Tigris | 会讲课教授是好老师 · X 阅读

Intel $INTC didn’t rally on vibes today. The market is slowly waking up to a simple but brutal shift: Apple, Google and Meta might push part of their advanced packaging to Intel. If that’s true, you’re not looking at a chip stock anymore. You’re looking at the next control point in the AI physical stack. But here’s the thing most people can’t hold a position even when the thesis is screaming at them. They don’t internalize anything. Three green candles create conviction, one red candle destroys it. This entire setup only makes sense if you understand the structural logic. So I put it all in the new deck. Start with page one.

Intel $INTC is becoming the only real geopolitical hedge in the entire AI stack. The decade-long mispricing is finally starting to break. I’ve pulled together a fund-grade memo, supply-chain maps, scenario paths, and the exact mechanics we use internally to price the re-rating. This isn’t a cheap-stock story. It’s the market waking up to the only thing that matters once AI turns into national infrastructure: on-shore advanced packaging and a credible U.S. foundry anchor. PPT: Gumroad link below https://t.co/egvh2zao2N Inside the Deck: Why hyperscalers are positioning Intel as the second anchor of the global compute stack; What the true inflection point looks like (and why Wall Street keeps looking at the wrong one); The structural mechanics behind the re-rating; When 18A can realistically cross economic breakeven; How far the U.S. reshoring agenda can push the upside; And where the actual risk boundary sits; For research use only.

我来解读一下，现在美股大家都说贵，理由很简单，标普500 的市盈率 22 倍，明显高于历史均值。 但问题是，大部分人连指数是怎么算的都没搞清楚。标普不是“500 家公司的平均”。 它是按市值权重算的。权重越大，对指数的估值拉动越强。 结果就是… 现在前 7 家巨头也就是 Mag7权重市值占市场38%，直接把整个标普的估值抬到 22 倍，看起来“全市场都很贵”。 可如果你用等权重来算，每家公司只占 1/500，估值立刻掉到 16 倍。 这才是真实的市场：大部分公司并不贵，是那几家超级巨头贵。 为什么会这么集中？前十家公司占了整个标普 42% 的市值。历史新高。 原因也不复杂。投资者知道在这轮科技革命里，能活下来、能吃掉全球算力红利的，就是最强的那几家。 市场不是盲目涌进去，而是看得非常清楚：强者恒强，当然随着ai传导，小企业一样也会有春天。

Investors keep asking for a clean long-term valuation range on $NVDA, so here is the framework that actually matters. If AI becomes an internet-scale platform, it will unlock roughly 20 trillion dollars of new value over the next decade. Semiconductors historically capture 15 to 20 percent of the compute layer, and Nvidia controls the critical layers of that stack: silicon, systems, software, and ecosystem lock-in. Even under conservative share assumptions, that translates into 2 to 3 trillion dollars of incremental enterprise value.This is what underpins the long-term optimistic band near 300 dollars per share. On earnings mechanics alone, the math is straightforward. Next year’s 6 dollars EPS at 40x gives around 240. A more optimistic 50x pushes toward 300. That’s the baseline for an internet-level innovation cycle. But the real story is non-linear. If AI compounds into robotics, autonomous systems, bio-automation, and agentic computing, the total value pool expands to 20–40 trillion. At that point, Nvidia’s advantage is not a higher multiple which is a structural position as the bottleneck supplier of compute. Paradigm shifts do not reward the average firm; they reward the few that become indispensable. This is where the reflexivity kicks in. Human FOMO pulls valuations forward. Markets price future dominance before the cash flow arrives. If investors begin anchoring to 2027–2028 EPS, the extreme upside becomes a function of psychology as much as fundamentals. Of course, every supercycle carries a risk. Two years of over-extrapolation can produce a bubble that wipes out the incremental premium. When the index moves 1 percent and Nvidia moves 3 percent, leverage and algorithms magnify fear and euphoria in both directions. Even under the most conservative industry assumptions, 4.6 dollars EPS in 2026 and 6 dollars in 2027, applying long-term semiconductor multiples of 30–40x still yields a fundamental floor around 180. Everything above that is the blend of structural dominance, macro liquidity, and investor positioning. Markets quote every day. You don’t need to trade every day. This range is a guide to the long-duration technology curve, not a short-term call. Personal research only. I hold $NVDA and may be biased.

