The White House just published a scorecard of every factory announced since January 2025.

Strip out the politics and there’s a real capex map inside.

Most of these projects haven’t been built yet.

That’s where investors can still position themselves.

You can read the full report here: American Manufacturing Renaissance Report (White House, Nov 2026).

I pulled the numbers out of the speech so we can read this objectively and spot the real opportunities.

What the report claims

The headline numbers:

The macro

• ISM manufacturing PMI back above 50 in January 2026, and above 50 for nine straight months.

• Construction employment rose 109,000 over the 12 months.

The sector map

1. Semiconductors: $159.8B

• TSMC’s $100B Phoenix expansion is the anchor.

• Micron ($25B) and the SpaceX/Tesla “Terafab” ($16.8B) are next.

• Six fab projects worth $149.3B. Three packaging and test plants worth $5.3B. Four supply projects (wafers, photomasks, probe cards, parts) worth $4.6B.

2. Pharma and biotech: $63B

• Eli Lilly owns the top three projects, worth $25.5B in total.

• Biogen is adding $2B.

3. Metals, minerals and materials: $61.3B

• Mesabi Metallics (private) is behind the biggest steel project.

• Korea Zinc is building a $7.4B critical minerals smelter

• Six rare earth magnet or processing projects total $4.2B: MP Materials, USA Rare Earth, Vulcan Elements, LS Cable, Energy Fuels and Cyclic Materials.

• Corning has a $2.5B cover glass line for Apple.

4. Aerospace, defense and ships: $29.9B

• Anduril’s $3.7B Arsenal-2 in Maryland and Hadrian’s $1.5B.

• Saronic’s $3.2B autonomous vessel yard.

• Boeing is investing $1B; L3Harris is expanding its satellite site for Golden Dome.

5. Vehicles and parts: $27.1B

• Mercedes-Benz: $4B

• Toyota: more than $3B for pickups.

The rest of the map

• Food and consumer goods: Chobani ($2.4B dairy) leads.

• Energy and grid equipment: Oklo, LIS Technologies and T1 Energy lead.

• Electronics: Amazon, Wiwynn, Prysmian and Celestica lead.

Subscribe now

The fine print investors should read

• Only about 85 of 522 projects are producing or under construction.

• One company (TSMC) is over a quarter of the total dollars. Semis are 42%.

And the bottlenecks the report mostly skips:

• Power. The report itself cites a 30% transformer shortfall.

• Labor. Factories now compete with data centers for the same electricians. T1 Energy raised its Texas solar fab budget 20%.

Who benefits, by sector

My read, from most direct to least:

• Chip fabs and their supply chain. The biggest dollars. Tool makers, packaging and materials ride the build.

• Electrical equipment and grid. Every plant needs power, switchgear and transformers. Supply is already short.

• Domestic metals. Steel and aluminum, plus anything rare earth or critical mineral with a defense buyer.

• Defense primes and shipyards.

• Construction and industrial automation.

How this hits my portfolio

Read more

Friday was ugly. Not just for stocks. For almost everything.

Stocks, gold, silver and crypto all fell.

When everything falls together, it’s usually not about one story. It’s about forced selling. And forced selling creates opportunities.

Let me walk you through what happened, what didn’t change, and what’s coming next.

Gold fell to a one-year low, under $3,900. Silver broke below $50 for the first time in 11 months.

That’s not normal. Gold usually goes up when stocks go down. This time it didn’t.

The biggest de-grossing since 2016

Hedge funds cut their gross exposure by 7.8% from the peak. That’s the largest de-grossing in this data since 2016.

What de-grossing means

When funds “de-gross,” they sell their longs and buy back their shorts. They shrink the whole book.

It usually happens when funds are losing money and need to cut risk fast.

Why it matters for you

Forced sellers don’t sell what they want. They sell what they can.

That’s why gold fell with tech. That’s why good companies fell with weak ones. The selling was about positions, not about each business.

Morgan Stanley’s data also showed hedge fund net leverage at 47% in late October. That’s near the bottom of its one-year range. Translation: a lot of the risk has already been taken off.

De-grossing can last longer than you expect. But history says these moments were better for buying than for selling.

Subscribe now

What happened this week

The Fed held

• The Fed kept rates at 3.75% to 4.00% on Wednesday.

• Traders now see a 70%+ chance of another hike in December

Big tech: spending more, cash flow falling

This was the big earnings week for the giant tech companies.

The message was the same everywhere. AI spending keeps going up.

• Amazon, Alphabet, and Meta now show negative free cash flow.

