Tigris | 会讲课教授是好老师 · X 阅读

The market asked for an AI audit. Micron $MU just dropped the receipt. $18.7B next-quarter revenue. $8.42 EPS. The reason most models missed it is simple: they still price AI by GPUs. The real denominator is 1GW. In my 1GW framework, GPU counts barely move. But the bill explodes where the system actually chokes: memory wafers. My slides called it months ago: Memory Wafers +139%. Tonight’s guide is what you get when the bottleneck migrates from “compute” to bandwidth + power. Don’t trade the headline. Trade the constraint.

No panic. No euphoria. Just audit season. The door isn’t closing, it’s being priced. Retail investor doesn’t need forecasts. It needs rules: presence, entries, and emotional immunity. This wasn’t an “AI is strong” earnings season. It was the first audit season. Whether data centers land on time. Whether hyperscaler capex becomes durable cash flow. Whether the backlog is future money, or future explanations that require financing, time, and a lot of narrative stamina. Rates are still high. That makes the game cruel. You don’t just need growth. You need growth that clears the rate line. You don’t just need a story. You need a business model that can survive being audited. That’s why the tape has been telling you the truth in plain sight. $ORCL can talk about massive commitments and still roll over from highs. $GOOGL can trade like a “winner” and still slip under 300. Semis can look unstoppable one month and suddenly trade heavy the next. That’s not “AI died.” That’s the market quietly forcing one question onto every balance sheet: show me the cash. Micron is a perfect window into this regime. $MU earnings are not a referendum on “memory strength.” They’re a referendum on whether the slope of the AI supply chain can still be believed. In audit season, guidance doesn’t have to be bad to get punished. It only has to be not strong enough. And here’s the part most people miss: institutions and retail are playing different games on the same table. Institutions are not constrained by conviction. They are constrained by not getting hurt. In a crowded, low-cash environment, any sentence that doesn’t sound airtight becomes a reason to de-risk. Their job is to cut tails, protect sleep, survive the next IC, and still have a story for LPs. Retail has a different edge. You don’t report to anyone. Your advantage is time. So the retail framework is simpler and more uncomfortable than most “earnings takes”: Stay present. Even if it’s small. You can’t control events, but you can refuse to be absent from the trend. Make adds rule-based, not emotion-based. When it sells off, ask one thing: is the slope being rejected, or are expectations being cleared? If it’s expectation clearing and the chain isn’t broken, that drawdown is a discount, not a disaster. And stop confusing complexity with intelligence. Options aren’t a personality. For most people they have one legitimate purpose: Helping you sleep. If you can live with volatility, spot is cleaner. What you need is a rule set: where you add, how many times you add, and where you stop. Audit season doesn’t teach you how to “call direction.” It teaches you which game you’re actually playing. And the biggest mistake retail makes in this regime is borrowing institutional fear and calling it “risk management.”

Everyone is asking the wrong AI question. It is not “can AI beat the market?” It is “how do I wire a model that never sleeps into a human who feels panic, greed and regime shifts, and still come out ahead?” I call that Vibe Investing. I turned it into a working playbook> AI is already elite at the boring middle,like compression, scenario trees, risk discipline. Humans still win where the distribution breaks: structural shifts, reflexivity, taking real PnL and career risk. This is exactly what my new Vibe Investing bundle on Gumroad is built around: >a deck on human–AI collaboration in real shock events >the Tigris IC Memo Master Template I actually use for investment committee >a few small tools for structural breaks, crisis prompts and trade logs. If you are a serious retail investor, it is like having a buy side CIO sanity: check your next panic decision. If you sit on the professional side, it is a way to plug AI into your process without becoming replaceable by it. Details and link in the first reply.

The Oracle dump is rational… and that’s exactly why it might be wrong. Yes, Phase 2 is a cash-flow audit of the AI buildout. Builders get hit first. But the market’s shortcut is dangerous: it’s treating $ORCL as a pure infra bet. ignoring what happens when AI value migrates downstream into enterprise software, databases, and workflows. That’s where margins live. If AI actually becomes enterprise productivity, Oracle doesn’t just rent GPUs, it monetizes the software layer sitting on top of the world’s private data. The selloff may be pricing near-term financing risk as permanent impairment. That’s a classic misread. Link in reply.