The memory sector has been the strongest performer this year. Prices are rising almost monthly and Micron is already up 2x. So when $MU sells off right after Google drops a new AI model, people assume the worst: “CFO raised Capex. Is this the start of another overbuild?” That reaction completely misreads where we are in the cycle. Micron is not dealing with a traditional DRAM boom-bust. It’s sitting inside a multi-year, AI-driven supercycle with demand coming from every direction. Model sizes are still compounding. Inference is just getting started. AI agents and on-device intelligence are about to enter consumer platforms. Cloud providers, AI chip companies, and even government HPC programs have already locked 3–5 year budgets. This multi-source, cross-industry demand pattern has never existed in memory history. The old 2017–2018 playbook doesn’t apply. HBM is not a business you can scale by simply “deciding to expand.” The stacking process is brutally hard and capacity takes years to qualify. The industry isn’t expanding recklessly. It’s expanding reluctantly. Micron’s Capex hike is not aggressive expansion. It’s forced expansion. Customers have already pre-booked most of 2026 supply. HBM is structurally undersupplied and will remain so. If Micron doesn’t invest now, the real risk isn’t memory oversupply. It’s a bottleneck across the entire AI supply chain: GPUs underutilized, compute nodes delayed, inference costs stuck too high, and model upgrades slowed for everyone. Capex going up is not a warning sign. It’s confirmation of overwhelming demand. We’re nowhere near the point where oversupply is a realistic concern. The winners of this cycle will be the companies willing to scale within a structurally constrained industry. This is the real signal from Micron’s latest commentary.

Everyone keeps trying to size AI by asking how much it can add to global GDP. That question is already too small. Global GDP is roughly 100 trillion dollars. If AI were just a writing assistant, then yes, the ten year productivity boost would be a few trillion at most. That is the linear scenario. But AI is not stopping at language. Once it penetrates enterprise workflows, rewrites decision chains, and automates operational layers, the curve shifts. You move from a simple efficiency tool to a new coordination technology. At that point, the contribution is not three trillion or eight trillion. It becomes a decades long compounding engine. Twenty to thirty trillion is the conservative range. The upside becomes far larger once AGI, robotics, autonomous systems, and industrial automation evolve. Here is the part most people miss. Macro totals do not drive investment returns. Concentrated profit pools do. Technological revolutions never reward a thousand companies. They reward the few that become the infrastructure. The internet did it. Mobile did it. AI will be even more extreme. Nvidia, Google, and other system level players are not showing incremental advantages. They are showing lock in driven by a full paradigm shift. Entire value chains reassemble around them. This is why AI’s economic impact will almost always be underestimated. Averages hide power laws. The big money is never in the mean growth rate. It is in identifying who becomes the profit gravity well. If you pick the future center of value capture, you are not trading an AI cycle. You are positioning at the base of a multi decade uplift in the trajectory of the global economy.

If there was a real bubble, it happened in 2023 and 2024, when a single early model like GPT-3 or GPT-4 was enough to send Nvidia up tenfold and push valuations into triple-digit territory. That period was speculation driven by imagination rather than earnings. What we are seeing in 2025 is the opposite. This year has been about validation, not excess. The AI buildout is now funded by four real buyers with real cash flow, like Microsoft, Google, Amazon, and Meta and their collective valuation sits near 26 times earnings, far below the seventy times we saw in the dot-com era. The market is also no longer treating AI as a single trade. Dispersion is everywhere. Google, which finally aligned hardware, software, and infrastructure, delivered a genuine rerating after years of underperformance and posted roughly fifty percent annual returns. In contrast, Amazon’s high multiple and slowing growth, and Meta’s strong growth but weak AI cloud monetization, left both stocks flat for most of the year. This is not what bubbles look like. This is what a rational repricing looks like when earnings momentum and monetization pathways decide who wins. Across all the downturns I’ve lived through, I have never seen a correction as orderly and disciplined as 2025. AI Capex is still expanding, and the business models are only beginning to show real commercial traction. You don’t disprove a technology cycle while the infrastructure is still being built. If anything, this phase is where high-quality names offer the best risk-reward, and these prices start to look more like a Black Friday sale than the end of a trend.