• Alphabet had its worst day on Thursday. $500 billion was gone in one session.

• Intel fell 29% on Friday. Its worst day on record.

• Meta is planning a record $97.5 billion fundraise.

Gold and silver: no safe place

• Gold had its worst October since 2008. About -6% for the month.

• Silver lost about 18% in October.

Why would gold fall in a selloff? Margin calls. When funds lose money on stocks, they sell what’s liquid to raise cash. Gold is very liquid.

• 82% of S&P 500 companies beat earnings estimates so far.

• Q3 earnings growth is expected around 30%.

• The economy grew 3.4% last quarter.

When prices move a lot and fundamentals move a little, that gap is where opportunities show up.

Every dip has a story. This chart is a wall of them.

From 2009 to now: flash crashes, credit scares, pandemics, rate hikes, AI panics. Each one felt like a reason to get out.

This week’s headline is Mag7 / AI earnings fear. It belongs on the chart. It does not change the long-term path.

Selling because it feels scary is usually the expensive choice.

What’s coming next week

A big week. Midterms, jobs and a lot of earnings.

A quick thank you

Thank you for the kind words lately.

Messages like these keep me going. I’m grateful you’re here, in the good weeks and the rough ones.

Portfolio Performance

Last 2 years: +148.5%

YTD: +64.3% vs. SPY +10.7%

Let’s dive in...

Read more

Marvell was up 197% this year by Wednesday’s close.

Then Bloomberg dropped a headline after the bell. Marvell is in advanced talks to buy Lumentum, a laser and optics maker worth about $83 billion.

Marvell fell about 9% overnight. Lumentum jumped.

So the market already gave its first answer. I want to understand the question first.

What does Marvell actually do? Why would it want a laser company? And can it even afford one?

Subscribe now

What Marvell actually does

Most people know Nvidia makes the AI chips. Fewer people know who connects them.

That’s Marvell.

An AI cluster is thousands of chips that have to talk to each other all the time. They talk inside a rack, across racks and across data centers. Every one of those links needs silicon. Marvell makes a lot of it.

It also helps the cloud giants design their own AI chips.

Today 79% of revenue comes from data centers. Five years ago Marvell was a storage and networking chip company. It’s a very different business now.

The Technology

1. Custom AI chips (custom silicon)

Amazon, Microsoft, Google and Meta don’t want to rent Nvidia forever. They design their own chips for their own workloads.

Marvell helps them build these chips. It brings the design team, the high-speed links, the memory interfaces, the packaging and access to TSMC’s newest nodes.

• XPU = the main AI chip (think AWS Trainium-type programs).

• XPU attach = the chips around it: network cards, memory expanders, storage and management chips.

• In August, Google signed a long-term custom deal with Marvell around its TPU ecosystem. Marvell gave Google a warrant for up to 59 million shares that vests as Google buys more.

• Management targets >$12B of custom revenue in FY29. In FY26, custom was around $1.5B (per Investor Day coverage).

2. Electro-optics (Marvell’s crown jewel)

Data moves between racks as light. Inside every 800G and 1.6T optical module sits a chip that cleans up and translates the signal. It’s called a DSP.

Marvell is a leader in these. It came from the Inphi deal in 2021. This is the business that has quietly carried the growth.

3. Switching

The Teralynx switch chips route traffic across the cluster. Marvell now has a 102.4 terabit switch. Broadcom is the giant here, so Marvell is the challenger.

4. Scale-up optics (the next big thing)

Today, the chips inside one AI rack talk over copper. At 200 gigabits per lane, copper only reaches about 2.5 meters (Investor Day figures).

So you can’t connect more chips without switching to light.

Marvell bought Celestial AI for this. It closed in February. Its “Photonic Fabric” connects AI chips with light instead of copper. Marvell expects it to reach a $500M annual run rate by late FY28 and $1B by late FY29.

This is why the Lumentum talks make sense on paper. More on that below.

5. Storage and memory

SSD controllers, CXL memory expanders and PCIe/CXL switches. Smaller, but they ride the same AI wave.

The numbers

Last quarter (Q2 FY27, reported Aug 27):

• Revenue: $2.74B, up 37% YoY. A record.

• Data center: $2.17B, up 46% YoY and 18% from the prior quarter.

• Non-GAAP gross margin: 58.9%. GAAP: 53.1%.

• Non-GAAP operating margin: 36.6%. GAAP: 16.8%.

• Non-GAAP EPS: $0.94. GAAP EPS: $0.33.

