Mauer Wang

[The Coastal Journal @1CoastalJournal]

The Fed’s $25B monthly QT cap lets it reinvest excess Treasury maturities.

A recent $34.8B in buys over 2 days raised eyebrows—if sustained, it could mean stealth QE.

Some call it a shift, others say it’s routine SOMA reinvestment.

The key is whether it continues.

[Wall St Engine @wallstengine]

Morgan Stanley sees AI infrastructure spend topping $3T by 2028. That includes $2.6T on data centers (chips + servers), $210–330B on new power generation, and likely hundreds of billions more for grid upgrades.

[宏观边际MacroMargin @MacroMargin]

金融三巨头发布会，政策落地情况，
如期以货币政策为主，政治局直接提及的政策，大部分都落地了：
①降准50BP
②OMO降10BP→引导LPR降10BP
③降再贷款等结构性工具利率25BP
③降个人住房公积金贷款25BP
④扩容科技创新和技术改造再贷款 5000E→8000E
⑤设立5000E 服务消费和养老再贷款
⑥优化两项支持资本市场的货币政策工具，互换便利5000E 再贷款3000E 合并为8000E

[Nick Timiraos @NickTimiraos]

The PCE price index (due at 10 am on Wednesday) is expected to show core prices rose just 0.08% in March, according to forecasters who map the CPI, PPI and import prices into the PCE.

Headline prices are anticipated to have been flat (-0.01%) in March

[qinbafrank @qinbafrank]

美股盘后谷歌亮点还是很多的：1、营收同比12%的增长，但是利润和每股收益同比46%的增长，��由现金流的增长增速也能跟得上利润增速。降本增效的效果还是很明显，也可以认为AI在里面起到了相当大的作用。

2、公布了700亿美金的股票回购计划，延续了过往的回购力度，进一步提振市场信心。

3、资本支出持续扩大，谷歌在数据中心、芯片和其他AI基础设施上的投入持续扩大。第一季度资本支出飙升至172亿美元，高于去年同期的120亿美元，也略高于市场预估的171亿美元。公司预计2025年全年资本支出将达到750亿美元，用于建设数据中心等项目，较2024年的530亿美元显著增加。

在当下的市场环境下、谷歌超预期的财报还是能提振市场情绪的，特别是25年资本开支的增速并未放缓，在当下政策不确定性企业自然收缩开支的大背景下，继续扩大支出也是有助于情绪修复。当然一季度只是预期和情绪影响市场、并未实质性体现关税落地后的真正影响如何。

[Glen Gilmore @GlenGilmore]

Apple Vision Air: Are AR glasses finally coming that we'll actually wear? https://t.co/j3o84ZnD2l #Apple #AR #Wearables #Tech

[Jukan @jukan05]

Interpretation of the 245% Tariff on Chinese Goods – Seolhwa Choi, Meritz Securities

Today, the 245% tariff rate became a widely discussed topic in China.

This is because the White House statement released on April 15 (local U.S. time) listed the tariff rate on syringes and needles as 245%.

That figure is 100 percentage points higher than the previously familiar rate of 145%. While China has stated that it will not retaliate, it seems that such external pressure has only strengthened nationalistic sentiment among the Chinese public.

https://t.co/OOP2NA8jqB

[Trinh @Trinhnomics]

Let's talk about Trump tariffs. They are up, it's a question of whether how much, to whom, which sector rather than whether.

I want to clarify a simple fact that needs to be nailed home - trade and investment go together. Tariffs are a friction to trade & if you just take the idea that tariffs are going up (we'll discuss soon the details) then INVESTMENT IS GOING TO BE RESHUFFLED.

Half of global FDI is US driven. Global investment will be reshuffled. Okay, let's talk about first w/ who is LEAST TARGETED & we got to who is MOST.