If you want to size $GOOG, you need a range, not fantasies. Investors get lost because they listen to every “10x” pitch and forget the math. Here’s how I look at it. We’re entering a 10 to 20 trillion AI value cycle. Alphabet can reasonably take 1 to 2 trillion of that. Moving from 3T to 4–5T is not storytelling. It’s value shifting toward whoever controls the workflow. The upside is never about the multiple. It’s about earnings velocity. If Google embeds AGI into Search, Android, and enterprise operations, the business model changes. You stop selling ads. You start charging an access fee to the digital labor economy. That means higher intent capture, higher ad yield, and a larger enterprise mix. Run the tape. If commercialization works, 2028 can land near 15 EPS. Put a 30 multiple on that, and you get a 450 stock and a 5.5T company. Of course there are risks.Regulatory breakups. AI cannibalizing core search. One headline can take the stock down 15 percent. But with Buffett anchoring the floor around 250, anything short of an extreme shock is just volatility, not a thesis break. In this market, you don’t follow stories. You follow earnings acceleration. And right now, Alphabet still has a real shot at owning the next leg of productivity growth.

I lived through every crash since 1987. They all looked different on the surface, but the pattern never changed. They arrived faster than anyone expected and wiped out what looked safe. Real investing was never about guessing moves. It was about building a system that protects you from yourself. Know the business, pay a fair price, and always keep a margin of safety. 2025 reminded everyone that high multiples without rising profits do not survive. Amazon proved even giants get punished when earnings lose momentum. This is not bearish. It is markets becoming rational again. Hype fades. Fundamentals stay. And investors who respect discipline are the ones who walk out stronger.

存储板块是今年表现最好，供不应求天天涨价，美光 $MU 已翻倍。 现在在谷歌推出AI模型还能继续跌，TPU也要用宽带HBM内存啊。 这是担心最新市场看到美光 CFO 在投资会议上说Capex要上调，第一反应就是担心扩产会不会带来未来的产能过剩。 我觉得担忧都来得太早了，本质上完全误读了现在所处的结构性阶段。 美光面对的不是那种传统 DRAM 的短周期，而是一条由 AI 需求推动、确定性极强的超级长周期。 大模型规模还在指数扩张，推理端刚刚起飞，AI 代理和本地智能开始进入消费互联网，云厂、AI 芯片公司、政府超算全部按三到五年锁定预算；这种多源头、跨产业的需求结构，在存储行业历史上从来没有出现过。 这意味着你不能再用“2017–2018 扩产过度、两年后杀价竞争”那套旧心智来理解这波周期。 HBM 堆叠本身的技术难度决定了它根本不是“想扩就能扩”的业务。这不是一场无序扩张，而是一场克制扩产。 而在这种克制的背景下，美光的 CapEx 上调不是公司冲动扩产，而是典型的“被动扩产”，因为客户已经把 2026 年的大部分供应锁死，HBM 供不应求的局面短期内根本无解。 此刻如果美光不继续增加投入，真正会出现问题的不是美光的 Capex，而是整个 AI 供应链，GPU 无法完全释放性能，算力节点延迟，最终影响到云厂的推理成本和大模型升级节奏。 美光现在上调 Capex，不是“扩产的风险”，而是“需求强到不扩不行”的信号。 担心过剩完全是想太早，现在仍然是全产业严重供不应求的阶段。 能够在克制行业里敢扩产的公司，才是下一轮周期的赢家。 我这里简单解读美光这次会议真正传递的核心信息。

不论你管的是国家基金、保险基金、养老金、主动管理、还是指数增强，谷歌都应该是结构性必配资产。 这个时代只要你想跑赢市场，甚至只是想拿到应得的回报，有个常识必须先接受 AI 已经写进未来十年的资产定价了。你不配谷歌，连最基本的认知线都是踩空的。 原因很简单，三个点就够了 第一，谷歌是唯一真正做到从硬件到模型到流量入口全链条打通的公司。TPU、数据中心、模型、搜索、Android、Chrome、Workspace，这已经不是生态，是操作系统级的护城河。 第二，现金流就是大模型时代的门票。没有百亿规模的自由现金流，根本没资格玩全栈。谷歌一年仅2000亿搜索利润能cover。 第三，最大的系统性风险只有司法部反垄断。而历史证明，反垄断打掉的是估值，不是价值，很多时候甚至能释放股东利益。风险本身都带着向上的预期差。 巴菲特也配置了，现在很多机构CEOCIO都在紧急学习反思，你不配能跑赢市场？不配谷歌，不是保守，那是真不懂世界。

