Tigris | 会讲课教授是好老师 · X 阅读

我不认同市场上那套特朗普“Art of the Deal”的老剧本解读。把它当作表演，就会低估它对定价机制的伤害。 如何看三重叠加风险？证据拆解？短中长的预测判断？具体应对策略？ 👇 上周末我把这周的盘面当成一场“三前线冲击”做了系统预测解读 格陵兰美欧关税冲突风险，美联储案件独立性风险，以及 OpenAI 自身在商业模式与现金流压力下的“找出口”。 三件事叠在一起，关税一旦和地缘政治绑定，风险溢价就不再按教科书走，你看到的就是波动率先行，指数高位下跌。 应对上我只强调一句：长期看好不等于短期可以乱冲或装死。仓位、止损线、最大可承受亏损，这些不提前写清楚，就不是交易，是赌博。小一点，规矩紧一点，别带 ego，先活过这段震荡行情。 目前市场集中避险，两条更接近现金流的线：存储是确定性更高的那条，价格链条在走，兑现路径更短。CPU 是“增量信息”的那条，AI 推理把通用计算重新拉回舞台，服务器升级带来的 CPU 需求重定价，是逆势的真正原因。 另外市场这周还绕不开一个地缘政治含金量很高的节点：Intel 财报（1/22 盘后）。我也已经把它放进产业链框架里做了点评，看看它到底是“故事修复”，还是“订单与供给的真实回归”。 当然不为浮云遮望眼，短线像过山车，但美国资产最硬的底层逻辑仍然是生产率。生产率回到长期趋势附近并继续抬升，才是长期回报的发动机。AI 这件事最终要在这里交作业，否则都只是叙事。 更完整的框架与证据链在我的英文相关推文以及article文章里。 想看更及时、更深度、带仓位纪律的更新，就走订阅链接。

Seatbelt on. Everyone will argue about Greenland tariffs and Fed independence this week. Important, yes. But the picture you’re looking at is the bigger tell. OpenAI CFO’s blog just did something rare: it put a hard spotlight on the real engine of the AI boom. Electricity. Compute. The meter. Strip away the hype and you see the uncomfortable shape of the business: an “AI company” that behaves like a power-hungry utility with a software logo on top. Training is a furnace you can’t turn off. Inference can get cheaper per token while the total bill still goes up, because usage explodes. Cheaper unit price, bigger volume, higher total cost. That’s how margin dreams die quietly. And that’s why the monetization debate is back with teeth. Ads, take-rate, enterprise agents paid on outcomes, maybe. But none of that is proven at scale yet. Until it is, the market has to price a simple risk: a compute burn rate chasing a revenue model that’s still being invented. Now stack the macro on top. Greenland tariffs and the US–EU trade fight are a risk premium machine. Policy uncertainty taxes multiples. Then the Fed “The Lisa Cook case “ adds another layer of noise: legal drama and independence anxiety make the referee look pressured, and markets hate that feeling. Put it together and you get the same outcome: higher volatility, thinner patience, and less forgiveness for story stocks. One takeaway: this week can punish size. If you don’t have a clear “I’m out” level and a max loss you can live with, you’re not trading, you’re gambling. Smaller bets, tighter rules, and no ego. Survive the chop, then swing when the dust settles. I’ll keep the public post clean. The positioning playbook and Intel $INTC checklist go to subscribers via the link in my profile.

$MU 股价涨了，大行不断调升目标，理由不外乎需求爆炸供给有限。这是卖方给散户看的。 我们要看什么？人和执行力，对于半导体高精尖制造业永远是“良率 + 封装 + 交付”这条最难的路。美国本土扩张才是被低估的战略溢价。 能赢，才能给PE估值，10，15，20的估值扩展，不然就是PB，3倍都嫌多。👇看懂

Intel 第一个装上 ASML 最先进 High-NA EUV，单台3.8 亿美元。 如果说AI壮丽如诗篇，芯片就是写它的纸。而EUV光刻机，是人类目前唯一掌握的“造纸术”。 2026，愿你成为光，也做那个不向“做不到”低头的人。 两份 Deck：中文把 EUV 讲清楚，英文用 ASML 季报看懂市场怎么给它定价和趋势。 Link见评论

TL;DR: A macro-architecture upgrade made trainable, proving the bottleneck is still bandwidth and memory.” DeepSeek just dropped a systems paper, not an “algo tweak.” Think of training as building a long highway. HC turns each layer into a multi-lane interchange: more routing freedom, more power, but it can destabilize deep training. mHC adds strict “traffic rules” so the mixing stays conservative and trainable at scale, while keeping the marginal overhead around ~6.7%. The real signal: the race in training isn’t over, and the bottleneck is still bandwidth + memory (the memory wall).