• Operating cash flow: $606M.

Next quarter (Q3 guide):

• Revenue: $3.15B +/- 5%. That’s about +52% YoY.

• Data center: up more than 20% from Q2 and about 75% YoY.

• Non-GAAP EPS: $1.10 +/- $0.05.

Growth is speeding up. Q1 data center growth was 27%. Q2 was 46%. Q3 is guided at ~75%.

Those are company targets. Not facts.

Balance sheet:

• Cash: $3.93B. Debt: about $4.96B.

• Nvidia invested $2B in convertible preferred stock in March.

• Buybacks of $200M and a small dividend ($0.06) every quarter.

What the stock did

• Oct 28: $251.65 close. Then ~$228 premarket after the Lumentum report.

Even after the drop, Marvell is still roughly 2.7x where it started the year.

“The next trillion-dollar company”

In June, Jensen Huang walked on stage at Computex in Taipei with Marvell CEO Matt Murphy. His first words:

“The next trillion-dollar company, ladies and gentlemen.”

Two things to keep in mind:

• Nvidia is an investor in Marvell. It put in $2B in March. Jensen is not a neutral analyst.

• Marvell is worth about $226B. $1 trillion is about 4.4x from here.

Is that realistic? I ran the math below.

Who runs it

• Matt Murphy, CEO since 2016 and also chairman. He turned Marvell into a data center company through deals: Cavium (2018), Inphi (2021), Innovium (2021), Celestial AI (2026).

• Dan Durn, CFO since June 2026. Previously CFO at Adobe, Applied Materials and NXP. He moved over from Marvell’s own board.

• Founded in 1995. Sold its automotive Ethernet unit in 2025 to focus on data centers.

Murphy is a dealmaker. Keep that in mind for the Lumentum news.

The good and the bad

Subscribe now

Yesterday’s news: Marvell and Lumentum

What was reported (Bloomberg, Oct 28):

• Marvell is in advanced talks to buy Lumentum (LITE).

• A deal could be announced as soon as next week.

• Talks are private. They could still fail. No price or terms reported.

• Lumentum was worth about $83B. Marvell about $221B.

• It would rank among the 10 biggest tech and telecom deals ever (Bloomberg data, via Mint).

How the market reacted:

• MRVL: about -8% to -9% after hours. ~$228 premarket this morning.

• LITE: about +10% after hours. ~$972 premarket.

• Bernstein told Reuters a deal this size would likely be impossible to fund fully with cash.

The questions I can’t stop thinking about

• How do you buy an $83B company with $3.9B in cash? Stock? Debt? Both? Each option hits shareholders differently.

• Why did Marvell double its credit line six days before the leak? Is that a hint, or just housekeeping?

• What does a deal do to earnings per share? Is it dilutive? By how much?

• What if they buy it? Does Marvell become the full-stack optics player Broadcom already is? Or does it overpay at the top of the laser cycle?

• What happens to everyone else? If Marvell owns a key laser supplier, do rivals lean harder on Coherent and other independents?

• What if they don’t? Does the stock bounce back to $250+? Or does the market worry about the next deal?

• Is Jensen’s $1 trillion even possible? What would the stock price need to be?

• What’s my price target, and what would make me change it?

Two possible outcomes. Six scenarios. My price targets.

Here’s how I’m positioning for each.

Subscribe now

Paid: Valuation, Targets & Plan

Read more

Trading in short ETFs just hit a record high.

So a lot of people are already positioned for a drop.

When everyone is on one side of the boat, it doesn’t take much to tip it the other way. Good news, or just “less bad” news, can force shorts to buy back. That can push prices up fast.

I’m not saying the bottom is in. Nobody knows that. But crowded fear has often been a better time to buy than to sell.

Join the subscriber chat if you haven’t.

I share data points, observations, and insights through the week. When I trade, I post the idea and my cost basis there too.

Portfolio Performance

Last 2 years: 149.7%

YTD: 67.4% vs. SPY 11.3%

Let’s dive into the portfolio changes…

Read more

If the last few weeks confused you, you are not alone.

The headlines say tech is booming. Yet many tech stocks just fell 20%, 30%, even 40%.

That does not add up. So people panic.

I want to help. This post explains what happened in plain English.

By the end, you will know how to read a sudden crash. And what to check before you sell anything.

Same selloff, very different damage

From Sep 30 to Oct 22, the chip ETF SMH fell 14.1%.

CoreWeave fell 36.2%. Nebius fell 32.5%. Nvidia fell only 7.1%.