大家都在说“铲子”一大把“金子”在哪里？也就是真正大笔AI的Capex投入哪里有回报。 我开解释为何这是大把金子的历史性地红利，56十年代的退休老年广场舞红利、7 8十年代有改革和炒房红利、21世纪这些年轻人不是洗洗睡了，而是星辰大海。 著名物理学家费米当年靠几张纸就能估出核爆当量，因为关键不是精准数字，而是抓准数量级。 今天预测 AI 对经济的影响，也必须用这种思路。传统模型是旧时代的工具，用来算 AI 会直接失真。 所以高盛错误在于那种“AI 贡献 8 万亿”的预测，本质是把 AI 当成互联网。 但 Gemini 3 已经证明这不是流量革命，而是知识劳动的工业革命和电气化级别。我从这个角度粗估未来，相信比所谓“精致计算”要真实得多。 粗估全球 GDP 如今是 100 万亿美金。如果 AI 只是写字工具，十年贡献生产率提升带来总的贡献累计确实可能只有几万亿。 但当 AI 进入企业流程、自动化任务、重写决策链条时，增速直接跳档。真正的数字不是 3 万亿，也不是 8 万亿，而是一个会持续几十年的结构性叠加：二三十万亿是低区间，长期可能更高。 这还不包括可能实现的AGI、机器人、自动驾驶和工业自动化等更大的上限。 从投资看更重要的是，宏观增量从来不是投资逻辑。技术革命的利润不会分给一千家公司，只会集中到少数几家。 互联网如此，移动互联网如此，AI 更极端。Nvidia、Google等核心头部科技的领先不是线性优势，而是范式切换后的“锁死式扩张”带动产业链的发展。 AI 的经济贡献会被严重低估，真正的机会不在平均增速，而在赢家通吃。 选对未来的利润分配中心，你不是抓住一轮行情，而是站在整个文明曲线抬升的起点。 这是历史性地红利，56十年代的退休老年广场舞红利、7 8十年代有改革和炒房红利、21世纪这些年轻人不是洗洗睡了，而是星辰大海。

@Real_China_O 我自己梳理一下主流所谓投行的分析，虽然错的离谱但是媒体广为流传。 这种内容肯定阅读难度和认知强度，远大于我随便写个什么赠手机和中国战狼的段子，虽然后者流量巨大。 但是相信真正读到读懂的人，回馈的是认知和财富。虽然我也不是肯定正确，但是相信能激发思辩

如果你刚买了辆特斯拉，你会突然发现满街都是特斯拉。市场也是这样：价格先动，认知才跟上，叙事最后被编出来解释价格。 近期市场下跌，不是什么AI泡沫，而是一次典型的流动性真空、信息真空、叙事反身性的叠加的小型危机。 全面解读当前市场背后的空头机制以及应该用何心态应对以及组合配置策略 美联储十月降息我第一时间解释鲍威尔会“无意”人为制造12月不降息资金危机 市场误读成“宽松回来”。但债市早已提示 11–12 月会进入“过山车式抽水期”。再叠加政府关门、数据停摆，市场突然失去信息。 当没数据时，价格本身就是数据。 于是反身性链条启动：价格跌，情绪波动、故事被创造来解释下跌 看看近期各种高盛之类的小作文（AI 泡沫？科技顶了？宏观危机？） 叙事再反过来加速抛售。机构获利回吐最爱的年底制造一次 mini-crisis 获利回吐，洗掉 FOMO、清杠杆、打压散户。 这轮下跌不是趋势反转，也不是泡沫破裂，只是现代市场的底层逻辑： 预期修正 to 流动性抽 to 信息真空 to 价格带情绪 to 情绪造叙事，叙事推价格。 市场不是因为“发生了坏事”而下跌，而是因为“没人知道到底发生了什么”而下跌。 但股票和数字货币或者黄金最本质差别在于，股票蕴含人类的价值创造力，现金流会把价格拖回现实，重新带回对未来的预期。 我之前提心过短期年底从组合角度提升现金比重和医药权重，但是未来一定在于AI，人工智能对于整个经济范式的改变，将永久提升增长上限 现在只是刚刚开始，没有任何一个泡沫在刚刚开始投入就破裂的。 资金短期紧张的反转随时可能发生，可能明天可能英伟达财报之后，或许Gemini 3震撼发布的任何事件性冲击。不用预测也无需预测。