今年过节不收礼，收礼只收DeepSeek的“大红包” 人家过节发钱，DS梁文峰发论文。这不是小修小补修个函数算法，这次是又一次震撼性的系统级的训练方面大升级 我尽可能浅显易懂讲清楚拆解出核心的对于26年AI训练和半导体意味什么？搞研究搞投资必读啊 以前训练大模型，有点像修一条很长的高架。 每一层高架只有一个接口，一辆车一路往前开，稳定是稳定，但你能做的路况调度很有限。 论文呢想法很激进： 把每一层高架改成多层立交，直接开四个车道，允许车在不同车道之间换道分流。信息流动更自由，理论上模型更聪明。 但自由是要付代价的。 立交越复杂，越容易“失控”：有的层把车流放大成拥堵，有的层把车流抽干成断流，层层叠加，训练就容易崩。 这篇论文最关键的价值，是给这套“多车道立交”加了一套硬规则：每次分流只能按比例分配，不能凭空把车变多，也不能把车搞没。结果就是两件事同时成立：训练更稳，效果也更好。 更重要的是，它没有停在“理论很美”那一步。这么搞整体训练成本只多了大约 6.7%，但换来的，是可训练性和性能上的提升。 这篇论文背后的信号是什么： 首先模型训练的竞赛远没有停滞。推理更重要但训练还在继续抬高上限可能性。 算法的时代结束了吗？没有。算法敢不敢设计得更激进，工程能不能把它撑住。谁能把“想法”变成“跑得起来的训练吞吐”，谁就拿到下一轮领先。 再次验证了一个老结论：今天最大的瓶颈依旧不是算力本身，而是带宽和内存。 对应到现实，就是高带宽内存、互联、通信、内核优化这些硬东西。 所以，这篇论文并不会让 HBM 需求消失。相反，它在告诉你：架构越进化，带宽和内存越成为核心约束，软件和硬件会继续互相逼迫、互相推动，形成滚动式升级。 把它当成 2026 的第一份开年礼就行。 它预示的不是“又一个模型更聪明了”，而是更大的趋势： AI 的天花板还在上移，而且这次上移的方式是系统工程把算法托上去。 新年第一份论文红包，收下。 今年 AI 还会继续把高度往上拔。

2025送给赢麻了人，大盘标普16.5%纳指21.6%，最好的策略三个字“Buy the Dip” 反复All in/All out抓波段大空头之类的，反复被打脸，输不在智商，在情绪，在认知： 因为他们拿着旧地图走进“新时代”。2026呢？也是三个字“Audit the Cash” 首先活着才有资格谈“复利”，不但是上万亿capex巨头也是投资者 回看2025我做只有一个，从头条看到Gap，从怀疑中找到拐点。 特朗普2.0的拐点是关税没通胀；美联储的拐点是k型复苏降利率；Ai的拐点是中国DeepSeek开源让推理带动了上万亿美金超级算力中心投资，以及HBM内存的超级周期 $MU ；美国的制造业回流是带动Intel全球地缘政治 $INTC ；搜索要死的拐点是谷歌的王者归来 $GOOG ；潮流文化去Woke化的拐点是让 $AEO 的牛仔消费回流； 这些都是我从年初一路关注分析分享的优质的投资标的，实现远超大盘的超额收益，股票投资的价值，不同于比特币/黄金，就在于增长和背后人的创造，能发现宏观/行业/和企业的拐点就意味着巨大的投资回报，但一切最终都在体现在创造背后实现的“现金流”。 2026年，更简单也更残酷，不亏钱，先活下来，不要再去抓“热点”，要找到现金流涌现的地方，那里是金矿。2026我只做一件事：把Buy the dip升级成Buy the survivors。 当然25年我个人最有成就感是，超越了4万关注者，谢谢你们的鼓励。 也开了订阅和分享我们买方角度投资Memo并用清晰易懂的方式deck展示，是“档案”证明，更是对自我的审计和复盘，也是激励不断的输出就是最好的学习，在社媒领域。 ㊗️各位朋友 2026 新年快乐 🎉 明年希望下图还能不是割韭菜的“绿油油”。

People keep turning $NVDA Groq’s SRAM story into a meme: “More SRAM = no more HBM.” - Wrong, Why? Quick question: have you ever cooked and wondered why you’re “slow”? I t’s rarely the stove. It’s the constant trips to the fridge. SRAM is the prep counter next to the stove. Put the hot ingredients and tools within arm’s reach, and the workflow becomes smoother, more predictable. That’s the real Groq signal: not “less thinking”, but less waiting inside the system. But a bigger prep counter doesn’t eliminate the fridge. HBM is the fridge + pantry: the capacity and bandwidth you need to serve many customers, long contexts, and high concurrency. So no, SRAM doesn’t “kill” HBM. SRAM buys determinism. HBM buys scale. In the inference era, if anything, scaling real deployments can make HBM even more important. $MU I turned this into a simple visual. Link in first reply.

