amanda

[amanda @amandaskylo]

@fupenglondon “逻辑的自我强化“指的是商业闭环的flywheel? 飞轮闭环（算力→模型能力→终端付费→再投入算力）如果真的转起来，受益最大的其实是模型/应用层的公司。下游商业化跑通、现金流充裕之后会自研芯片，绕过英伟达。

[amanda @amandaskylo]

@fupenglondon “大模型训练能不能切换到算力商业化盈利”，这是说��来很多年现金流能不能按照现在的斜率持续增长，也就是分母端的g 吗。可是表外承诺增加信用风险，r 也上去了

[amanda @amandaskylo]

@tuolaji2024 Fake news.
Commerce Secretary Howard Lutnick has publicly warned Apple against buying memory from Chinese suppliers including CXMT, and confirmed the administration communicated its concerns directly to Apple

[amanda @amandaskylo]

@jukan05 “Fits Nvidia’s interest” assumes the wrong objective function. More rack shipments only helps Nvidia if it doesn’t cannibalize revenue/margin. HBM is high-margin, per rack — swapping it for volume isn’t obviously “rational” unless you know the margin math on the trade.

[amanda @amandaskylo]

@jukan05 Legitimate maybe as a narrow caveat on the growth assumptions priced into HBM-exposed stocks, but it would need to survive the pricing-power counterpoint to even support that narrow claim, and it doesn’t scale up to justify a blanket bearish call on the whole memory sector.

[amanda @amandaskylo]

@AndrewM61370616 @jukan05 Why not the opposite. Algorithmic efficiency reduces hardware demand. Every efficiency gain is reinvested into training even larger models.

This is essentially Jevons paradox .

[amanda @amandaskylo]

@tablescraps42 @KobeissiLetter Or we don’t know.

[amanda @amandaskylo]

@fi56622380 @TiezhuCrypto Suppose non-depreciation OPEX is 200*.35=70B, depreciation 87.5B, we have EBIT = 42.5 B. At 15% tax rate , NOPAT is 36.125 B. 对于一个投入 $350B 资本开支的项目 ROIC = 36.125/350 =10.3%. It passed the WACC hurdle rate, which for hyperscalers typically ranges between 7.5% and 9.5%

[amanda @amandaskylo]

@fi56622380 @TiezhuCrypto 这个算力 GW 对应 CapEx 的模型很清晰。Back of Envelope calculation for required revenue growth to justify the Capex. GCP Q2 2026 单季营收为 $24.8B, annual run rate about 24.8*4 =99.2 B. 按 4 年折旧算，单年折旧 350/4~$87.5B。

[amanda @amandaskylo]

@fi56622380 @TiezhuCrypto With 100% increase of revenue growth, depreciation percentage of revenue is 87.5/200 =43.8%. Typical non-depreciation operating expenses (power, water, and facility maintenance) for cloud vendors represent around 30% to 35% of revenue, leaving about 20%+ 的 EBIT margin.

[Damnang @damnang2]

https://t.co/0Xal4VwfH7

[Macro_Lin @LinQingV]

这一轮AI半导体设备周期的重心在存储。按新增晶圆产能算，DRAM、NAND、HBM合计占89%，先进逻辑只有11%。这个数出自Bernstein 7月20号的深度报告，标题就叫"要达到AI涅槃，我们需要多少WFE"。

设备三巨头从高点回撤了25-30%，年初至今的涨幅还在80%以上。报告把AI数据中心的建设规划换算成半导体设备需求，再对照跌完的股价，看里面还剩多少预期没兑现。

美国数据中心活跃容量6月底50.7GW，同比增29%，这部分是已经通电运行的。管线容量338GW，同比增179%，包含了规划、拿地、审批中的所有项目，最终能落地多少没人说得准。设备股的多空分歧就在这338GW上。

换算框架值得一看，Vera Rubin机架每台消耗约65片晶圆，逻辑只有8片，HBM 19片、DRAM 22片、NAND 16片。机架外面还有agentic CPU、每颗CPU配套的0.5TB DDR、数据中心NAND。加总下来，每增加1GW/年的算力部署能力，需要4.6万片/月的新增晶圆产能，对应75-80亿美元设备投资。

存储线和逻辑线的设备构成差别很大，沉积刻蚀占比高，EUV层数少，光刻强度远低于先进逻辑。所以这轮的受益格局和上一轮先进制程驱动的周期不一样。应用材料存储设备敞口全行业最高，被列为首选，闪迪这种NAND纯度最高的标的也拿到了很高的目标价。

总量测算的基准情景，是到2030年数据中心的年新增部署能力比2026年高出50GW。当前基线约20GW/年，也就是2030年单年部署70GW。要支撑这个速度，2027-29累计需要7,360亿美元WFE，逐年2,000亿、2,450亿、2,910亿。2,000亿前两年还被当成行业的理想化天花板，这份报告里成了2027年的起点。乐��情景给到增量100GW，对应2029年单年5,420亿。

估值这块，按当前市场共识的盈利预测，三巨头2029年市盈率在24-29倍。换成基准情景推出来的盈利，EPS比共识高40-60%，市盈率降到15-21倍。增量75GW的情景下EPS翻倍，市盈率只剩11-15倍。现在的估值贵不贵，取决于你信哪个情景。

[Damnang @damnang2]

This is an article every memory investor should read at least once.

Full article: https://t.co/eBvRzLf5Hi

[Bg2 Pod @BG2Pod]

BG2 w/ Gavin Baker. The SpaceX IPO, Fable 5 / Mythos, AI Capex Update & Market Check. 🚀💰 @BG2Pod @altcap @GavinSBaker @_clarktang

[amanda @amandaskylo]

@its_mizuko_ 21

[Anatoli Kopadze @AnatoliKopadze]

instead of watching 2 hours of Netflix tonight, watch this Stanford lecture

it's the clearest explanation I've seen of how ChatGPT and Claude actually work

useful whether you've never touched AI in your life or have been using it every day for the past year

I took the key ideas and turned them into a practical guide on how to actually get 100% out of Claude

find it below

[amanda @amandaskylo]

@ThePowerAudit @Eyalo365 Very insightful. Thanks.

[amanda @amandaskylo]

Simple but useful

[amanda @amandaskylo]

@confinedape You’re right that this might be expensive underwriting with understated jump risk, but we don’t know yet, and that uncertainty is being weighed against the opportunity cost of being out. You’d rather underwrite 10-year IG paper for a company with $70B in FCF than sit in cash.

[amanda @amandaskylo]

@confinedape The IPO/unlock wave is your most concrete catalyst, and probably your best point. But international and money market capital is sitting on the sidelines at scale. Rate cuts don’t have to precede the bid; anticipation of cuts is enough to rotate $6T in money market funds.