Berkshire Hathaway? Up 9.8% over the same days.

Meanwhile, the businesses kept growing:

• Meta grew revenue 26% last quarter, to $64.6 billion.

• Meta also raised the floor of its AI spending plan, to $135 billion this year.

• That spending flows straight to chip makers and AI cloud builders.

So what broke? Mostly, the plumbing of the market.

Three reasons a stock falls

1. The business breaks

Sales shrink. Margins fall. The story changes.

Nike is a clear example. Revenue fell 4% last quarter. The company expects a high-single-digit decline this year. The stock is down about 79% from its 2021 peak.

That kind of drop can be permanent. Selling is often the right move.

2. The market pays less

Profits stay the same. Investors just pay a lower price for them.

This happens when fear rises. Think rates, war or worries about a whole sector.

3. Someone has to sell

This is the one most people miss.

Picture a great house on a nice street. The owner gets a job abroad and must sell by Friday.

The house did not get worse. The seller ran out of time.

Big funds can end up in that position. When they must sell, they sell what they can, not what they want.

Who was selling this month?

Not regular investors.

• Hedge funds cut tech at the fastest pace in Goldman Sachs records going back to 2016. That was the five days to Oct 15.

• By late October, they were shorting stocks five times faster than buying.

Retail investors went the other way. One market maker saw $9.1 billion of net retail buying in a single week.

Big money was selling hard. Small money was buying the dip.

So why were the pros selling? It comes down to three steps.

Step 1: Borrowed money

Most big funds do not just invest their own money. They borrow more, then buy more.

Subscribe now

Then comes the margin call

The lender holds the fund’s stocks as a guarantee.

When those stocks fall, the lender calls. “Add cash, or sell.” That is a margin call.

If the fund has no spare cash, it sells. Fast. At almost any price.

That is a forced seller. The selling has nothing to do with the companies.

This already happened in July

Situational Awareness was one of the hottest AI hedge funds of 2026.

It grew to about $45 billion in assets. Reports put its leverage as high as 400%.

Then AI chip stocks fell in July. Margin calls hit. The fund sold most of its stocks to Citadel at a discount.

It lost about 67% in one month.

When a fund that big sells, it pushes down the same stocks everyone else owns. Then the next fund feels the pain.

Step 2: Bigger swings, smaller bets

Volatility means how wildly a price swings. It says nothing about where the price ends up.

Two stocks can finish the year at the same price. One gets there calmly. The other rides a roller coaster.

Over the past year, Berkshire swung about 15% a year. The chip ETF swung about 40%. Nebius swung over 100%.

Many big funds budget for swings, not just dollars. A common rule:

Stock exposure = target swing ÷ market swing

Say a fund aims for 15% yearly swings:

• Market swings of 20% mean 75% in stocks.

• Market swings of 30% mean 50% in stocks.

• Market swings of 40% mean 37.5% in stocks.

Nobody had to turn negative on AI. The model saw bigger swings and sold.

Step 3: Short sellers

Short selling means borrowing a stock, selling it today, and buying it back later. The short seller wins if the price falls.

• Sell at $100. Buy back at $80. Keep $20, minus fees.

• Sell at $100. Price rises to $120. Lose $20.

The best case is a 100% gain. The worst case has no limit, because a price can keep rising.

Why shorts make drops worse

A short is still a sale. It adds more selling to an already falling market.

This month, hedge funds shorted about five times more than they bought.

Why shorts make bounces explosive

When prices turn up, short sellers face growing losses. So they rush to buy back.

That buying pushes prices higher. Which forces more buying. That is a short squeeze.

Remember April 9, 2025? After days of tariff panic, the Nasdaq jumped 12.2% in one day.

Moves like that rarely come from good news alone. They come from crowded bets unwinding at once.

Putting it together: two loops

This is why markets can fall like a stone, then rise like a rocket. Same engine. Different gear.

Subscribe now

Why some stocks fell more than others

When a fund must cut risk, it sells the jumpiest stocks hardest. They use up most of its risk budget.

As of Oct 23, a typical day moved:

• CoreWeave about 11%.

• Nebius about 10%.

• Nvidia under 3%.

The ranking roughly matches the damage. Nvidia fell least. CoreWeave fell most.

Five AI stocks can be one bet

Owning five AI stocks feels diversified. Often it is not.

Over the past three months, Nebius and CoreWeave moved together with a correlation of 0.86. A score of 1.00 means they move in perfect lockstep.

So when funds dump a basket of AI cloud stocks, every name in it gets hit. Strong or weak.

