HAP retweeted
一堆投资大佬都在发出泡沫警告，但霍华德·马克斯却不以为然，而恰恰马克斯的观点是我认为最值得关注的。聊下马克斯的核心观点：他目前并不认为市场已经达到真正意义上的泡沫状态。他的核心逻辑在于：泡沫的本质特征之一是投资者对风险的“浑然不觉”（obliviousness），而当华尔街上下都在警惕泡沫风险时，这种“浑然不觉”恰恰是缺席的。“如果足够多的人都在谨慎地追问某个领域是否存在泡沫，那么完全有理由认为，即便它是泡沫，市场也会表现得很克制，”霍华德·马克斯说。
马克斯把泡沫描述为“非理性繁荣（irrational exuberance）、轻率冒进（recklessness），以及浑然不觉（obliviousness）”。对照这一框架，霍华德·马克斯目前的评估是：投资者确实表现出对进一步上涨的“错失恐惧”（FOMO），但对泡沫风险的持续讨论本身，说明市场并未陷入“浑然不觉“的状态；而在另一个信号——投资者鲁莽行为方面，信号同样是混杂的。
为什么马克斯的观点值得关注？作为橡树资本的创始人，《周期》这本书的作者，他对基本面的理解还是市场情绪应该是顶级的。
更关键的是，过去20年里他辉煌的战绩：
1）2000年1月，马克斯发了他那篇后来被广为流传的《bubble.com》备忘录，当时纳斯达克尚未见顶，马克斯在文中将当时科技股狂热与历史上的南海泡沫（South Sea Bubble）等进行类比，警告市场存在明显的投机过热迹，00年3月纳指见顶；
2）2007年7月，马克斯发布《It’s All Good》备忘录，是他在全球金融危机爆发前最重要的预警备忘录。07年10月美股开启了为期一年半的暴跌；
3）2008年9月雷曼破产后，市场陷入末日叙事，认为金融体系可能崩溃。马克斯在08年9月的《Nobody Knows》）和10月《The Limits to Negativism》两篇备忘录中指出，悲观已过度。并且他的橡树资本动用此前准备110亿美元“储备基金”开始分批逐步减仓抄底；
4）2012年3月，2000–2011年标普500年化回报仅约0.55%，GFC后负面情绪浓厚，马克斯建议对股票采取更积极态度（相对当时极度负面的环境），不期待2000年式大涨，但认为正面情景不难想象；
5）2020年3月新冠爆发后市场快速下跌（标普500从2月高点到3月低点跌约1/3）。马克斯在20年3月中《Nobody Knows II》及3月19日客户更新中指出，未来充满未知，但无知不等于必须无所作为。橡树明确拒绝“等到底部再买”，而是在能以低价获得价值时买入。恐慌性抛售（包括被迫卖出）创造了机会，在市场最混乱时开始配置。
马克斯自己也承认，这五次大判断是他约50年职业生涯中的主要成功案例。
而他的《周期》那本书则是我认为可以被反复拿来细读的一本投资圣经，很早也在X上xitter.cf/qinbafrank/status/1837…有推荐过。
这次他从市场反身性的角度来看市场并没有到泡沫状态，昨晚的推文也聊到对这次美股调整的定性是：是高速增长叙事下的正常市场波动，而非泡沫破裂的序曲。殊途同归
之前Grok说我有马克斯的周期概率的痕迹，那可太荣幸了，xitter.cf/qinbafrank/status/2092…
过去七个交易日内美股的这一波抛售是如何发生的？10月8日前后开始，市场对AI高增长叙事和资本开支可持续性的质疑逐步升温，多个事件叠加形成共振：
1）10月8日左右：OpenAI营收口径差引发第一波抛售
OpenAI向投资者披露，截至9月底年化营收约500亿美元，低于市场此前流传的约700亿美元水平。口径差直接打压了“AI变现超预期”的乐观预期，科技股高成交量下行，纳指与标普出现明显回撤；
2）紧接着就是产业端“资本是瓶颈”表态
英伟达、博通等CEO在公开场合强调，AI发展瓶颈不仅是产能，更是资本（融资能力与资本开支节奏）。博通被报为Anthropic和OpenAI提供近千亿美元融资担保，叠加SpaceX等巨额芯片采购计划，市场开始担心AI基建的债务与现金流错配；
3）10月13
9月CPI同比升至3.7%（汽油价格大幅上涨贡献过半），核心也偏强，市场对年内加息概率短暂飙升，科技股进一步承压；
4）同一天
Anthropic投资者日披露三季度营收，增速明显放缓。公司称调整后盈利、现金流为正、计算效率改善，但ARR推算低于部分投资者此前年底逾1000亿美元的乐观预期，IPO估值讨论从约2万亿美元向1.5万亿美元附近下修。亚洲芯片股与美股AI/半导体板块同步重挫。
