# AI capex — X 热门讨论 (2026-09-11 04:01 UTC)
## @onechancefreedm (EndGame Macro) · 09-10 22:41 · ♥33 ↻7 💬8 Basically, Oracle has to turn its $664 billion backlog into enough real cash over the next 15 to 19 years to cover about $125 billion of debt, $260 billion of data center lease commitments, roughly $32 billion of power and infrastructure obligations, plus future capex. That is already more than $417 billion of major obligations, before interest, dividends and normal operating costs.
There is no corporate precedent for this exact mix of scale, leverage and funding structure. Oracle spent $28.5 billion on capex in one quarter and still relied on $11.4 billion of customer financing plus nearly $20 billion of new equity. The simple point is that AI demand cannot just stay strong. It has to stay extremely strong for years, because the obligations do not disappear if demand slows. > 引用 @onechancefreedm: Oracle’s AI Empire Is Being Built With Customer Money, Equity and Debt
Oracle’s Q1 FY2027 results look exceptional at first glance.
Revenue rose 30% to $19.3 billion. Cloud revenue reached $11.6 billion, IaaS surged 121% to $7.4 billion, GAAP EPS rose to $1.56 and non GAAP EPS reached $1.92. RPO climbed to $664 billion.
The growth is real.
But the economics underneath it are changing.
Most of the incremental growth is now coming from infrastructure rather than Oracle’s traditional high margin software business. Infrastructure requires GPUs, data centers, electricity, depreciation and enormous upfront capital.
The Cash Flow Needs Context
Oracle reported $23.1 billion of operating cash flow.
But $11.36 billion came from customer prepayments containing a significant financing component. That represented almost half of reported OCF.
Oracle simultaneously spent $28.5 billion on capital expenditures, leaving reported free cash flow near negative $5.4 billion.
If the customer financing inflow is analytically removed, operating cash flow falls to roughly $11.7 billion and free cash flow approaches negative $16.8 billion.
That is not Oracle’s GAAP presentation. It simply shows how much cash the operating business generated without customer financing.
Oracle also sold nearly $20 billion of common equity during the quarter.
That matters.
The AI expansion is being funded by a combination of Oracle cash flow, customers and capital markets.
The Backlog Is Not Cash
The $664 billion RPO figure is impressive, but it is not near term cash sitting on the balance sheet.
At fiscal year end, only about 12% of the prior $638 billion RPO balance was expected to convert into revenue within 12 months.
Oracle therefore has to buy hardware, build capacity, secure power and commit to long term leases well before much of the associated revenue arrives.
That creates a duration mismatch.
It works extremely well if utilization remains high and major customers perform.
It becomes dangerous if demand weakens after the infrastructure has already been built.
The Hidden Leverage
Oracle carries roughly $125 billion of reported borrowings, with quarterly interest expense already at about $1.43 billion.
But conventional debt understates the commitment.
Oracle has also disclosed roughly $260 billion of additional long term data center lease commitments, largely spanning 15 to 19 years, alongside major power and infrastructure purchase obligations.
The old Oracle primarily sold high margin software.
The incremental Oracle is increasingly becoming a capital intensive infrastructure operator.
The Late Cycle Vulnerability
Oracle added 850MW of data center capacity and delivered more than 300,000 GPUs.
That creates enormous fixed costs and power requirements.
If energy prices rise and Oracle cannot fully pass them through, margins compress.
If it does pass them through, customers face higher compute costs and utilization can weaken.
The biggest breakpoints are straightforward.
• Higher power costs • A major AI customer delaying capacity • Utilization falling while depreciation and leases remain fixed • Credit spreads widening while financing needs remain high
Oracle has proven that AI demand is enormous.
It has not yet proven that this buildout can fund itself without continued customer advances and outside capital.
The most important combination this quarter was not just 121% IaaS growth or $664 billion of RPO.
It was $28.5 billion of capex, $11.4 billion of customer financing inside OCF and nearly $20 billion of newly issued equity.