Not all AI stocks carry the same risk

“AI” is not one trade.

• Chip leaders like Nvidia, Broadcom and TSMC sell the tools. Their customers are cash-rich giants.

• Big tech platforms pay for AI from huge profits. But even Meta’s free cash flow turned negative $2.1 billion last quarter.

• Debt-funded builders are the most fragile. CoreWeave’s bonds yield around 15%, a sign lenders are nervous. Oracle sits one notch above junk at S&P.

For the debt-heavy names, a bounce may not stick until rates fall or credit eases.

For cash-rich leaders, a forced-selling drop is more likely to pass.

Bounce or real recovery?

The classic mistake is calling the bottom on day one of a bounce.

Look at the last two weeks. The chip ETF rose 1.6% on Oct 19 and 1.9% on Oct 21. Then it fell 6.0% on Oct 22.

That was a bounce. Not a recovery.

Three simple checks:

Check 1: Is the trend repaired?

The 50-day average is a simple short-term trend line. The 200-day is the long-term one.

The Nasdaq is still 2.4% below its 50-day. It is still 3.9% above its 200-day.

So the long trend is intact. The short-term damage is not fixed yet.

Check 2: Are the fear gauges calming?

• VIX measures how much swinging traders expect in the S&P 500. It closed at 19.45. Its one-year middle is about 17.

• VXN is the same idea for the Nasdaq 100. It closed at 28.3. Its one-year middle is about 23.

Both are still above normal. Funds are likely still in “cut risk” mode.

Check 3: Did the fear story change?

The big worries right now:

• War-driven oil prices. Brent crude traded above $116 on Oct 23.

• Higher inflation, and a 10-year Treasury yield near 5%.

• Doubts that AI spending will pay off fast enough.

Until one of these eases, rallies can fade fast.

Right now, all three checks read “not yet.” That does not mean sell. It means be patient with new buys.

Subscribe now

Should you sell? Three questions

1. Who is selling?

If funds are selling because of borrowing and risk rules, the selling ends when those rules are met.

Funds have already cut their net bets on stocks to a one-year low. That leaves room to buy back later.

2. Did the business change?

Look at the company, not the chart.

Are sales still growing? Are customers still ordering? Is the debt manageable?

If yes, the drop is probably about the seller. If no, the drop may be telling you something.

3. Can you hold through it?

This one is personal.

• Do you have debt, or borrowed money in your account?

• Will you need this money soon?

• Can you watch a 30% drop without panic?

If the answer is no, cutting risk is a valid choice. Even if the stock recovers later.

Never copy someone else’s portfolio. You can copy their stocks. You cannot copy their cash needs, debts or nerves.

The honest risks

Forced-selling drops usually pass. Not always.

• Technical can turn real. If selling cuts off a company’s funding, the business can suffer. Debt-heavy names are most exposed.

• Funds can sell more. Tech is still a large part of hedge fund books.

• Dip buyers used borrowing too. Leveraged chip ETFs drew heavy inflows. Those buyers can become forced sellers next.

• The economy can get worse. Higher oil, inflation and rates make investors pay less for every dollar of profit.

• Bounces fool people. We just saw one fail in a single day.

The opportunity

Forced sellers do not care about price. They sell what they must, when they must.

That can hand patient investors great businesses at prices that have nothing to do with their value.

The key is knowing which drop you are looking at:

• Broken business? Respect it. Step aside.

• Forced seller? Watch the checks. Get ready.

• Debt-heavy and forced? Be extra picky. Wait for proof.

Next time the market crashes while the businesses look fine, ask one question first.

Who has to sell?

Subscribe now

If this helped

This post is free because a lot of new investors feel lost right now.

If it helped you, please share it with a friend who is confused by this market.

And subscribe so you get the next one.

Paid subscribers also get my full portfolio and every move I make, each week.

See you in the next one.

Not financial advice. I may own some of the stocks mentioned. Data can change quickly. Do your own research.

Coinbase closed around $178 on October 21, 2026. That’s roughly 60% below the July 2025 high of $444.

Bitcoin is near $84,500 — still well below its October 2025 peak near $126k. Coinbase has fallen harder than the underlying asset, which is typical for a leveraged exchange, but the business underneath has changed in important ways.

The One-Sentence Thesis

Coinbase is shifting from a pure trading toll booth into regulated infrastructure for crypto, stablecoins, and on-chain activity. It now earns a meaningful share of revenue even when trading volumes are soft. At ~$178 you are paying roughly 7.5x trailing sales for a company with a fortress balance sheet and growing non-trading engines.