[外汇交易员 @fxtrader]

https://t.co/EwkvsFTzRk

[财经数据库 @caijingshujuku]

网友火爆的长文：大的要来了！（收藏）
���先声明，个人川普黑，从来没说过他好话，甚至发表过他比咋们主席还傻逼言论，个人目前希望国内经济不要崩，大家还在国内的还能生活好一些，但这次川普关税我认为对中国可以说是一步杀招，拜登政策可以说是温水煮青蛙，以下都是���共产党客观评价，并且很多对于共产党评价还是亲戚体制内党员跟我说的，本人也不恨共产党，单纯讲下关税问题，会先介绍中国目前经济现况，经济结构，内需不足是如何产生的？
如何解决内需不足？中共会去解决内需问题么？内需不足情况下贸易战有什么影响？，大萧条失业下为什么中国不会再出现起义和革命党？这是共产党面临的转折点？为什么说报复关税是当下通缩经济萧条下不理智政治决策。中国人会受到多大的关税影响。
如果你对这些问题感兴趣可以看下，本人也不是专门媒体写文章，只是分享一下自己看法。
现在国内经济三驾马车，出口，消费，房地产，只有出口这两年还在支撑，因为国内不断内卷价格，工人工资也不涨，工作时间加长，但起码还有口饭吃，国内的消费能力除了公务员几乎没有了，中产随着失业降薪不稳定因素消费降级甚至开始用拼多多。

底层更不用说了，几亿中国人月薪不到3000，公务员事业编退休金七八千上万，几亿农村养老金一两百（农民都是交过公粮，粮食税，无偿去干活过的，三提五统就是从农民收入收取三次用于社会其他建设，等于把农民剥削到骨头里），中国大部分钱都通过垄断国企石油烟草卖土地银行（世界500强企业，中国上榜前几的都是国有银行），整个社会收的税收又拿去过度大基建（过程中可以大量贪污），在偏远地区云贵川��区没什人地方建高速，全国各地一天到晚修路，维稳费用接近或超过军费（肯定会有粉红觉得不信，事实在胡温时期共产党声誉最高时候，当初财政部公布过公共安全支出报告。

因此经济学人，中国经济学者一些智库人员，美国学术研究机构等做出研究对比过，得出中国维稳在胡温时期就超过军费，现在习主席时期，言论监管维稳哦，明显加强了，而且习近平上台后就不公布这些数据了跟青年失业率一样，直接不公布了，等修改符合中国特色统计才可能放出来），全国各地单位一天到晚派人去北京截上访，一次差旅费就得几万，这个也是我体制内亲戚给我讲他们单位真实发生的，说全国各地都这样，你想一年光是入北京就得花纳税人多���钱，更不要说守天桥或者异议人士小区家门口，配合着老干部的天价医疗，特供食品，维持这套体制还需要高工资让公务员听话（其实证明科举制度是中国历史集权统治的伟大发明。

考公就是科举另一个版本，中国年轻人都去考公了，都希望自己能加入这个体制从而可以剥削其他人），一个个国企都是关系户，世世代代相传领导岗位，小国企亏钱也不倒闭，财政补贴，还有就是补贴各种新能源芯片公司，不断有骗补贴公司，整个社会在这种榨取制度下，普通底层兜底为0，只能不断存钱抵御风险，所以没人敢消费。

消费占gdp比例只有40多，美国有80，一般国家都是六七十，人均gdp差不多的马来西亚，却拥有非常好的医疗兜底，免��140国家，中国人均gdp已经不低了，只是财富被共产党和听从党指挥的资本家白手套瓜分完了，如果让垃圾低效能国企淘汰，维稳费用去了，减少公务员工资，取消特供医疗食品，中国免费医疗和人均收入提高都有了，这样内需消费也起来了，也不用看美国脸色。

因为就没那么多贸易顺差，现实情况是政府企业是合作关系，员工什么权利都没，不能罢工，假的工会，没有选票，所以政府为什么要帮你严格执行劳动法，一个共产党国家竟然是世界上工作时长第一，公务员退休金跟农民差百倍，有点地狱笑话了，房地产就不用说了，都快入土了，除了上海等一线城市核心地区还暂时能稳住，二线房价等着腰斩，三线四线随着工作减少老龄化人口去世，人口萎缩房价等着归零吧。

[Christophe Barraud 🇫🇷 🇲🇨 @C_Barraud]

🇨🇳 #China Factory Activity Expands Before Expected Trump Tariffs - Bloomberg
https://t.co/T74p099u5R

[Jukan @jukan05]

250325_Server TAM updates: Revising Down AI Training Server Outlook - Goldman Sachs

(1) Goldman Sachs revised down its forecast for rack-level AI server shipments (144-GPU Equivalent):
•2025: From 31K → 19K
•2026: From 61K → 57K

(2) This revision reflects compounded uncertainties on both demand and supply sides.