Full Alphabet IC memo (Gumroad): Written for people who already know $GOOG is a core asset, but refuse to trade the tape like a mood swing. It does three things: -Pins scenario valuation ranges on GOOGL so +10% / -10% is context, not drama -Ties those ranges to explicit assumptions and catalysts, so you know what has to be true -Marks the zones where price and fundamentals decouple, which is where alpha lives, not comfort Link:

There is a little pattern in the tape that has become almost funny. Every OpenAI launch triggers the same reflex: the demo drops, the timeline melts, and Google trades down like it just lost another limb. If you think every OpenAI launch means Google is dying, you’re trading the surface story. OpenAI and Google aren’t even valued on the same axis anymore. One is drifting toward frontier infra plus agent layer economics. The other still earns its keep as the world’s biggest attention marketplace. The tape reacts as if OpenAI is taking Google’s lunch; in reality they are barely playing the same game. And yet Google is not some cornered incumbent. It runs its own stack end-to-end: silicon, models, cloud, distribution. When the frontier moves, Google isn’t just a victim of the headline. It’s also a beneficiary of the underlying economy the model boom feeds: more tokens, more workloads, more infra demand, including on its own cloud. The real question for investors has nothing to do with “coolness” or who grabbed the news cycle. It’s whether the market is still pricing Google purely as an ads machine, or whether you’re getting its entire AI optionality for free every time panic selling hits the tape. If you can tell the difference between a narrative drawdown and an actual earnings threat, you’re at least playing the stock on the right dimension.

分享一个我常看的指数，市场上钱多少比听鲍威尔和看头条靠谱。 只要是负值，趋势还不变，不用担心什么AI泡沫和Capex投资小作文，市场上钱多还是会追和托底高杠杆、高 capex、押注 AI 远期现金流的资产。

Forget Fed's rate cut and Powell's script. In the Trump–Fed knife fight he is basically a caretaker on borrowed time. If you want to know whether risk still has a bid, you watch Chicago Fed’s NFCI and ANFCI, not the press conference. Both indices are sitting around −0.5. Zero is “normal,” positive is tight, negative is loose. NFCI, a composite index of risk, credit and leverage that maps the real liquidity regime. NFCI < 0 already tells you there is no broad credit crunch, no systemic “funding winter.” ANFCI is even more negative, which means that after you strip out the macro cycle, financial conditions are still looser than the economy alone would justify. Translate that into plain portfolio language: high policy rates do not equal tight financial conditions. As long as NFCI and ANFCI stay below zero, the system is still willing to fund stories with long duration cash flows and heavy capex. The AI capex wave, the four-to-five-hundred-billion-a-year data center buildout, high-multiple semis and software names all sit on top of that liquidity bedrock. So my playbook for tomorrow is simple. After the meeting, I am less interested in whether Powell sounds “hawkish” or “dovish” and more interested in where NFCI and ANFCI drift over the next few weeks. If they stay negative while the Fed is already cutting, then monetary easing is not being neutralized by markets, it is being amplified. That is how you extend an equity boom, not how you kill it.

H200 isn’t about geopolitics. It’s about whether $NVDA keeps its edge, how HBM/CoWoS get repriced, and whether this AI bubble bleeds out or detonates inside your book. That is where the alpha really lives. China builds indigenous silicon regardless of H200. The only real variable is whether that progress compounds inside CUDA or outside US control and that’s what actually reprices NVDA, HBM names, and the entire AI cycle. “Treason vs survival” is the wrong frame for H200. People say NVDA is “selling out America” or DC is “gifting China compute.” None of that matters. With or without H200, China will build its own silicon. The real trade is… how much edge the US gives back, and on what terms, to keep control of the game. If your mental model is “block everything or lose the lead,” you’re still reading the cartoon version. You cannot stop indigenous silicon any more than you could stop fabs in Taiwan or Korea. The only question is whether that progress compounds inside your standard or outside your reach. Cut China off completely and you force a hard break from CUDA. Painful now, but long term you invite a parallel stack that never needs to listen to Washington again. Ship H200 under caps, tariffs and high prices, and you compress the gap but keep their best engineers and applications living inside CUDA while their money funds your next node. Underneath that sits the real border: packaging. CoWoS, HBM and binning create a stratified compute world where different blocs get different silicon, different thermals, different latency budgets. US hyperscalers get top shelf; China gets a monetised, constrained tier. So H200 isn’t about exports. It’s a live test of NVDA’s true moat, process vs ecosystem tax , and how much compute spread policymakers are willing to trade for a controlled landing of the AI cycle. Full buy-side breakdown, compute, incentives, bottlenecks, portfolio impact is in my earlier tweet.