熟悉长期看我推特的朋友都知道，从川普关税一出来，我的立场就很清楚，关税压制需求不会导致通胀。 这与“主流经济学家”和美联储的关税不确定拉高通胀唱反调。因为我看清他们不是不懂而是站在“政治”角度为了反对关税而反对。 最新的旧金山联储研究150 年美国关税史，严谨实证，结论非常简单… 关税对总需求的打击大于对成本的提升。它的净效应偏向通缩，而不是通胀。 所以现在最大的荒谬是什么？ 鲍威尔还在把通胀压力往关税身上推。在数据面前还继续坚持这个叙事，不是认知问题，就是态度问题。 眼下的正确方向其实更清晰， 不是继续维持高利率，而是加快落地降息，给美国的投资、创新和居民消费一个稳定的利率环境。 你不能一边说美国要重建产业链、推动能源投资、推进 AI 和基础设施升级；一边又让利率卡在几十年最高位。 关税不是问题，错误解读关税才是问题。 美联储如果继续坚持旧故事，那对美国经济的伤害才是最大的通胀源。鲍威尔这种不下岗等这过年？

真正的好公司，总是在低谷时被质疑，AI时代唯一同时握有搜索、模型、芯片、应用、数据五大底牌的公司。 我早在今年2月就系统分析过谷歌的投资价值，也一直坚定持有。 所有的后见之明都不值钱，真正难的是在上涨前，能穿越波动和质疑，守住自己的判断。 投资从来不是预测涨跌，而是先有深刻认知，再让财报去印证。 作为自身的用户，夜从AI时代红利一路关注见证它的模型发展、云计算市占率、分拆风险、搜索引擎广告的现金流。 $GOOG

Everyone loves “buying the dip”直到真正的 dip 出现。 因为历史里的 dip 都很便宜，现实里的 dip 都很恐怖。 真正的理由只有一个： 下跌永远裹着恐慌、怀疑、自我怀疑。等你看懂，它已经反弹了。 而且那些嘲讽“标普才跌 3%，算哪门子 dip”的人…你陈述了事实，但错过了重点。 在成长板块，三十多个点的回撤已经摆在那。这才是真正的 drawdown。而且还可能更深。 但机会永远属于一种人： 早就想好自己要买什么的人。 别人慌，你有清单。别人问是不是底，你关心是不是错杀和低谷 市场没变，是人的心态先崩。 上涨不断放大情绪，下跌也是如此。 真正的买点从来不是“看上去安全”，而是你能看穿情绪表面，看到真正的成长股被低估错杀，敢在别人不敢的时候动手。

别只用AI，学会造AI。 课程网站：https://t.co/VI3K85anqN 材料会陆续开放：作业、自评系统、讲义、视频全套。 想真正理解AI工作原理，而不是只会prompt，这门课值得收藏。

既然大家认识AI都是从ChatGPT开始，那何不先学清楚这东西到底怎么工作的？ 终于有人把“现代AI”真正教对了。不是一堆论文数学符号逻辑，而是手把手教你从PyTorch开始造一个Chatbot。 CMU的春季新课《Intro to Modern AI》授课是CMU的机器学习主任教授Zico Kolter NLP领域顶级专家。 码住，学起来

$GE $CAT 今年都涨了快80%了，如果AI泡沫真要破，绝不会是Burry嘴里GPU折旧，而是因能源电力。 美国在建数据中心需求45GW，相当于40座核电站，总投资2.5万亿美元。钱不是问题，电才是问题。 电网排队接入要7到10年，巨头们只能自建电厂。目前方案？燃气轮机、内燃机、柴油机组。 泡沫破在哪里？ 现在主要燃气轮机设备涨价50%、交货两年起步，但是能用钱解决的都不是事，很多解决不了在于，没人干活。 懂高压电施工的工人，比GPU还稀缺。 这才是AI资产现金流的最大黑洞。没供应，多大需求也没有现金流，而且也影响风险折现率，自然影响股票估值。 或许这也是黄仁勋说的，中国AI竞争能赢的可能性所在。不在于算力、在于能用的算力。未来要抄电表。