(3) Consequently, we have revisited our outlook on Taiwanese Server ODMs and cooling system-related value chains, lowering target prices (TP) for Quanta, Hon Hai, FII, Wistron, AVC, and Auras by approximately 7-21%.

(4) Specifically, Quanta has been downgraded from “Buy” to “Neutral.”

(5) 2Q25 is expected to be a strong quarter as rack-level AI servers ramp up significantly after a prolonged waiting period.

(6) However, annual shipment expectations remain somewhat conservative due to potential shipment pressures arising from product transitions in 3Q25.

(7) Meanwhile, the forecast for High-Power AI Servers (8PU Equivalent) remains relatively unchanged:
•2025: From 437K → 423K
•2026: From 424K → 423K

(8) The slight decline to flat forecast for High-Power AI Servers reflects customers’ continued preference for flexible baseboard solutions and indicates these servers will not be fully replaced by rack-level AI servers.

(9) Demand for AI inference servers and general-purpose (GP) servers remains strong, maintaining a positive outlook.

(10) Therefore, expanding AI-driven applications (AI inference) and the recovery of aging server replacement cycles (GP servers) will act as growth drivers.

(11) Customer CAPEX projections:
• US: 2025 YoY +41%, 2026 YoY +15%
• China: 2025 YoY +26%, 2026 YoY +4%

(12) AI Training Servers show continued growth momentum, but compounded factors including slower shipment growth and uncertainties from product transitioning are impacting both demand and supply.

(13) In particular, shipment volumes may temporarily decline in 2H25 due to NVIDIA’s GPU platform transitioning to its next-generation model (Blackwell).

(14) Additionally, complexities in rack-level AI server production and assembly introduce uncertainties during production scaling, alongside ongoing debates about sustained training demand beyond DeepSeek.

(15) Tariff issues may also become significant variables.

(16) These dynamics necessitate geographic diversification in production lines for major Server ODMs, requiring additional time to stabilize yields in new regions.

(17) Ultimately, we forecast AI Training Server revenues to reach $160bn in 2025 (+30% YoY) and $260bn in 2026 (+63% YoY):
• 2025: Revised down from $179bn → $160bn
• 2026: Revised up from $248bn → $260bn

(18) Our market outlook revision reflects a 10.6% downward adjustment for 2025, but a 4.8% upward adjustment for 2026 compared to previous forecasts.

(19) AI Training Servers - Full Racks market forecast:
• 2025: Revised from $88bn → $54bn
• 2026: Revised from $182bn → $156bn

(20) AI Training Servers - High Power market forecast:
• 2025: Revised from $90bn → $106bn
• 2026: Revised from $66bn → $104bn

(21) Goldman Sachs significantly adjusted the 2026 forecast downward by 14.3% for the Full Racks market but substantially raised the High-Power forecast by 57.6% from $66bn to $104bn.

(22) AI Inference Server revenues are projected to grow YoY by +41% in 2025 and +39% in 2026, primarily driven by expanding AI applications.

(23) GP Servers are expected to see gradual recovery from aging replacement cycles, coupled with new CPU platform introductions, projecting modest YoY growth of +6% in 2025 and +4% in 2026.

[Time Horizon📊 @Time_HorizonX]

近几年“更积极的财政政策”对比

[Fred Chen @DrFrederickChen]

NVIDIA roadmap, thanks to SemiAnalysis https://t.co/accwwMGjD3

[Jukan @jukan05]

250318_Bi-weekly AI Thematic: NVIDIA’s CPO Switch - Nomura

(1) Semi Vision recently reported that NVDA is likely to unveil its first CPO switch at GTC 2025.

(2) We expect that the CPO switch will receive increasing attention over the next few years because it plays a crucial role in reducing power consumption and enhancing cost efficiency in AI Data Centers. *Based on Meta’s 24k-GPU AI cluster, computing power (AI GPUs and servers) accounted for 68.2% of CAPEX in 2024, while networking/interconnection accounted for 23.9%.

(3) Although major global companies such as NVIDIA, Broadcom, and TSMC are taking proactive stances on the CPO technology roadmap, due to (1) still high prices and (2) deployment issues, it is unlikely that we will see significant high growth in 2025.