Intel $INTC quietly did something with Amkor $AMKR that matters more than any AI day-trade chart. Everyone saw AMKR spike. The real story is what this says about Intel’s role in the next phase of the AI cost curve. Putting EMIB into Amkor’s Korea line is not “Intel giving up control”. It is Intel admitting two things at the same time: >internal packaging capacity is already at the limit of what AI is asking from it, >keeping EMIB locked inside its own fabs caps its strategic value. Once you teach an external partner your most political packaging stack, you are not selling a service, you are placing a piece on the board. You are telling hyperscalers and governments: there is now a second channel for advanced AI packaging that does not begin and end in Taiwan. That changes behavior. Cloud buyers suddenly have a credible way to split risk between two queues instead of fighting over the same CoWoS slot. Policy makers see a concrete object they can fund and protect, not just another 18A slide. Vendors stuck behind TSMC’s backlog have a real option other than “wait and pray”. The tape still trades Intel on PCs, quarterly foundry logos and whether 18A “catches up”. It barely prices the fact that packaging has become the real bottleneck and Intel is slowly moving from “struggler in wafers” to “necessary second pole in packaging”. Will a lot of early EMIB volume be overflow rather than pure wins? Probably. Can AI capex overshoot and turn shortage into overcapacity later? Absolutely. But that is not the point today. The point is that Intel just turned EMIB from an internal toy into a networked standard with a geography and a partner that fits the new risk map. Most people saw a pop in $AMKR. I saw Intel taking one quiet step closer to being the only acceptable Plan B in the AI hardware stack.

For people asking what the deck actually is: It is not an AI prompt pack and it is not a beginner course. It: >uses the DeepSeek event and Nvidia’s full price path as a stress test for AI systems, analysts and traders looking at the same shock, >treats Alpha Arena as a real lab for Grok, GPT 5.1, Gemini and others under identical prompts and data, >forces both into a three layer human AI stack for investing: information, reasoning, execution, >ends with two playbooks: an “AI shock” protocol for event days and a weekly human AI workflow for research and risk. Written in the same format I would use for an internal training with a research, PM and trading team. Gumroad link:

“Guess we finally figured out how to pay for all those GPUs lol.” Musk was joking. He was also selling Grok 4.20 as the smartest trader in the room. The timeline heard a different promise: plug an AI into your broker, fire the analysts and traders, get rich？ Reality: Alpha Arena is a two week, stateless trading lab. The platform scrapes news and sentiment, turns them into structured prompts, and asks different models to react under identical constraints. It shows what frontier LLMs are already good at in markets, and where they still hit the wall. My read is simple: >AI is already elite at the boring middle of the stack: digesting flows, mapping scenarios, enforcing risk discipline. >Humans are still needed at the edges: regime shifts, game theory, and deciding when a model’s world view is too narrow to trust. That is the gap I try to close in a new deck, “Investing in the Age of Intelligent Machines: Human AI Collaboration as a Source of Sustainable Alpha”. It uses the Grok vs GPT vs Gemini experiment and the DeepSeek shock to build a concrete workflow for PMs, analysts and traders. Details in first reply ⬇️

普通人对“复利”的低估，远比对任何一只股票的低估严重得多。很多人一辈子忙于还账单、赶通勤、处理琐事，从来没机会认真想过。 真正决定阶层差距的，不是谁更辛苦，而是谁更早学会让“资产”和“时间”一起工作。 穷人的悲剧在于，从小活在复利的“场”里，却从来没人告诉他们，这东西可以被自己利用。 没有人跟他们说过：你的时间用来换工资，资产用来换未来。前者是体力，后者才是杠杆。 所以，复利不是理财技巧，而是一个从小就该被安装的底层认知操作系统。绝大多数人缺的，不是自律，而是这块心智软件从来没被更新过。 这是我《第八大奇迹：复利的八个底层规律》里的第 1 个概念： “ 人类系统性低估复利和指数增长的威力，也低估了“拥有资产”这件事本身的意义。” 我把完整的图和例子整理成一个小课件，有兴趣可以系统看一遍，或者讲给你的孩子听。