$MU 圣诞快乐🧑‍🎄🎄 今年我从年初一路分析跟进的核心AI时代的大礼包 前两天也全面从具体操作交易层面分享心得 今天复刻我的预测。市场给了多头最好的礼物：一个干净的、伴随回踩确认的历史新高突破。 不要被 RSI 的超买吓到，也不要被节前的缩量回落干扰。 现在的结构非常简单：天花板 ($277) 已经变成了地板。 只要站在地板上，我们就继续仰望星空。 不要追高，关掉软件，去过节。现在的波动不值得你浪费节日时光。 AI大时代 主要靠运气🍀 顺便看能力。Merry Christmas 🎄 全面深度看👇两个链接内容

一般说价值投资的都瞧不起技术分析，这是误区，也是投资行业中“交易员”的价值。 比如看 Intel $INTC 我是长期投资看好，作为全球半导体“第二极”的美国战略价值，这是终点，但是路径是坎坷，你什么时候买？是不是大涨要卖？ 资本市场是无数投资者的无沟通协调博弈。比如一个城市约陌生人又不能联系，大概率不约而同选地标，比如上海东方明珠或者香港时代广场纽约中央车站。 市场也一样，资金会在最显眼的价位附近集中挂单，流动性、止损、算法触发都堆在那里。 找到最简单所有人知道最“丑陋质朴”的那些线均线、黄金分割、RSI，这些就是“集合点”，周一就要吹上攻集结号了么？ 个人观点赔率1:5。仅供教育分享。个人持有。👇分析

Most “value investors” hate charts. They shouldn’t. If you invest on a 5-year horizon, today’s RSI is mostly noise. But if you ignore technicals completely, you’ll keep overpaying for your own thesis. My $INTC view is structural, but the short game is execution. In a fractured world, the U.S. needs a credible domestic manufacturing pole. That’s the long game. But where can I add exposure with cheap risk, and where am I wrong? Markets are a coordination game. If we had to meet in NYC with no phones, we’d both end up at Grand Central under the clock. That’s a Schelling point: the default meeting place. In a market of millions of strangers, nobody can call each other and say: “Buy here.” So the crowd converges on the same obvious landmarks: round numbers, the 50/200DMA, prior highs/lows, and yes, the lazy-looking Fibonacci zones. A Schelling point. Grand Central for liquidity. $INTC is a clean example of “cheap risk”. 36.3 (Fi 0.618) held as the first defense. 36.9 (50DMA) is the next tell. 39.8 (Fi 0.786) is the real wall. Clear it with volume, and 44 becomes a map. Fundamentals tell you WHAT to own. Technicals tell you WHEN to size up, and where the thesis is invalid. Conviction is long-term. Execution is tactical. Skip the fancy lines. Use the levels everyone watches.

This is a buy-side log, not a news feed. I post audited causal chains when the denominator shifts. Major events: live tape + options + structure. Different cost bases, same mechanics. Commentary, not advice. Full IC deck archive in comment。

虽然不保证每天都有“重大发现”。但每次出手我都给你一条可审计的因果链思考和实战价值。 你看我时间线就知道，只要遇到我真正关注的重大事件，比如美联储、行业估值拐点、公司定价被冲击关键时刻，我会在盘前盘中实时拆解。 买方怎么盯盘，怎么用期权和量价去判断短线强弱、波动和噪音。当然… 每个人成本和风险偏好各不相同，为自己下单负责。但我做的是把买方交易台的变量摊开，结构是否被接受，关键区间是否被守住，期权墙与对冲流如何改写路径，强弱到底来自公司还是指数。 你看到的是实战，不是课本。目标很简单，少被故事小作文牵着走，少把好仓位捐给洗盘，把纪律留在盘面里。 但也别误会，散户优势不是天天盯盘，是时间不是频率。你不需要向出资人汇报，也不需要用复杂期权证明自己很聪明。你只需要一套能让你睡得着的规则，让时间的复利替你工作。 同时对我也是一种纪律和记录。社媒公开表达本身就是一种约束。 把判断写下来，是在强迫自己对逻辑负责，对时间负责。这是一种记录，也是一种记忆，更是一种可回溯的验证方式。 这也是这个账号存在的原因：在关键时刻，留下真实的判断和时间戳，让市场，最终给出答案。 授人玫瑰，手留余香。分享不是牺牲，是对自己思考质量的长期投资。鼓励有更多中推的专业人士分享中英文高质量内容，现在推特日益有小红书化趋势。

