Fractile AI晶片評測：Anthropic砸2.5億美元賭SRAM推理 | Fractile Chip Review: Anthropic's $250M SRAM Inference Bet

By Kit 小克 | AI Tool Observer | 2026-08-25

🇹🇼 Fractile AI晶片評測：Anthropic砸2.5億美元賭SRAM推理

Fractile評測熱門話題來了：這家英國新創剛拿下Anthropic約2.5億美元的AI晶片訂單，估值一口氣衝上65億美元，是5月才募資時10億美元估值的6倍以上。這筆交易讓「SRAM推理晶片」成為本週最多人搜尋的AI硬體關鍵字，也讓Fractile擠進Groq、Cerebras、Etched這幾家專攻AI推理的晶片新創戰局。

Fractile是什麼？為什麼Anthropic願意砸2.5億美元？

Fractile是2022年由牛津大學機器人學者Walter Goodwin創立的英國AI晶片新創，專注解決一件事：讓大型語言模型「推理」（也就是產生答案）的速度更快、成本更低。Anthropic這筆初步訂單金額約2.5億美元，未來可能擴大，但晶片要到2027年才會正式出貨——換句話說，Anthropic現在賭的是還沒生產出來的矽晶片。

SRAM推理晶片跟傳統GPU差在哪？

關鍵在記憶體架構。傳統GPU得把運算單元和記憶體分開，資料要在兩者間來回搬動，這個瓶頸是AI推理變慢、變貴的主因之一。Fractile的做法是把運算和SRAM記憶體直接整合在同一顆晶片上，跳過搬資料的延遲。Fractile自己宣稱這套架構能讓大型語言模型跑得比現有硬體快上100倍，同時把營運成本砍低90%——這數字目前只是廠商說法，還沒有第三方基準測試驗證，但已經足以說服Anthropic下訂單。

這筆交易對AI晶片市場代表什麼？

Fractile不是唯一一家想啃Nvidia市占的AI晶片新創，Groq、Cerebras、Etched都在做類似的事——用客製化架構專攻推理，而不是像Nvidia GPU那樣訓練推理通吃。但Fractile這次特別的地方，是Anthropic在晶片還沒量產前就願意簽單，而且直接把估值推高6倍以上。這反映出一個現實：隨著AI晶片需求暴增，加上Nvidia先進製程產能被TSMC吃光、Samsung也調漲晶圓代工報價，各家AI實驗室已經開始「賭未來晶片」，而不是只買現貨。

值得關注的三個後續

晶片能否如期2027年量產：目前一切都還在紙上談兵階段
第三方基準測試：100倍速度、90%成本降低是否經得起驗證
Anthropic是否會擴大訂單：初步合約已預留擴大空間

對開發者跟企業用戶來說，這則新聞短期不會改變你用Claude的體驗，但長期如果Fractile的架構真的兌現承諾，代表AI服務的推理成本有機會大幅下降，間接反映在API定價上。

好不好用，試了才知道。

🇺🇸 Fractile Chip Review: Anthropic's $250M SRAM Inference Bet

Fractile just became the AI chip story everyone is searching for this week: the UK startup landed a roughly $250 million inference chip deal with Anthropic, and its valuation reportedly jumped to $6.5 billion — more than 6x the $1 billion mark it hit just three months ago. That puts Fractile in the same conversation as Groq, Cerebras, and Etched as the specialized inference-chip startups chasing Nvidia dominance.

What Is Fractile, and Why Did Anthropic Bet $250M on It?

Fractile is a UK chip startup founded in 2022 by Oxford roboticist Walter Goodwin, focused entirely on making AI inference — running an already-trained model to generate answers — faster and cheaper. Anthropic's initial commitment is around $250 million, with room to expand, but the chips aren't expected to ship until 2027. Anthropic is effectively betting on silicon that doesn't exist yet.

How Is SRAM Inference Different From a Regular GPU?

The core idea is memory architecture. Traditional GPUs separate compute from memory, forcing data to shuttle back and forth — a bottleneck that drives up both latency and cost during inference. Fractile instead puts compute directly on the same die as SRAM memory, skipping that round trip. Fractile claims the design can run large language models up to 100x faster than existing hardware while cutting operating costs by 90% — numbers that come from the company itself and have not been independently benchmarked yet, but were apparently convincing enough for Anthropic to sign.

What Does This Deal Mean for the AI Chip Market?

Fractile isn't alone in chasing Nvidia's share — Groq, Cerebras, and Etched are all building custom silicon optimized specifically for inference rather than general-purpose training-plus-inference GPUs. What makes this deal notable is that Anthropic signed before the chips are even in production, and that single contract alone re-priced Fractile's valuation more than 6x. It signals that as AI chip demand outpaces supply — with TSMC's advanced nodes sold out and Samsung raising foundry prices — labs are now betting on future silicon, not just buying what's available today.

Three Things Worth Watching

Whether the chip actually ships in 2027 — everything right now is still on paper
Independent benchmarks — can the 100x speed and 90% cost claims hold up outside Fractile's own numbers
Whether Anthropic expands the order — the initial deal reportedly leaves room to scale up

For developers and businesses, this won't change your Claude experience today. But if Fractile's architecture delivers on its claims, inference costs across the industry could drop meaningfully — and that tends to show up eventually in API pricing.

好不好用，試了才知道。

Sources / 資料來源

常見問題 FAQ

Fractile是什麼公司？

英國AI晶片新創，2022年由牛津大學機器人學者Walter Goodwin創立，專注用SRAM架構加速大型語言模型推理。

Anthropic為何投資Fractile？

看中其SRAM推理晶片架構，宣稱能大幅降低AI推理成本與延遲，即使晶片要2027年才量產，仍先簽下約2.5億美元訂單。

Fractile晶片跟Nvidia GPU差在哪？

Fractile把運算單元與SRAM記憶體整合在同一顆晶片上，省去資料搬運延遲；Nvidia GPU則是訓練與推理通吃的通用架構。

Fractile估值為何暴漲6倍？

5月募資時估值約10億美元，8月因拿下Anthropic訂單，傳出以65億美元估值募資6億美元，漲幅超過6倍。

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

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

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士