“Your labor pays today. Your assets pay the rest of your life.”If someone had sat me down early and said. That single sentence would’ve saved me more than most hedge fund courses ever did. People dunk on Dell’s donation because they look at the number and not the mechanism: If you think 250 dollars can’t change a kid’s life, you’ve clearly never met compounding or poverty. Sure, 250 compounding to 1,400 isn’t life changing. Everybody with a Bloomberg terminal knows that. But here’s the part people with smooth lives always miss: in low income households, the idea of money working for you is not just absent, it’s unimaginable. Kids grow up learning that money only moves when you bleed for it. No assets. No ownership. No leverage. Just hours, shifts, and bills. When survival is the OS, investing never installs. The first time a kid sees their balance move because the market moved, something rewires.Curiosity kicks in. Agency kicks in. That’s the real return. And this is where people underestimate compounding, the math is secondary, the mindset is the multiplier. You don’t need every kid to become Buffett. You just need a few to realize “this world isn’t closed to me.” Wealth gaps don’t explode because rich people are evil or poor people are lazy. They explode because one group grows up watching compounding in real time, and the other group grows up watching their parents juggle overdue bills. Different inputs. Different loops. Different outcomes. It’s system dynamics, not morality. So installing the compounding mental model is the point. That’s the real eighth wonder of the world. Everything else is just arithmetic. If even a small minority walks away with three things -a visceral understanding of compounding -a habit of owning productive assets -the memory that “my money worked while I slept” If you want to actually understand this force the way buy side people do, not the TikTok version, I builded a compact module breaking down compounding as a physics problem, not a slogan. It’s free on Gumroad. Download it. Study it. Then share and teach it to someone younger than you. That is how you quietly change a balance sheet, and a life, over time. Link in comments

The Eighth Wonder: Compounding and Investment Principles Link:

Many people aren’t long or short Google. They’re long or short the last 10% move. This IC memo is not another target. It is a discipline for $GOOG How to size, when not to sell, and when price action is mispricing instead of information. Link in comments.

Full Alphabet IC memo (Gumroad): Written for people who already know $GOOG is a core asset, but refuse to trade the tape like a mood swing. It does three things: -Pins scenario valuation ranges on GOOGL so +10% / -10% is context, not drama -Ties those ranges to explicit assumptions and catalysts, so you know what has to be true -Marks the zones where price and fundamentals decouple, which is where alpha lives, not comfort Link:

“Nvidia now, Google later” is a good headline Musk gave you. I built the IC memo that underwrites it on cost of intelligence, not vibes. Full Alphabet IC memo (Gumroad): Written for people who already know GOOGL is a core asset, but refuse to trade the tape like a mood swing. It does three things: • Pins scenario valuation ranges on GOOGL so +10% / -10% is context, not drama • Ties those ranges to explicit assumptions and catalysts, so you know what has to be true • Marks the zones where price and fundamentals decouple, which is where alpha lives, not comfort Link:

One week ago I published a buy-side style Intel memo around $36, priced at $4.99. Today the $INTC stock is ~$43 after a +10% day and a +6% day. The point of Tigris Research is not “I called the bounce”, it’s to put a clear, one-read framework on paper before the move: packaging, foundry, Apple 18A, and the “post-Taiwan” re-rating path. Current price for the memo is $7.99. Link in the comments.

You are a fan, not an investor. If the only thing you hear from that Musk interview is “I don’t buy stocks, I only build”. The real tell was when he quietly told you who gets to be AI king today and who is being groomed to run the empire next. Today Nvidia is king. Tomorrow, the most valuable operator over the next decade is Google. Why? Most people heard “Nvidia is strong, Google is also good.” That’s not how decision-makers hear it. Any CEO with a few billion in AI capex on the line heard something very different: What keeps them up at night is not “will my product sell,” it’s “will my supply chain hold.” Right now 80–90% of global AI acceleration sits on one vendor. That’s not a moat, that’s a single point of failure. That’s why you see the so-called contradiction that isn’t a contradiction at all: everyone is still scrambling for Nvidia’s parts, while every major buyer quietly builds a second leg – in-house ASICs, TPUs, custom inference chips. Even 5–10% penetration is not about saving money this quarter; it’s buying a survivable alternative. The moment a “backup” becomes a production-grade line, the gameboard changes. Valuation follows that physics. For Nvidia $NVDA, the last two years have been priced on almost perfect assumptions: share north of 80%, gross margins in the 75%, ten years of cash flows discounted as if nothing structural intervenes. But if multi-sourcing is a physical necessity, that curve cannot extend linearly. Volume can grow, but share and margins will be mechanically squeezed; the imagination premium in the DCF bleeds back toward normal. Flip the frame and Google $GOOG becomes the underpriced side of the same story: it builds the models, designs the accelerators, owns the cloud and the data-center economics. That is exactly the profile the market is still valuing like “an ad business with AI optionality.” That gap is the real upside. I’m not interested in the emotional “can Nvidia be killed” debate. The real question is more primitive and more predictive: who can rewrite the cost of intelligence. Once you put power, packaging, optical interconnect, topology and utilization into the same equation, this stops being a chip duel and turns into a long chain about how system-level TCO (Total Cost of Ownership)is going to be redistributed between two architectures. I’ve built a full cost-of-intelligence framework that pulls Nvidia and Google onto the same coordinate system. The full IC memo is where I open that chain end to end and show exactly how the curve bends. Link in comments.