一覺起來，外資那邊也炸了...昨晚財報+業績會發酵了一波之後，外資開始大肆吹阿里（甚至上週關於ant的小作文，也有人郵件裏提了一嘴...）...

比如下面高盛的這篇，現在昨晚的阿里業績會後，有些說法開始把“ABC”塑造成“A for AI, B for BABA, C for China”

圖片

外資吹起來，比我們還猛 ...慢慢的，可能老外也要FOMO了...

這篇裏面問的是下一步能不能看20x PE, 我簡單摘抄一些；

圖片

昨天另外比較收到關注的是微軟CEO Satya Nadella的一個podcase，裏面說到的很多東西，和阿里管理層頗爲相似（比如雲計算是目前最明顯的AI貨幣化模式）。我這裏用AI簡單翻譯一些。

1/ 納德拉認爲雲計算公司將是AI的贏家（"在此之後，局面變得有些模糊。有人會說贏家通喫——但我並不這麼認爲"）。潛臺詞是微軟並不需要依賴模型本身，真正的變現能力在於基礎設施。

He views Hyper-scalers as the winner in AI world (‘Then after that, it becomes a little fuzzy. You could say, hey, there is a winner-take-all model- I just don't see it.’).  The subtext is MSFT doesn't need the models but infrastructure is the true monetization.

2/雖然大模型日益同質化，但AI對算力的"飢渴"不會消退——不僅訓練需要算力，推理階段和Agents的指數級應用擴張同樣需要。模型仍需運行在需要狀態存儲+存儲+計算的超大規模算力上。你需要在邊緣部署"艦隊式"分佈式數據中心，因爲"不可能所有服務都從得克薩斯州提供"。這對微軟、亞馬遜和Cloudflare（NET）等類型公司構成利好。

LLMs are getting more commoditized but AI remains ‘hungry’ for compute workloads, not just training but test-time compute and then Agents expand use 'exponentially'.  Models will still need to run on hyper-scale compute which requires state + storage + compute.  You need a 'fleet' of distributed Data Centers at the edge because you 'can't serve everything from Texas'.  Bullish for MSFT, AMZN & NET types.

3/ 最犀利的觀點："宣稱達到某個AGI里程碑只是無意義的基準測試投機。真正的基準是推動全球增長+10%"。納德拉認爲可能迎來新一輪工業革命，推動發達市場增長+7-10%（或經通脹調整後+5%）。最大贏家將不限於科技公司，而是更廣泛採用這種新型AI的行業。

This is the fire quote:  'Claiming some AGI milestone is nonsensical benchmark hacking.  The REAL benchmark is a world growing +10%'.  Nadella sees another Industrial revolution possible that drives +7-10% growth in developed markets or +5% inflation adjusted.  And the big winners will NOT just be Tech companies but broader industries that are using this new abundant intelligence commodity.  Bullish for AI Adopters MSXXAIAD.

4/ 微軟明年資本支出將超千億美元，但相比Sam、馬斯克、扎克伯格和孫正義等"大力出奇跡"派，納德拉的表態仍最理性。此前微軟放棄"星際之門"預訓練項目已見端倪。他承認超前投資必要性，但也強調最終需供需匹配，若供給端過度炒作卻無法轉化爲客戶價值就會"脫軌"。

MSFT spending $100B+ capex next year but Nadella still sounds most rational on capex relative to Sam Altman, Elon, Zuck and Masa types in the build build build camp.  We saw this with MSFT passing on the Stargate pre-training build so not a surprise.  Nadella says they've invested a lot ahead of demand but at some point you have to match supply and demand and you can 'go off the rails if you're hyping yourself up' on the supply side without translating that to customer value.

5/ 過度建設風險？他明確使用這個詞："會出現過度建設...所有人都在賽跑"，雖然微軟仍需"可能指數級"投入，但他對2027-2年前後產能上線後成爲"出租方"的前景感到興奮——這意味着算力價格將下降。 (180K: 我們本土現在摸着北美投資節奏過河，ROI會更高一些，也會少一些浪費）

Overbuild risk?  He uses that word, saying ‘there will be overbuild .. everybody’s going to race’, and though MSFT will still need to invest, 'maybe exponentially', he's 'so excited to be a leaser' with capacity coming online in 2027-2, meaning prices will come down for compute.

6/ 儘管與OpenAI關係出現裂痕，但大語言模型不會贏家通喫。他認爲消費市場可能贏家通喫，但企業客戶必然需要多供應商，模型領域也將如此。

LLMs are not a winner-take-all despite the (fraying?) OpenAI relationship.  He says consumer markets can be winner-take-all but anywhere where the buyer is an enterprise they will want multiple suppliers so he thinks that will happen on the model side.

7/ 傑文斯悖論："任何突破（如deepseek的能效邊界突破）改變單位token性能曲線時，需求反而會激增"。

Jevon’s Paradox:   ‘Anytime there's a breakthrough, like even what DeepSeek did, with the efficient frontier of performance per token changes, the curve gets bent, and the frontier moves. That just brings more demand.’