Fundamentals & Numbers That Matter

Q2 2026 total revenue was $1.22 billion, down ~19% YoY.

• Transaction revenue: ~$599M

• Subscription & services: ~$555M (48% of net revenue)

• In Q4 2024 the S&S share was only 29%.

Coinbase also states that 88% of net revenue now comes from sources other than pure bitcoin spot trading (versus 45% in 2020). That is the structural change.

Trading market share hit a record 10.3% of global crypto volume in Q2. Institutional transaction revenue grew while retail was softer. Twelve different products each generate over $100 million in annualized revenue. Coinbase One crossed 1 million paid subscribers.

Balance sheet: Cash and cash equivalents $8.6 billion. Total debt in the mid-$6 billion range. Net cash positive. Adjusted EBITDA remained positive at $208 million even in a soft quarter. GAAP net loss was $359 million (third consecutive), driven largely by mark-to-market items.

Growth rates have been cyclical, as expected. Full-year 2025 revenue was roughly $7.2 billion. The multi-year story is the rising quality of revenue, not straight-line top-line growth every quarter.

How Coinbase got here

Coinbase is 14 years old. It survived two crypto winters, an SEC lawsuit and the FTX collapse.

Subscribe now

What Coinbase actually is

Most people still think “crypto app.”

That’s maybe a third of the story now. Coinbase runs eight connected businesses on one regulated platform.

• Retail app: buy, sell, stake, earn USDC rewards. 7.6M monthly transacting users in Q2.

• Coinbase One: the subscription. Over 1 million paid members, a record in Q2.

• Institutional & Prime: trading, financing and prime brokerage for funds and companies.

• Custody: stores more crypto than any other platform.

• Derivatives: Deribit plus perpetual futures. Record global share.

• USDC: co-founded with Circle. Coinbase gets roughly half the economics.

• Base: its own blockchain. The #1 chain for stablecoin volume.

• Everything Exchange: stocks, prediction markets, pre-IPO perpetuals, commodities.

The key point: these businesses feed each other.

How big can this get?

Coinbase sells into some of the biggest markets in finance.

Crypto itself

• Total crypto market cap: $2.84T (CoinGecko, Oct 21).

• Assets on Coinbase: $246B at the end of Q2. That was 11.2% of all crypto.

Stablecoins

• Total stablecoin supply: $304B (DefiLlama, Oct 21).

• USDC is $72.6B of that. Coinbase holds about $20B of it in its products.

Derivatives

• Crypto derivatives trade roughly $110T a year.

• Coinbase did $4.2T over the last 12 months. That’s only about 4%.

• So there’s a lot of room to grow share.

Tokenized assets

• Tokenized real-world assets: $39B today (rwa.xyz, distributed value).

• McKinsey’s base case: about $2T by 2030. Its range is $1T to $4T. That’s an estimate, not a fact.

Payments

• Cross-border payments were about $130T in 2019 (McKinsey). Stablecoins want a piece of that.

You don’t need Coinbase to win all of this. A small slice of each market is enough.

Subscribe now

USDC: the quiet cash engine

This is the part most investors underrate.

USDC is a dollar stablecoin. Circle issues it. The reserves sit in Treasuries and cash. They earn interest.

Coinbase gets a big cut of that interest.

• Average USDC held in Coinbase products: $20B in Q2. It was $2B in Q2 2023.

• More than 30% of all USDC in circulation sits in Coinbase products.

• Over the past year, Coinbase captured about 50% of all USDC economics.

• Stablecoin revenue: $292M in Q2 alone.

• The Circle deal auto-renewed on the same terms. No renegotiation risk for now.

Here’s the flywheel:

• More USDC on Coinbase means more reserve income.

• Coinbase shares part of it as USDC rewards.

• Rewards attract more USDC. Users then trade, pay and lend with it.

One more thing. The Fed hiked rates in September. Higher rates mean more interest on USDC reserves. BofA pointed to higher stablecoin revenue in its October note.

Where institutions keep their crypto

When BlackRock or Fidelity launches a bitcoin ETF, someone has to hold the coins.

That someone is usually Coinbase.

• Coinbase holds 84.5% of US spot bitcoin ETF bitcoin as sole or main custodian (Jul 31, 2026).

• Assets on platform: $246B at the end of Q2.

• In April, the OCC gave conditional approval for a Coinbase national trust bank.

Custody is sticky. Funds don’t switch custodians every quarter. It’s a boring business. I like boring under a volatile one.

