohu在Transformer推理任務上比H100快20倍。目前Etched已累積超過10億美元的客戶合約，累計募資逼近20億美元，還從Nvidia挖角了不少晶片工程師。
Jane Street實際用起來如何？
Jane Street上個月拿到第一批Sohu機架，已經投入實際交易相關工作負載，不是單純測試機。對一家以毫秒級效率為生的高頻交易公司來說，願意把核心運算換到一顆新創晶片上，本身就是很強的背書訊號。
ASIC路線的風險是什麼？
老實說，把整間公司押在「Transformer會繼續是主流架構」這個賭注上，風險不小。Etched AI晶片的效能優勢建立在Transformer結構不變的前提下——如果之後Mamba、SSM或其他新架構真的取代Transformer成為主流，這顆客製化ASIC可能一夕之間變成一塊沒用的矽。相較之下，Nvidia的GPU雖然單項效能輸給專用晶片，但通用性讓它能跟著架構演進持續有用武之地。這也是為什麼即使Etched估值飆到210億美元，多數雲端業者目前還是傾向把Etched當作H100的「加速選項」而非全面替代品。
常見問題
Q: Etched的Sohu晶片可以取代Nvidia GPU嗎？
目前還不行。Sohu只能做Transformer推理，訓練、圖形運算等其他工作仍需依賴通用GPU，兩者比較像互補而非全面替代。
Q: 一般開發者現在能買到Sohu晶片嗎？
還不行，目前僅開放給簽約企業客戶如Jane Street，尚未有公開販售或雲端API管道。
Q: Etched估值暴漲是因為技術突破還是市場炒作？
兩者都有：Jane Street實際部署是硬指標，但AI硬體投資熱潮也讓市場願意用更高倍數給估值，需持續觀察後續客戶擴張再判斷。
好不好用，試了才知道。
🇺🇸 Etched AI Chip Review: $21B Valuation Doubles, Takes on Nvidia
The Etched AI chip startup just saw its valuation double from $10.3B to $21 billion in under a month, and it's coming for Nvidia's grip on the AI inference market with a chip that does exactly one thing. The new $700M round was led by returning backers Sequoia and a16z, plus trading giant Jane Street — which is also Etched's first real production customer. That's the part that matters: not a benchmark slide, but an actual company running live workloads on it.
What Makes Etched's Sohu Chip Take On Nvidia?
Etched's bet is simple: trade generality for raw speed. Sohu is an ASIC (application-specific integrated circuit) designed from the transistor level up to do one job — transformer inference — instead of juggling training, inference, and graphics like Nvidia's H100. That narrow focus is what lets Etched claim a 20x inference speedup over the H100 on transformer workloads. The company has booked over $1 billion in customer contracts, raised nearly $2 billion total, and has been quietly poaching chip engineers away from Nvidia.
How Is It Actually Working for Jane Street?
Jane Street received its first Sohu rack last month and is running it against real trading-related workloads, not just a test bench. For a high-frequency trading firm that lives and dies by milliseconds, putting core compute on an unproven startup chip is itself a strong signal.
What's the Risk With Betting on an ASIC?
Honestly, betting an entire company on "transformers stay the dominant architecture" is a real gamble. The Etched AI chip's whole performance edge assumes transformer architecture doesn't change — if Mamba, state-space models, or some other successor architecture eventually replaces transformers, this custom silicon could become dead weight overnight. Nvidia's GPUs lose head-to-head on raw speed, but their generality means they stay useful as architectures evolve. That's likely why, even at a $21B valuation, most cloud providers still treat Etched as an H100 accelerator option rather than a full replacement.
FAQ
Q: Can Etched's Sohu chip replace Nvidia GPUs entirely?
Not yet. Sohu only handles transformer inference — training and graphics workloads still need general-purpose GPUs, so the two are complementary rather than a full swap.
Q: Can regular developers buy a Sohu chip today?
No. It's currently limited to enterprise contract customers like Jane Street, with no public sales or cloud API access yet.
Q: Is Etched's valuation surge driven by real tech or hype?
Both — Jane Street's production deployment is a concrete signal, but the broader AI hardware funding frenzy is also inflating multiples, so it's worth watching how customer adoption plays out.
好不好用，試了才知道 — you don't know if it's good until you try it.
Sources / 資料來源
- Inference chip startup Etched raises $700M more at $21B valuation - SiliconANGLE
- $21 billion AI chip startup Etched takes on Nvidia, poaches its engineers and lands Jane Street - Tech Startups
- AI chip startup Etched doubles valuation to $21 billion in under a month - Reuters via Investing.com
常見問題 FAQ
Etched的Sohu晶片可以取代Nvidia GPU嗎？
目前還不行。Sohu只能做Transformer推理，訓練、圖形運算等其他工作仍需依賴通用GPU。
一般開發者現在能買到Sohu晶片嗎？
還不行，目前僅開放給簽約企業客戶如Jane Street，尚未有公開販售或雲端API管道。
Etched估值暴漲是因為技術突破還是市場炒作？
兩者都有：Jane Street實際部署是硬指標，但AI硬體投資熱潮也讓市場願意用更高倍數給估值。
延伸閱讀 / Related Articles
- Cursor Origin評測：GitHub當機潮下的AI原生代管平台 | Cursor Origin Review: AI-Native Code Hosting Rivals GitHub
- ChatGPT青少年版評測：AI年齡預測真的擋得住嗎 | ChatGPT for Teens Review: Does Age Prediction Work?
- OpenRouter收購案評測：Stripe砸70億買下AI模型路由 | OpenRouter Acquisition Review: Stripe Pays $7B for AI Router
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
張貼留言