自我改進AI新創Recursive Superintelligence簽4.1億美元AWS大單 | Self-Improving AI: Recursive Signs $410M AWS Compute Deal
By Kit 小克 | AI Tool Observer | 2026-07-30
🇹🇼 自我改進AI新創Recursive Superintelligence簽4.1億美元AWS大單
自我改進AI新創公司 Recursive Superintelligence 於 2026 年 7 月 28 日宣布，與 AWS 簽下一筆多年期、金額高達 4.1 億美元的雲端運算合約，用來訓練「能自己找出弱點並改進自身能力」的 AI 系統。這是今年目前規模最大的自我改進AI基礎建設投資案之一，也讓 Recursive Superintelligence 從 5 月才低調曝光的新創，一躍成為業界關注焦點。
什麼是 Recursive Superintelligence 的自我改進AI？
簡單說，這是一套讓 AI 自己評估表現、自己修正弱點的系統，目標是減少人類反覆調校模型的工作。Recursive Superintelligence 由前 Salesforce 首席科學家 Richard Socher 領軍，2026 年 5 月才以 6.5 億美元募資曝光，背後金主包括 Alphabet 旗下創投 GV、以及晶片巨頭 Nvidia 和 AMD。公司目前仍處於研發階段，尚未有正式上市產品。
AWS 4.1 億美元合約買了什麼？
這筆合約占了公司 6.5 億美元總資金的大部分，換來的是 AWS 客製化的運算基礎設施，而非傳統的人力擴編。重點包括：
- 多年期雲端運算容量保證，支撐自我改進AI所需的龐大訓練與推論算力
- AWS 表示會針對 Recursive 的需求「共同開發」基礎設施，未來可能複製給其他 AI 公司
- 公司規劃在 2026 年 10 月前推出第一批具體產品，驗證自我改進技術是否真的能用
自我改進AI為什麼值得關注，也值得存疑？
值得關注是因為「AI 自己改進自己」如果成真，等於把研發週期從人月縮短到自動化迴圈，對整個產業是質變。但值得存疑的地方也很明顯：這類說法過去已有多家新創提過，實際成果往往雷聲大雨點小；Recursive Superintelligence 目前只有資金和合約新聞，還沒有任何可驗證的產品或 benchmark 數據，10 月的交付承諾能不能兌現仍是未知數。自我改進系統一旦失控，也牽涉到 AI 安全上「能力遞歸提升」的老問題，業界對這類系統的監管與稽核機制目前仍不成熟。
Kit 小克怎麼看？
大手筆簽 4.1 億美元的雲端合約，說明資本市場還是願意賭「自我改進AI」這個故事，但錢砸下去不代表技術一定成立。比起「募了多少錢、簽了多大合約」，更該看 10 月產品上線後的實測表現——到時候才知道這套自我改進機制是噱頭還是真本事。
好不好用，試了才知道。
🇺🇸 Self-Improving AI: Recursive Signs $410M AWS Compute Deal
Self-improving AI startup Recursive Superintelligence announced on July 28, 2026 a multi-year, $410 million compute deal with AWS to train AI systems designed to identify their own weaknesses and improve their own capabilities. It is one of the largest infrastructure bets on self-improving AI so far this year, and it thrusts a startup that only left stealth in May into the center of industry attention.
What Is Recursive Superintelligence's Self-Improving AI?
In short, it is a system meant to let AI evaluate its own performance and fix its own gaps, cutting down on manual human tuning. Recursive Superintelligence is led by former Salesforce chief scientist Richard Socher, emerged from stealth in May 2026 with $650 million in funding, and counts Alphabet's venture arm GV, Nvidia, and AMD among its backers. The company is still in R&D — it has no shipped product yet.
What Does the $410M AWS Deal Actually Buy?
The deal consumes most of the company's $650 million war chest, and it buys custom AWS infrastructure rather than headcount. Key details:
- Multi-year guaranteed cloud compute capacity to power the training and inference self-improving AI needs
- AWS says it will co-develop infrastructure tailored to Recursive's workload, which could later be reused for other AI firms
- Recursive plans to ship its first tangible products by October 2026, the real test of whether self-improvement actually works
Why Does Self-Improving AI Matter — and Why Be Skeptical?
It matters because if AI can genuinely improve itself, it compresses R&D cycles from human-months into automated loops — a real step change for the industry. But skepticism is warranted: several startups have made similar claims before with underwhelming results, and right now Recursive Superintelligence has funding headlines but no verifiable product or benchmark data. Whether the October delivery promise holds up is still an open question. Self-improving systems also revive the old AI-safety concern of recursive capability gains, and industry oversight for this category is still immature.
Kit's Take
Signing a $410M compute deal shows capital markets are still willing to bet on the self-improving AI story, but a big check doesn't prove the technology works. Watch the October product launch, not the funding headline — that's when we'll know if the self-improvement mechanism is real or just a pitch.
好不好用，試了才知道 — the only way to know if it works is to try it.
Sources / 資料來源
- TechCrunch: Recursive Superintelligence signs $410M compute deal with Amazon
- AWS 官方新聞稿：Recursive Signs $410 Million Multi-Year Collaboration with AWS
- Data Center Dynamics: AWS secures $410m compute agreement with Recursive Superintelligence
常見問題 FAQ
Recursive Superintelligence是誰創辦的？
由前 Salesforce 首席科學家 Richard Socher 領軍，2026 年 5 月才以 6.5 億美元募資曝光，背後有 Alphabet 創投 GV、Nvidia 和 AMD 等金主。
4.1億美元的AWS合約跟公司總資金有什麼關係？
這筆合約佔了公司 6.5 億美元總募資的大部分，主要用於購買運算基礎設施，而非人力擴編。
自我改進AI現在有實際產品可以用嗎？
目前沒有。Recursive Superintelligence 仍處於研發階段，公司規劃在 2026 年 10 月前推出第一批產品。
這場交易為什麼值得存疑？
因為公司目前只有資金與合約新聞，沒有可驗證的產品或 benchmark 數據，過去也有多家新創喊出類似口號但成果有限。
延伸閱讀 / Related Articles
- Pacing the Frontier公開信：1100名AI員工籲美國踩煞車 | Pacing the Frontier: 1,100+ AI Workers Urge US to Slow Down
- FCC封殺中國人形機器人：Unitree後門疑慮掀美國禁進口 | FCC Bans Chinese Humanoid Robots Over Backdoor Risk
- MCP新規範解析：告別Session，AI工具協定轉無狀態 | MCP 2026-07-28 Spec: AI Tool Protocol Goes Stateless
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