我会有时间更新，不可能每天都有“重大发现”。但每一次更新，都是围绕潜在超越市场共识、可能产生 alpha 的那条因果链。 我分享的内容，本质上都是买方视角的深度研究，而且往往是重大性的、范式转移级别的判断。 很多东西，说得直一点，在全网都是第一次系统性写出来的。 这和卖方投行那套满大街天天吹故事、天天当头条，涨了就出来喊逻辑，跌了再找补不一样，本质上都是事后叙事验证“自己英明”。 我做的恰恰相反。所有判断，都是提前的、超前的，而且不是从财报、新闻或管理层表态出发，而是从最底层、最核心的逻辑和物理约束出发，去推演未来可能发生什么。 所以理解起来，确实会有一定难度。很多内容用英文写，也是因为这些问题本身就不是“情绪层面”的，而是系统性的。 真正难度关键的不是语言，而是你是否愿意建立系统性思维，用一整套框架去看问题。不然给你十倍股，也拿不住。 如果你一路看下来，会发现我做的其实是同一件事：用底层逻辑，去拆不同行业公司的范式变化。 比如英特尔的双极化定位，谷歌作为“全战型 AI 平台”的地位，美光的超级内存周期，Oracle 从 AI 建造商向云服务收费站的迁移，以及英伟达生态的延续性与竞争边界、美联储决议、复利概念等。 不是让你买入挣钱因为都是假设存在时间验证和证伪，但肯定存在独一无二的价值。 日积跬步以致千里。共勉。 要系统性的完整解读报告，都在我发布的 memo deck 里备查。关注

The regime changed quietly: efficiency stopped being a virtue. It became a liability. Taipei. An Ivy campus. Australia. Different contexts, same structure. High-order systems are now colliding with unmodelable, individual tail shocks. Not because the world “lost control.” Because we optimized control into fragility. For three decades, we engineered the mean. We centralized supply chains, standardized processes, compressed redundancy, and celebrated utilization. It worked, until the tail started pricing itself into daily life. Deglobalization is not a mood. It is risk budgeting. Political consolidation is not stability. It is correlation. Identity and belief hardening into tribes is not culture. It is narrative volatility. Semiconductors are the cleanest mirror. The world concentrated critical compute, advanced packaging, and geopolitical exposure into too few places. In the mean, it looks like dominance. In the tail, it is a single-point failure with a global blast radius. This is where Intel matters, even if you dislike the stock. Not as “the next TSMC,” but as a second pole. A partial US-based stack changes the distribution: more routes, more optionality, less world-wide correlation to one island and one strait. In a tail-driven regime, partial redundancy is not a nice-to-have. It is a valuation input. The market is slowly repricing from growth premium to resilience premium. Not because resilience is inspiring. Because it is scarce, measurable, and survivable. The question is no longer whether a shock happens. The question is whether you still have options when it does. Company resilience is redundancy, multi-path supply, and onshore capability. Portfolio resilience is liquidity, hedges, position sizing, and detoxing from single-node dependence. Mean returns are easy to model. Tails decide careers.

The retail edge isn’t timing. It’s structure. Here’s the only framework I trust in “AI audit season” core+tactical. $MU Most people treat earnings and macro like a casino: call the move or confess you’re wrong. That’s not investing. That’s dopamine management. Buy-side doesn’t get paid for being clever. It gets paid for staying solvent through regime shifts. Core is your right to be there. It’s held on the thesis clock, not the candle clock. Tactical is path management. It exists to make sure volatility doesn’t tax you out of a good idea. Two rules. No extra religion. When the tape is strong, I harvest a little certainty. Not a full exit. Not a victory lap. Just enough to lower my emotional cost and stop overtrading. When the tape is weak, I only add if structure is intact. Not “it’s red.” “it’s red and getting absorbed.” That’s the difference between a discount and a knife: A discount drops, then stabilizes. A knife drops, then accelerates and fails to reclaim. And stop worshipping levels like buttons. “Did it break 260?” is the wrong question. The question is: did the market accept above it by the close? Wicks are noise. Closes are verdicts. If you internalize that, you stop donating good positions to the wash cycle. You stop confusing motion with information. AI audit season doesn’t pay you for predicting candles. It pays you for staying on the right side of structure long enough for the thesis to compound.