5）10月14号
9月PPI大幅超预期，能源是主因，但核心也偏高，显示通胀有扩散迹象。这加剧了“滞胀”或通胀顽固、央行被迫更紧的担忧。
6）也是当天 同步的AI信用/融资担忧进一步加重
CoreWeave等债务沉重的AI基础设施公司CDS（信用违约互换）大幅走阔至1100基点以上（隐含较高违约风险），债券收益率一度冲到14%-16%。甲骨文CDS也升至约300基点附近，创纪录水平。科技板块CDS整体走阔，反映市场质疑巨额AI资本开支的回报与融资可持续性。
7）而10月13之后几天
霍尔木兹海峡油轮/商船攻击明显增多（伊朗相关导弹、水雷等），油流量受扰。布伦特原油一度触及110美元附近（为5月以来首次），进一步推升通胀担忧并压制风险资产。这又成了市场额外的避险与成本压力。
这些因素并非孤立：
1）OpenAI收入口径问题 + CEO资本瓶颈警示 +Athropic的营收增速放缓 →质疑AI回报与融资；
2）叠加CPI/PPI超预期（能源驱动+扩散）与霍尔木兹攻击推升油价 → 强化通胀与利率担忧；
3）但最近两个交易日Coreweave的债券被抛售、市场对AI资本开支的可持续性担忧情绪暴增。
这是“营收口径差 + 增速放缓信号 + 资本/债务担忧 + 通胀与油价冲击”的共振，市场从“AI高增长无限持续”的定价，转向对变现速度、融资成本和资本开支回报周期的重新审视。结果就是对高估值、高资本开支的AI相关股票的集中抛售。
9号推文xitter.cf/qinbafrank/status/2109…聊到认为近期高Beta容易📉的风险、算是比较早有预警，
而这两天crwv债务问题无疑成了市场对阵眼，昨天这里xitter.cf/qinbafrank/status/2111…有聊过crwv的真实压力和经营状况，不得不去思考13-15号crwv的债券收益率暴涨也一丝操纵的味道xitter.cf/qinbafrank/status/2111…。
但毕竟抛售已经发生。
对此次的抛售的定性是什么？
个人角度此轮波动是高速增长叙事下的正常市场波动，而非泡沫破裂的序曲。
核心是什么？
1）核心是产业最终要回到巨大的资本开支未来能不能赚回来？多快赚回来？高债务公司能不能扛到现金流为正的时候？
其实二季度云厂商的财报季已经显现了AI商业化是在加速的迹象，看微软、谷歌和亚马逊的财报就能看出来。这本质是上是一个时间差xitter.cf/qinbafrank/status/2081…资本开支先花钱，商业化逐步跟上
2）但有一点，市场还没转过来，闭源大模型厂商的营收已经不能代表AI商业化的全部了，最早7月就有聊到超大云厂商更能衡量AI商业化的全部。前两天这里xitter.cf/qinbafrank/status/2110…又再聊一次，可以看到详细逻辑。
3）所以之前一直聊到产业上的关键就是几家超大云厂商的财报，但现在市场要看的已经远不是收入增速。
二季度微软、亚马逊和谷歌财报数据有力支持商业化和云需求正在扩张。但
也提高了市场预期门槛：投资者已经看到加速，接下来要求更强的持续性和现金回报证据。
如果云巨头的云收入、客户token消耗量快速增长且大幅度超预期，营收增速开始超过capex增速，继续增本开支，现金流没有超预期恶化。这种情境下市场情绪才能得到真正修复。
如果投入继续增加、回报却开始滞后，市场就会从“增长叙事”转向“更严苛的资本回报审查”。
当然个人认为前者是大概率的
4）短期来看市场可能还没稳住、17、18号周末两天中东局面并不稳定、伊朗继续袭击油轮的力度在加大。再加上悲观的情绪的发酵，也会引发市场的进一步担忧。19号开始新一周的前半周市场可能还是不稳。
而20-21spacex、meta的财报就是本周的关键事件。28-29号微软、谷歌和亚马逊财报是此次核心
cnn恐慌贪婪已经跌到了17了xitter.cf/qinbafrank/status/2111…，个位数恐惧极值一步之遥，如果再往下杀杀市场情绪恐惧到极值、那么我们就不应该再恐惧了。
正如16号推文xitter.cf/qinbafrank/status/2110…聊到、已经能看到低买的机会了。
本条由@bitget_zh赞助，「Bitget 买美股：秒级入场，丝滑交易
HAP retweeted
There is a MASSIVE and growing compute glut.
