# AI capex — X 热门讨论 (2026-09-22 08:15 UTC)
## @ShanuMathew93 (Shanu Mathew) · 09-21 14:50 · ♥30 ↻9 💬2 Mauboussin and Callahan’s paper presents a favorable aggregate outlook for hyperscaler returns, with incremental returns troughing around 23% versus an estimated 8% cost of capital. But the company paths differ substantially, and the projected recovery depends on consensus being right.
Forward figures below use consensus as of August 31, 2026. >ROIC: returns on the accumulated capital base >ROIC compares after-tax operating profit with invested capital. It helps assess overall business economics, although consolidated returns can conceal differences between existing operations and new AI investment. >Amazon: recovery from mid-2027. >Alphabet: approximately 30% through 2031. >Microsoft: remains high throughout. >Meta: approximately 40% in early 2025 falls toward 20% by 2031. >Oracle: 13% in Q2 2026, then nearly doubles from the lows. >Nvidia: approximately 48% in Q2 2023 rises toward 200% in 2028. >Micron: approximately 89% in Q1 2027 falls toward 55% by end-2028. >Apple: remains elevated, with recovery during 2026–28.
ROIIC: returns associated with additional investment Incremental returns help assess what additional capital is producing. The paper uses three-year changes in NOPAT with a one-year investment lag. >Amazon: ~35% in Q1’26 → ~15% in Q4’27 → ~45% in 2031. >Alphabet: ~70% in Q4’23 → ~30% in Q4’27 → ~35% in 2031. >Microsoft: ~23% trough in Q2’27 → ~60% in 2030. >Meta: ~70% in Q1’26 → ~16% in Q4’27 → ~13% in 2031. >Oracle: −3% in Q2’24 → ~50% in 2030. >Nvidia: >600% in Q1’26 → below 200% in H2’29. >Micron: >1,000% in Q4’26 → ~35% by end-2029. >Apple: volatile, generally near or above 100% from Q3’25 through late 2028.
If elevated capex persists longer, capacity takes longer to monetize, or incremental cash generation disappoints, both the timing and magnitude of the recovery change. The combined outlook looks attractive conditional on those forecasts; evidence that new capacity is earning the expected returns still needs to arrive. https://x.com/ShanuMathew93/status/2102047714649206869
## @j2kithforever (J2KH_forever) · 09-21 19:58 · ♥30 ↻1 💬14 $IREN is not currently a cloud company; rather, it is essentially a promissory note that won't mature until 2027.
The core of $IREN’s fundamentals isn't about "current profitability," but whether the combination of "5GW of power capacity + sold 2026 capacity + customer prepayment financing" can translate into auditable AI cloud revenue by late 2026 or 2027. What has already been validated: delivery of Horizon 1, Q4 AI revenue surpassing mining revenue, rising contract unit prices, and financing covering the majority of GPU costs. What remains to be validated: on-schedule delivery of Horizons 2–4, achieving 0.8GW of IT capacity by 2027, avoiding severe dilution from massive CAPEX, and realizing healthy profit margins after scaling up.
This is a high-growth, high-leverage narrative stock: the path to fundamental improvement is clear, yet that path is extremely sensitive to construction timelines and financing conditions. It is not "already a mature cloud company," but rather a company "midway through transformation, with the payoff just beginning."
This post is just the starting point. If any of the three key factors—timelines, dilution, or profit margins—change, I will provide updates—so be sure to follow me if you're interested! https://x.com/j2kithforever/status/2102125446687306074
## @youngstockuser (美股投资young) · 09-21 19:48 · ♥30 ↻1 💬14 $AVGO这轮最容易被忽略的风险，不是NVDA突然来抢它的生意。
Broadcom现在AI Semiconductor已经做到单季167亿美元，公司甚至把FY2027 AI芯片收入预期拉到大约1150亿美元，2028年看到2300亿美元。增长已经非常夸张。
问题也出在这里。
这门生意越来越依赖Google、Meta、OpenAI这类Hyperscaler继续疯狂做Custom Chip。
只要这些公司AI CapEx继续涨，AVGO的Custom Accelerator和Networking都能一起吃。
但如果其中几家客户开始觉得自研芯片ROI没那么高，或者订单被$MRVL之类的竞争对手抢走，市场对AVGO未来两年的盈利预期会很快重新算。
Google最近已经把一部分Custom Chip订单给了Marvell，这就是很现实的竞争。
所以Broadcom现在最大的优势也是最大的风险：
订单特别大，但客户也特别大。
几个Hyperscaler一张订单就能让Revenue多几十亿美元，同样的，一次设计丢单也不会是小事。
VMware当然能提供现金流托底，所以AVGO不是纯AI Story Stock。
但现在这个估值下，股价真正需要的是AI业务继续超预期。
未来看$AVGO，与其天天盯NVDA，不如多盯：
Google、Meta、OpenAI下一代芯片最后交给谁做。 https://x.com/youngstockuser/status/2102122693076136378
## @kubotamas (久保田 雅也@Coalis) · 09-21 22:40 · ♥33 ↻4 💬0 AnthropicのAI減速宣言への痛烈批判
OpenAIやAnthropicは世界最高モデルを競っている間に、それを持続的なモートへ変換する時間が足りない
だからフロンティア開発を減速すれば、computeをinferenceへ回し、 価格を下げてProductを作り、収益化してCapexを回収する時間が得られるのだと
更に、OpenAIが再びフロンティアをリードし始めたタイミングで政府介入による減速を主張。安全性への懸念はリアルでも、その政策提言がAnthropicの競争上の利益とあまりにもきれいに一致しているとの皮肉
https://t.co/Fo9QCx07ix https://x.com/kubotamas/status/2102166060678238551