That is where the real risk now sits. https://x.com/onechancefreedm/status/2098180144720384122
## @ElonEconomyX (ElonEconomy) · 09-11 00:22 · ♥30 ↻10 💬3 The SpaceX AI thesis is getting very fucking real.
> $100B+ annualized revenue target by year-end.
> Well over 2 GW of compute by the end of this year.
> 5–10 GW deployed next year.
> A new hosting deal worth roughly $1.1B/month starting December.
> Compute CapEx paying back in under a year.
> First orbital compute satellites targeted for next year.
And this is happening while Starship is moving toward full reusability. > 引用 @ElonEconomyX: Bret Johnsen just gave a pretty insane look into what SpaceX is actually building.
There is so much happening inside SpaceX right now. A few things he said really stood out:
> SpaceX’s entire strategy is vertical integration. Rockets, satellites, compute, power, software and eventually the customer.
> Starship is the foundation for everything. Cheaper, fully reusable Starship unlocks massive Starlink expansion, direct-to-device and orbital compute.
> Flight 14 is expected to be the first revenue-generating Starship flight, carrying production V3 Starlink satellites.
> SpaceX is targeting recovery and eventual rapid reuse of both Starship stages as soon as next year.
> Terrestrial AI compute is already becoming a massive business. SpaceX expects well over 2 GW by year-end and 5–10 GW deployed next year.
> A new hosting deal adds roughly $1.1B/month starting December, putting SpaceX on track for roughly $100B ARR annualized by year-end.
> New compute investments can pay back in under a year. Johnsen says they are seeing roughly $30–$50 per watt in monetization.
> Orbital compute could reach cost parity with terrestrial compute surprisingly soon because SpaceX controls the rockets, satellites and infrastructure.
> First orbital compute satellites are targeted for next year, with the system scaling significantly afterward.
> The Cursor acquisition is about speed. AI is moving too fast for SpaceX to rely entirely on organic development.
> Grok 4.6 is already ahead of 4.5, with 4.7 coming soon. SpaceX wants the best AI product, not just the best infrastructure.
> Starlink has evolved from “better than nothing” into enterprise-grade connectivity, with huge opportunities in aviation, maritime, rail and other mobility markets.
> Physical AI could create another enormous demand driver. Robots, autonomous cars and aircraft will need constant connectivity.
> Direct-to-device is progressing toward full 5G-quality service, with the next generation of satellites already being developed.
The bigger picture is fucking insane:
Starship → cheaper launch → massive satellite scale → Starlink + mobile + orbital compute → AI infrastructure → AI products → physical AI.
SpaceX isn’t building separate businesses.
They’re building one giant vertically integrated machine where every piece makes the others more powerful. https://x.com/ElonEconomyX/status/2098205577813905690
## @IEObserve (Intl Econ Observe) · 09-11 02:59 · ♥36 ↻6 💬1 Anthropic 公布的安全威脅報告其實有蠻多看起來很扯的內容，其中很多跟中國有關，某種程度上可以理解為什麼A社看起來很paranoid而且動不動就在喊中國威脅。 ---- #狸貓換太子：Moonshot（Kimi）與 DeepSeek 的中間人竊取與嚴重洩密
靜默轉發（Silent Relay）： 月之暗面（Moonshot）與 DeepSeek 均被發現將其自身產品使用者的對話請求，在用戶完全不知情的情況下偷偷轉發給美方的 Claude 模型處理，再將結果回傳給用戶。
兩家公司藉由這種中間人代理架構，一方面免去自身伺服器的運算成本，另一方面暗中側錄使用者的真實問答作為蒸餾訓練材料。
思維簽名重放攻擊（Thinking Signature Replay Attack）： 為了防止蒸餾，Anthropic 僅在 API 中回傳加密的「思維簽名」（Thinking Signature）。
Moonshot 與 DeepSeek 研發出跨會話重放技術，將保存的加密簽名在新會話中重新誘使 Claude 解碼為明文思考鏈，藉此完全瓦解技術防護。
不可控的敏感數據外洩： 由於這兩家廠商將使用者的原始輸入無差別轉發至美國伺服器，直接導致其國內用戶與企業的極度機密資訊裸奔：
#解放軍與國防軍工監視影像： 一名隸屬解放軍的用戶將成都市數百個天網監視器畫面（包含解放軍設施及中國電子科技集團 CETC 研究所周邊）輸入 Kimi 進行「異常行為分析」，該即時監控數據直接被 Moonshot 轉發至 Anthropic。
#俄羅斯國防部與外國藥廠機密： DeepSeek 轉發的請求中，包含了俄羅斯國防部外包 IT 工程師輸入的政府資料庫連線即時憑證；以及某跨國大藥廠在東南亞四國高達數千萬美元的資本支出（CapEx）工廠擴建預算明細。
#中國公安警務大數據： 公安局技術人員在開發「利用身分證號碼進行軌跡比對」的警務案件系統時，原始代碼與資料庫結構全數經由 DeepSeek 轉發至境外。
#次級黑市轉售：商湯（SenseTime）與 MiniMax 報告指出，中企對西方算力與模型能力的渴望催生了龐大的灰色「API 轉發站（Transfer Stations）」黑市。MiniMax 被查出設立無明顯關聯的海外空殼公司，專門搭建只提供 OpenAI 與 Anthropic 模型存取的轉發中繼服務，目的純粹是攔截真實使用者的提示詞來訓練自己的模型
而商湯科技（SenseTime）則直接向第三方黑市數據商採購這類被非法側錄的 Claude 對話數據，用於架構其自身的蒸餾訓練管線。
阿里巴巴（Alibaba / Qwen）的 #超大規模蒸餾 阿里巴巴發動了 Anthropic 監控史上規模最大的蒸餾攻擊。透過注入固定提示詞，強制模型在輸出終端答案前以標籤寫出完整的內部思考步驟。
這些思維鏈數據隨後被轉化為監督微調（SFT）資料集，直接灌入其主力開源模型 Qwen 3.5、3.6 與 3.7 的訓練流程中，涵蓋底層核心開發、長期複雜推理與 Agent 工具使用能力。
#非法的工業級蒸餾 所謂非法蒸餾，是指未經授權透過自動化程式向高階「教師模型」（如 Claude Opus）發送數百萬次具針對性的複雜邏輯提示，竊取其內部的推理軌跡（Reasoning Traces），並將這些高品質數據用於訓練本土的「學生模型」，從而在極短時間內、花費極低算力成本取得頂級代碼與推理能力。
#跨國鎮壓與涉台輿情情資處置（GTG-14021 與 GTG-14022） 地方公安局網警與警校研究生利用 Claude Code 操作輿情監控系統，追蹤海外異議人士（如知名社群帳號「李老師不是你老師」）。
#管控名單生成： 操作者透過反覆誘導（Re-prompting），規避模型的安全防護，直接由 AI 產出對境內 10 名維權訪民的具體管控方案，包括截訪、強制「約談」及通訊定位限制。
#涉台輿情重構： 承包商（GTG-14022）每天處理數十篇外媒與台灣媒體報導，將台灣的外交與文化活動自動標註為「對主權之威脅」，並將涉及人權的詞彙套上諷刺引號，轉化為向中共高層匯報的「三戰」（心理戰、法律戰、輿論戰）專報。
#敘利亞維吾爾人的跨國誘捕（GTG-14010） 一名不具備阿拉伯語能力的中國外包人員，利用 AI 監控逾 100 個 WhatsApp 群組與 Telegram 頻道，篩選出在敘利亞具有軍事背景或生活困頓的維吾爾族人。
透過 Claude 扮演中東方言軍事顧問，該操作者以流利的敘利亞阿拉伯語進行多日臥底對話，企圖以虛擬貨幣或資金報酬招募線人，並鎖定其在新疆境內家屬以施加壓力。 https://x.com/IEObserve/status/2098245092708679831
## @orrdavid (David Orr) · 09-10 23:57 · ♥38 ↻1 💬6 The reason hyperscales ran with max AI capex:
They realized 1+ year ago that ultimately electricity / building slots that use lots of electricity would be the bottleneck.
Once those locations are all locked up, which they probably are today already, the race will end. https://x.com/orrdavid/status/2098199315223064898