Anthropic continues to mark all unsold compute as R&D to keep up the illusion of a shortage and the adjust out massive losses to pretend to have profitable inference.
You are a moron if you can’t see through this nonsense.
If I'm not mistaken, Anthropic's 85% gross margin and >40% compute margin from this Bloomberg story implies that 70-75% of their compute is now going to R&D.
bloomberg.com/news/articles/…
HAP retweeted
AI’s funding equation is starting to crack.
Hyperscaler's FCF is expected to turn negative in 2026 and getting worse next year, as AI capex heads past $1T. Meanwhile, semis prices hit 20-year highs and credit conditions tighten—most visibly in Oracle’s surging CDS.
Oracle’s 5Y CDS has gone from ~50bp to ~280bp in a year, and IG tech spreads are widening with it.
The AI buildout was supposed to be self-funding. It isn’t anymore.
Who will be the first on the hook?
$ORCL $MSFT $GOOGL $AMZN $META
Jefferies: The AI cash-flow squeeze.
The AI spending boom is starting to leave a serious mark on cash flow.
Jefferies notes that aggregate free cash flow across the four major hyperscalers collapsed from $60B in Q4 2025 to just $7B in Q2 2026.
Consensus expects it to turn negative at $13B in Q3, while quarterly capex is projected to hit $200B.
The spending is real. The cash generation is increasingly under pressure.
The next question is how much longer investors will tolerate the divergence.
$AMZN $META $GOOG $MSFT #DataCenter #AI #Hyperscalers
HAP retweeted
Two follow-ups on the power supply-demand analysis (thanks for all the feedback and DMs):
1/ There are several different ways to define and measure data center capacity (GW). It's important to distinguish between them:
A. Global total DC capacity: ~130 GW by YE2026, with annual additions of 30/50/70 GW in 2026/27/28E, based on SemiAnalysis estimates. [Chart in original post]
B. CSP + Neocloud capacity: ~63 GW by YE2026, with annual additions of 22/33/37 GW in 2026/27/28E, based on BofA estimates. [Chart below]
C. AI DC capacity: ~37 GW by YE2026, with annual additions of 19/43/35 GW in 2026/27/28E, based on SemiAnalysis estimates. [Chart below]
D. US total DC capacity: ~57 GW by YE2026, with annual additions of ~17/23 GW in subsequent years, based on Funda estimates. [Chart in original post]
2/ Global supply vs. where capacity actually gets built
Many agree that globally, ~50 GW of new DC capacity could come online in 2027, with the US contributing ~20 GW.
However, several people raised some important points that I 100% agree with:
A. National interests: The US historically accounted for ~70% of new DC capacity. There are strong national security and economic incentives to keep a larger share of future capacity onshore.
B. Latency: esp. on scale across training.
C. Data privacy
The distinction is between theoretical global supply and where capacity can realistically be deployed and utilized. The rest of the world may have enough power and development capacity to support the projected buildout, but geopolitical, economic and technical constraints could lead to a very different outcome.
AI Infra: Are we confusing a U.S. power shortage with a global supply glut?
One of the biggest debates around the AI buildout is whether we have enough power and data center capacity to absorb all the compute coming online in 2027.