$MU 这种级别跳空高开，短线如何有纪律去操作，我昨晚分享了。 现在问题是背后是几乎所有的市场预测都被打穿。机构分析师肯定是模型算错了，都知道股价是未来现金流的折现。 那么拿什么折现？取决于你看不是今天的涨幅，而是未来能拿住么？能追涨么？如果回落需要加仓么？ 我把这新模型系统分析 👇 明年 $NVDA 新一代芯片会带来什么样的估值冲击，对于整个AI价值链？哪里还有超越市场预期的超额收益和风险？ 还在数 GPU 的人会一直看不懂。真正的分母是 1GW：带宽、HBM、封装、供电冷却才是定价权。

美光 $MU 大概率高开，年化汇报超1.6倍了。 从年初我首提超级内存周期，一路跟到现在定期评论分享，难在如何跳出传统的周期性行业，不要去线性代数推到看问题。 特别是明天可能的几种短短的应对纪律和策略拆解，以及长期周期的逻辑演变下一步在哪里？如下 ⬇️

Don’t treat Micron’s $MU earnings as a victory lap. This is audit season, not celebration season. The easiest way to die in memory is to mistake a loud quarter for a long cycle. You think you’re holding “the future,” then wake up trapped in a perfectly familiar loop with a bigger amplitude. I’ve been long this memory cycle since January, and I’ve been updating the thesis through multiple prints. The return so far is roughly 1.8x. But P&L is the easy part. The hard part is not giving it back right when the crowd starts getting emotional and the market starts getting judgmental. Because the real question is not “did $MU beat.” The real question is whether this is a durable shift in pricing power, or just a well-lit peak inside a cyclical business. Memory is ruthless that way. It rewards people who refresh inputs, not people who fall in love with a chart. So keep the model honest. Stress-test the same core dilemma every time: super-cycle, or fake long cycle. In plain English, can HBM supply discipline and pricing power stay tight beyond the excitement of this print, or does supply, mix, and competition quietly loosen the grip while everyone is still tweeting victory? Now the part that matters most for retail: discipline beats prediction, especially here. Before you think about tomorrow’s open, define two numbers about yourself: your real mental cost basis, and the drawdown you can tolerate without breaking your process. If you don’t define your limits, the market will do it for you. Usually in the least forgiving way. As for tomorrow, I’m not interested in pretending I can predict the tape to the dollar. I care about walking into the open with pre-committed behavior, so I’m not trading adrenaline at 9:31. Scenario one: gap up, push, then fade. This is the most common because headline gaps attract fast profit-taking. The retail mistake here is swinging between greed and panic, either dumping everything on the first red candle or chasing back in higher because it “felt like it got away.” The clean move is simple: trim some into strength to reduce risk, then let the rest ride without you needing to be a hero. Once some cost comes off the table, your mind gets quieter, and quiet minds make better decisions. Scenario two: gap up that holds, dips get bought, structure stays tight. This is where institutions are re-rating, not retail celebrating. The retail mistake is trying to sell the exact top and then discovering the train did not come back. If the post-gap support holds, let it run. Watch structure, not ticks. When support breaks, reassess. Until then, don’t micromanage a trend that’s doing your job for you. Scenario three: no fireworks, maybe even a dip first. This is actually retail-friendly because it forces the market to cool itself and show real demand. The retail mistake is catching the first knife because it “looks like a deal.” Don’t. Let it stabilize, let buyers show up, then add in small steps. You’re not buying the bottom. You’re buying probability. Whatever path we get, the tells are the same and you don’t need to turn them into a religion. After the initial move, is there real buying underneath, or does it go hollow. Does volume persist, or does it evaporate after the first push. Does Micron lead memory and semis with it, or is it partying alone. And does macro flip the table through rates, risk appetite, or index pressure. See those clearly, and you won’t get shaken out by a couple of noisy candles. I’ll end where I always do. Long-term money here is not about guessing tomorrow’s open. It’s about conviction in HBM pricing power plus supply discipline, and the willingness to keep updating the thesis as facts change. Options are tools at best, mostly to help you sleep. They decay. Don’t turn tools into beliefs. The core question never changes: can this part of the AI stack keep converting demand into cash flow. If yes, don’t hand over a long-term edge just because the short term got loud.

If this guide shocked your model about $MU, you’re using the wrong denominator. Stop counting chips. Start pricing 1GW. Full deck: the migration of pricing power from GPUs to HBM, packaging, and the grid.