Subscribe now

Where the next growth comes from

Coinbase has more shots on goal than ever.

1. The Everything Exchange

One account for crypto, stocks, derivatives and prediction markets.

• 12 products already make $100M+ in annualized revenue each.

• Coinbase International Exchange lists 180+ markets for non-US institutions, including stock perpetuals.

2. Tokenized stocks

• Live on Base since August 24, 2026 for eligible non-US users.

• Real shares, held 1:1 by a regulated custodian.

• Apple and NVIDIA were among the first. They trade 24/7 and work as DeFi collateral.

3. Prediction markets

• Over $100M annualized revenue already.

• Contracts and revenue more than doubled in Q2.

4. Stablecoins and payments

• The GENIUS Act takes effect on January 18, 2027.

• The OCC published its final rule for it today, October 21.

• Clear rules should bring more banks, fintechs and merchants onto USDC.

5. Base and AI agents

• AI agents need a way to pay each other. Over 90% of those stablecoin payments already settle on Base.

• It’s early. But if agents become real customers, Base is the default rail.

6. International

• India: full launch in June 2026 with direct rupee deposits, spot and perpetuals.

• Europe: MiCA license since June 2025, valid across the EU.

• Coinbase says India is its only new country launch for the next 12 months. So it’s a focused bet.

7. Derivatives

• About 4% of a market I estimate at $110T a year. Every point of share is a big number.

Subscribe now

Main Risk-Regulations

The big one passed. The GENIUS Act was signed on July 18, 2025. It sets rules for dollar stablecoins like USDC.

The other big one stalled. The CLARITY Act, the market structure bill, failed a Senate vote 49 to 50 on Sep 15, 2026. It needed 60.

One sticking point is stablecoin yield. Banks don’t want stablecoins that pay rewards, pulling deposits away.

That fight matters for Coinbase. USDC rewards are a key part of the flywheel.

Valuation, Targets & Plan

Read more

Why it happened

1. AI debt fears. The trigger was one of the biggest AI cloud providers. It funds its data centers mostly with debt. Its bond yields jumped to about 15% to 16% this week. The cost to insure its debt passed 1,000 basis points. That’s distressed territory.

Coreweave fell about 25% in a week. The fear spread to every company that rents out AI compute. Then to chips, memory and power.

One analyst called this group “the subprime tranche of the AI trade.” Gartner said most of these AI clouds won’t stay independent by 2030.

2. Hot inflation. September CPI came out Wednesday, Oct 14. Prices rose 0.6% in a month. Inflation is now 3.7% a year. It was 3.4% in August. Gasoline drove more than half of it.

The 10-year Treasury yield spiked to 5.33% that day. Oil jumped again on Friday.

3. The Fed is hiking, not cutting. The Fed raised rates in September, its first hike since 2023. It meets again Oct 27 to 28. Fed Chair Kevin Warsh gave no signal on Friday. Futures still price some chance of another hike this month.

Higher rates hurt anyone who borrows to build. That’s the AI build-out right now.

4. AI revenue doubts.

• Oct 8: a report said OpenAI’s annualized revenue was about $50B. Investors expected $68B to $70B. Part of the gap is accounting. The market didn’t wait for details.

• Oct 15: Reuters reported that Anthropic’s investors are weighing a lower valuation for its IPO. About $1.5 trillion, down from $2 trillion.

Extreme fear

• CNN’s Fear & Greed Index closed Friday at 18.

• A week ago it was 46. Neutral.

• 6 of its 7 signals are in extreme fear.

So what actually changed?

Two things are real:

• Money costs more. Inflation is up. The Fed is hiking. Debt-funded builders pay more.

• The market wants proof. AI revenue has to catch up with AI spending.

What hasn’t changed so far: the demand. The big tech companies haven’t cut their AI spending plans. We’ll find out in the next two weeks, when they report.

Portfolio Performance

Last 2 years: 153.8%

YTD: 72.6% vs. SPY 11.7%

This week in the account

• Sold 1 stock completely.

• Opened 2 new positions.

• Added to 3 existing positions.

Let’s dive into the portfolio changes.

Unlock Full Portfolio Access

• 📊 Full portfolio transparency

• 🎯 Real-time buy/sell updates

• 🔍 Stock research and investment insights

The price will increase soon. Join now to lock in $180/year for as long as your subscription remains active.

Read more

The market just went on sale.

Some of the best growth stocks fell 25% to 50% from their highs.

The trigger was a hot CPI print and fresh AI worries. The Fed meets Oct 27 to 28.