I think we're missing an important distinction: U.S. supply constraints ≠ global supply constraints.
Global demand: ~50-55GW in 2027 based on most estimates.
[chart below]
Global supply: ~50GW. Dylan Patel / @SemiAnalysis_ forecasts roughly 50GW of new capacity coming online, including 10-20GW attributable to @OpenAI and @Anthropic.
[chart below]
U.S. supply: ~20-25GW. @FundaAI estimates ~17-23GW. If the U.S. accounts for ~70% of global demand (~35GW), we're looking at a potential 10-15GW power shortfall. That gap is real, and I don't dismiss it.
[chart below]
But here's what I think the market is overlooking: the rest of the world.
Based on channel checks, Southeast Asia could potentially bring online ~15GW, Japan ~5GW, Europe another ~5GW, with additional capacity across China, Canada and increasingly Latin America.
These are indicative estimates, not guaranteed deliveries. But the point is that global capacity to absorb AI compute may be considerably larger than a U.S.-centric analysis suggests. Of course, power availability doesn't automatically translate into operational data centers. Permitting, grid connections, equipment deliveries and construction timelines all matter. And compute demand isn't perfectly fungible across geographies.
Still, a U.S. power shortage doesn't necessarily mean the world cannot absorb 50GW of compute.
Another important data point: $TSMC.
Of all the companies in the semiconductor supply chain, $TSMC is probably the one whose judgment I trust most. It is famously conservative and has exceptional visibility into customer demand.
On its latest earnings call, CEO C.C. Wei said TSMC has gone as far as sending its own teams to verify data center construction progress and power availability in the U.S.
"The feedback is so far so good, and the CSPs have been well prepared."
My take: I'm cautiously optimistic that sufficient capacity exists globally to absorb the incoming compute supply. The RoW opportunity is larger than many appreciate.
My biggest concern remains execution: whether enough land, power and data center shells actually get delivered on time. Delays could leave chips temporarily idle and cause near-term disappointments for certain suppliers.
But that's a very different bearish thesis from structural global oversupply.
++
Better formatted: robonomics.substack.com/p/ai…
HAP retweeted
From the @FT: "Cracks are appearing in America’s debt-laden $10tn AI infrastructure boom as project delays, public backlash and shaky investor confidence bring a reckoning for the data centre build-out underpinning the US economy."
IMO, what makes all these issues even more interesting is that no matter how they evolve, there are significant economic, financial, social, and geopolitical implications
#economy #markets #tech #ai
HAP retweeted
Why does a company with $104 billion in contracts still lose money?
CoreWeave sells GPU computing for AI. In Q2 2026 it brought in $2.58 billion in revenue and reported about $104 billion in backlog. It still posted a $626 million net loss.
Its net interest expense that quarter was $640 million.
Before interest, the business was close to breakeven, with an operating loss of $49 million. Interest turned a small loss into a large one.
Why so much debt? GPUs are paid for before a cluster goes live, while customers pay over several years. On the income statement, that equipment cost is spread out as depreciation. Much of the cash spending comes before the customer payments. Loans help bridge that gap. CoreWeave borrows against signed customer contracts and the GPUs themselves, so one contract brings in revenue and backs the loan.
A neocloud is a cloud company specialized in GPU computing for AI. Nebius, Lambda, Crusoe and Nscale are others, each funded differently.
Think of a landlord who signs the tenant, borrows to build, then repays the bank with the rent.
That is why financing conditions matter. This week, some CoreWeave bonds traded at yields around 15%, signaling pressure on future borrowing costs. Existing fixed-rate interest payments do not automatically rise.
A large backlog helps secure financing. It does not guarantee a profitable buildout. The real test is whether customer payments recover the equipment investment and financing costs, while arriving in time to meet debt payments. Not investment advice.
HAP retweeted
According to MS, AA-rated hyperscalers are "mostly price insensitive to higher interest rates," since they start from interest coverage above 30x. The problem is everyone else. In an illustrative exercise, MS has the high-quality hyperscalers issuing $400 billion at 7% next year, which takes coverage from 30.7x to a still-comfortable 11.9x. But give Oracle and SpaceX $120 billion at 8.5%, and their combined coverage collapses from 4.0x to just 1.8x.