(4) The global network switch market is structurally growing due to LLM training and data center expansion, and the shipment volume of high-speed switches—such as 800G models—is expected to drive future growth.

(5) Regarding the penetration rate of CPO in the AI network market, LightCounting expects that initial shipments of 1.6T optical interconnects will be small, and that by around 2028–2029, 3.2T will dominate the market.

(6) We estimate that the revenue-based penetration rate of CPO switches in the AI Data Center switch market will reach 2% in 2025, 22% in 2026, and 32% in 2030.

(7) In 2023, Broadcom launched the Tomahawk 5 Bailly, a 51.2T Ethernet switch CPO product.

(8) In December 2024, TSMC, in collaboration with Broadcom, successfully completed pilot production (3nm) of one of the key CPO technologies, the MRM (Micro Ring Modulator).

(9) Meanwhile, Semi Vision reported that NVIDIA is expected to launch its first CPO switch, the Quantum 3400 X800—based on InfiniBand—at GTC 2025, with mass production likely to start from July 2025.

(10) It was also added that the Ethernet-based CPO switch, Spectrum 5 X800, is expected to enter mass production in December 2025.

(11) The key components of a CPO switch include:
1) lasers,
2) optical chips,
3) optical engines,
4) switch ASICs,
5) optical components (Fiber Array Units),
6) shuttle boxes, and
7) MPO (Multi-Fiber Push-On) connectors, among others.

(12) We are focusing on leading companies in the FAU market, such as Suzhou (300394 CH EQUITY), which we believe will be direct beneficiaries of the long-term CPO trend.

(13) In addition, T&S (300570 CH) is recognized as a major supplier of shuttle boxes and MPO connectors for CPO switches, and Unisplendour (000938 CH) as well as Accelink (002281 CH) are actively participating in CPO switch solutions—factors that are expected to benefit from the expansion of AI infrastructure investment in China.

(14) Additionally, Broadcom’s CPO product roadmap is as follows:

(15) 1st Generation CPO: Tomahawk 4 “Humboldt”
■ In 2022, Broadcom, in collaboration with Tencent, launched the 1st generation CPO system, Tomahawk 4 “Humboldt”.
■ It was directly coupled (Co-Packaged) with four 3.2-Tbps Silicon Photonics Chiplet (SCIP) optical engines, with each optical engine containing approximately 250 optical components.
■ It is equipped with 32 transmit (Tx) and 32 receive (Rx) channels, supporting a signal rate of 100Gbps per channel.

(16) Latest CPO: Tomahawk 5 “Bailly”
■ Tomahawk 5 “Bailly” is a 51.2T Ethernet switch CPO product, competing in the same segment as Marvell’s Teralynx 10 (51.2T, supporting 64 ports at 800G).
■ It integrates eight 6.4-Tbps Silicon Photonics-based optical engines, coupled with the StrataXGS Tomahawk5 switch chip.
■ Compared to traditional plug-in optical transceivers, it reduces power consumption by 70% and improves silicon area efficiency by 8 times.
■ Each optical engine includes about 1,000 optical components (a 2× increase over Humboldt).
■ It integrates millions of transistors into a single optical engine on the same substrate as a complex logic chip, thereby minimizing the need for additional signal-conditioning circuitry.

$NVDA

[宏观边际MacroMargin @MacroMargin]

野村陆挺的这篇观点，标题很贴切。
我在年初的简报也分析过，一季度在宏观上完全没有压力，在产业的乐观情绪下，可以暂时不care宏观风险。但从二季度开始，国内政策效应衰减、“抢出口”消退、贸易战可能升级等等，市场将被迫提升对宏观风险的关注度

[Christophe Barraud 🇫🇷 🇲🇨 @C_Barraud]

🇨🇳 #China’s #Inflation Declines Below Zero for First Time in Year - Bloomberg
https://t.co/HmeaIhsyqq

[Eric Jhonsa @EricJhonsa]

Added a bit to my $AVGO and $CLS positions ahead of Broadcom's earnings. Think Marvell's issues are specific to its $AMZN relationship (namely, Alchip winning Trainium3 and Amazon's accelerator buys skewing towards $NVDA GPUs, because that's what most AWS customers want).

[Jukan @jukan05]

Morgan Stanley’s latest data on company-specific CoWoS demand…