Nvidia Gets $690BN, The Other 499 Get $340BN: Tracing Every Dollar Of 2027's AI Capex zerohedge.com/markets/nvidia…
HAP retweeted
AI capex这个数被翻来覆去地算，税这块我几乎没见人提。
美国去年把100% bonus depreciation写成永久的了。服务器、GPU、交换机都算5年期资产，买的当年就全额抵掉。几家hyperscaler一年几千亿美元的采购，按21%的联邦税率，当年就是几百亿的税不用交——不是免掉，是往后推。你每花一块钱买GB200，山姆大叔先替你垫两毛。
麻烦的是这玩意儿会上瘾。递延税只有在capex一直涨的时候才永远不用还。哪年真把采购砍下来，抵扣没了，收入还是平的，税单自己先跳起来，自由现金流反而更难看。
假设你是MSFT的CFO，明年很想在财报会上说一句"capex纪律"。先把那张税单算一遍：少买的那部分，那两毛谁来垫？
HAP retweeted
卧槽，原来这周砸崩 AI 股的 CoreWeave，玩法跟借钱拿地、盖楼、卖期房的开发商一模一样。
这事的导火索在本周二。Chamath 在 X 上指出 CoreWeave 一笔债的收益率到了 15%，马斯克在底下回了一个词“yikes”（哎哟，吓人），直接上了一堆新闻。第二天 CoreWeave 和同类公司全线大跌，CoreWeave、IREN、Nebius 这类“新云”公司这周都跌了约 20%。周五 CoreWeave 又跌 11.9%，一周跌了约 25%。
硅谷最火的科技投资播客 All-In 最新一期，Chamath、Sacks、Jason 加上代班的 a16z 联合创始人 Ben Horowitz，把它扒了个底朝天。套路很眼熟：签下微软大单、拿合同去借钱、买英伟达芯片收租。已经借了约 350 亿美元，还得再借约 350 亿，75% 的收入绑在三家客户身上。
更憋屈的是它把“期房卖早了”。长约都是固定价格，现在现货算力价格涨上去了，它多赚不到。Chamath 甚至建议它干脆赔钱毁掉部分合同、改卖现货，说这中间差了两三百亿美元。
现在债主先慌了。Chamath 算了笔账：它的违约保险这周涨到约 953 个基点，5 年违约概率约 55%，“基本是抛硬币”，“这不是我说的，是保险市场说的”。一笔 GPU 贷款的利率已经涨到 15%。他也承认表面上不危险，资产约 780 到 800 亿美元，债约 350 亿；但他团队估算，离触发债务条款只差正负 50 亿美元左右，这是一场“信心游戏”。
Sacks 觉得这只是波动：它实际付的利息还是 9.6%，二级市场打折影响的是以后借钱的成本。最该盯的是回本周期，CoreWeave 要 2.5 年，马斯克的 SpaceX 卖现货算力约 1 年就回本。
Ben 则说 a16z 一家这类公司都不投。1999 年带宽每兆 1500 美元，只能签三年长约，到 2001 年跌到 50 美元，借钱囤带宽的“直接死了”。不过他也承认现在不一样，算力严重短缺，亚马逊 CEO 说“2028 年以后都卖光了”。
最后 Chamath 点破了真正的死穴：10 年期美债要是涨到 5.5% 到 6%，“所有靠借钱赚利差的生意都会被碾死”。接下来一个月见分晓，就看它还能不能借到那 350 亿。
HAP retweeted
Gartner says that by 2030, 80% of neoclouds will have ceased operations, that AI data center owners would need to generate $8.5 trillion in revenue to cover the current spend and hit an ROIC of 12%, and that $4tr+ in assets face writedown risk by 2032.
gartner.com/en/newsroom/pres…
HAP retweeted
Bloomberg: "Virtually every major category of costs associated with building a data center is ballooning...The cost of construction materials such as steel and cement has skyrocketed because of trade wars, fuel expenses and geopolitical conflicts. The AI boom itself has driven up expenses for everything from chips to copper. The price of short-term memory used in data centers is more than five times what it was just a year ago, data compiled by Bloomberg Intelligence show. Demand for equipment that generates and carries electricity is so high because of the building frenzy that suppliers have backlogs stretching beyond 2030."
It is important to keep in mind that only $492 billion of the $2.3 trillion data center development pipeline is currently under construction. The remaining $1.8
trillion is tied to planned and precommitted projects. And those projects are facing rapidly inflating costs. This inflation is something I dove deep on in my report on "The AI Trade" (sageroadresearch.com/product…). To quote the report:
"The question remains, can hyperscalers enable agents at a price enterprises will pay that’s margin positive for hyperscalers? That’s clearly not the case today. As Nvidia VP of applied deep learning Bryan Catanzaro acknowledged in June: “The cost of compute is far beyond the cost of employees.” And the cost of compute is inflating.
As the FT noted in late July, over the past two decades, spending on semiconductors averaged 0.47% of world GDP per year. This year, total spend on semiconductors is projected to equal 1.25% of the world economy. And next year, that spend may exceed 1.5% of global GDP—roughly three times higher than three years ago. From demand increasing far faster than supply has come “chipflation”. “On-demand GPU capacity is effectively sold out,” as Apollo noted in mid-June. TSMC, which owns roughly 72% of the global chip foundry market, is reportedly sold out until 2028.
Yet, it’s not just “chipflation” that matters. IT equipment, which includes chips, accounts for 64% of the total cost of building a data center. Infrastructure accounts for another 29%, which includes electrical systems, cooling, cabling, and fire prevention. Those costs are also inflating, up an estimated 40% over the past two years. Then there’s the cost to actually build the physical structures, which accounts for 7% of data center building costs. Inflating data center construction costs have been the primary driver of the PPI for non-residential construction spiking since 2021. Overall, the average per megawatt cost of building a data center has increased from roughly $7 million pre-pandemic to $11.3 million this year. For AI-optimized data centers, the average cost per megawatt is expected to land between $15 million and $20 million this year. According to iRecruit.co, 50% of data center executives report project bids rising by 6% to 15% over the past year.
We believe the unit economics of agentic AI are one of the challenges most neglected by AI optimists today. As has been made clear this year, enterprises will not spend heavily on agentic capabilities unless the models improve at the same time costs come down. Maybe in an ideal world, hyperscalers would slow down the buildout as they wait for model efficiency breakthroughs and chip advancements to drive down token costs and agentic token load. However, competition won’t allow that. So, they’re racing to build out the data center capacity needed to improve the models, but that effort is only increasing the cost of building data centers, which is increasing the agentic price tag, at least in the near to medium term."
Bloomberg link: bloomberg.com/features/2026-…
HAP retweeted
推荐大佬写的Claude Code 防封工具箱，你想成为哪国人使用CC 就可以变成哪国人。
1. 系统环境设置 —— 一键切换时区、系统语言、区域格式（选个地区就套用），并能打开定位等系统权限开关
2. 内置的"指纹 Chromium"浏览器可切换
3. VPS 网络搭建，俗称翻墙（可选）
4. 精确清理本地claude数据等
HAP retweeted
WSJ just reported that Amazon is in talks to buy Decart for about $7B, after earlier talks with Nvidia and Anthropic collapsed.
Decart, founded in 2023 by Dean and Orian Leitersdorf and Moshe Shalev, builds software that lets developers move AI workloads between different chips.
Its DOS optimization stack claims to make models run up to 8x faster on the same hardware, including Nvidia GPUs, Google TPUs and Amazon Trainium. So basically with this acquisition, Amazon must be targetting to lower Trainium’s cost per token below Nvidia hardware.
Amazon has a head start on the technology, since AWS gave Decart early access to Trainium3 in 2025, and Decart adapted its model architecture to fit Trainium’s chip design.
Decart also builds real-time world models: Lucy for immersive experiences such as live video, and Oasis for physical AI and robotics simulation.
The price would sit about 75% above the valuation Decart set in May, when it raised $300M at an estimated $4B valuation.
Anthropic, Amazon’s close AI partner, wanted Decart’s technology to make its existing computing infrastructure handle more demand, but abandoned a $6B deal after due diligence.
The deal pays off for AWS only if Decart’s speedups hold on customers’ models, beyond Decart’s own video systems, and