Here’s my view. Pullbacks like this create a lot of opportunities. Some stocks are really cheap now.

This could be a good time to buy, especially in AI. But I’m watching other sectors too.

My best buys came in moments like this. My biggest Nebius buy was at $24, in April 2025. The market was in a panic.

So here’s my shopping list. Five names.

#1: Nebius (NBIS)

Why it’s great

• AI cloud built for GPUs. Demand is still bigger than capacity.

• Founder Arkady Volozh built Yandex. He has done this before.

• Customers pay ahead. Nebius expects more than $9B of prepayments in 2026.

• AI cloud adjusted EBITDA margin: 50%. This is a real business.

The growth

• Revenue Q2 2026: $582.3M, +454% YoY. +46% vs Q1.

• ARR: $3.0B in June. It was $551M in September 2025.

• FY2026 revenue guide: $3.0B to $3.4B. Reaffirmed in August.

• ARR target for end of 2026: $7B to $9B.

• Adjusted EBITDA: $236M, vs -$21M a year ago.

The worries: huge capex, possible dilution, and the GPU depreciation debate.

The chart

• Price is below the 50-day average ($224). It’s still above the 200-day ($173).

• RSI is 35, based on the Oct 14 close. Close to oversold. The July low came at a similar reading.

My levels

• Watch zone: $160 to $175. July 17 low ($164) and the 200-day average.

Next catalyst

• Q3 2026 earnings: November. Date not announced yet.

• The number I care about: progress toward $7B to $9B ARR.

The other 4 on my list

• #2: Revenue up 379%. Trading at about 5x forward earnings.

• #3: My pick for humanoid robots and physical AI. Down 51% from its high.

• #4: My safe pick if the crypto bull cycle continues.

• #5: AI networking. Revenue up 115%. Down 44% from its high.

Read more

I get asked the same question a lot.

“How do you find these before they run?”

There’s no magic screener. Just a few stories that taught me what to look for. Some of them hurt first.

I respect your time. So the short version is right below: the checklist I use.

But read the details too. They’re what make the checklist work.

The short version

My best multibaggers were small or hated stocks. The business kept getting better while the price got worse.

That gap is the opportunity. Your job is to spot it, size it right, and stay calm while it closes.

Subscribe now

Palantir: the business didn’t fall, only the stock did

I first bought Palantir in the COVID era. Somewhere around $20 to $30.

Then it went down. And down. By late 2022 it was trading near $7.

I was down a lot.

But here’s what I kept noticing. The business wasn’t going down. It kept growing.

Everyone talked about Palantir as a “government company”. That was the label. But the commercial side was growing too, and the market just wasn’t looking at it.

And Palantir had something nobody else really had back then:

• Deep access to governments.

• Software that companies could use as their own personal AI.

• AI tools that were genuinely unique at the time.

So at ~$7, I didn’t sell. I bought. That was my key buy.

What came next:

• Q4 2022: first profitable quarter. Revenue was up 18%.

• Commercial growth took off.

• Palantir joined the S&P 500.

At $7 the whole company was worth around $15B. Today it’s around $500B.

The lesson: Find what the company does uniquely.

Subscribe now

Selling Palantir at $100

In early 2025, Palantir hit $100. That was about 14x from my $7 buy.

I sold. And I rotated that money into Nebius.

Why? Palantir was now a $200B+ company that everyone wanted to own. Nebius was hated, much smaller, and growing much faster. The upside from there looked bigger.

Was I right? Partly.

Palantir kept going. It’s about $209 today. That’s roughly another 2x after I sold.

But Nebius did better. Even from around $40, it’s up more than 5x.

The lesson: Selling a winner isn’t a mistake by itself. Accept that the stock you sold might keep running.

Subscribe now

Nebius: people saw “Russian”, I saw the founder

So where did the Palantir money go? I found Nebius around $40. I wrote about it on X many, many times.

The pushback was always the same:

“It’s Yandex.” “It’s Russian.” “It’s not profitable.”

But look at what was actually there.

The company was run by Arkady Volozh. He’s the founder who built Yandex, the “Google of Russia”. Now he was building an AI cloud in Europe and the US. It had billions in cash from the split. And the growth targets were huge.

In February 2025, Nebius gave a big target. Run-rate revenue was $90M in December 2024. They guided to $750M to $1B by the end of 2025. That’s roughly 8x to 11x in a year.

The market didn’t care. It kept selling.

I had bought higher. I watched it drop anyway.

The stock fell to $24. Then to $20 in early April 2025, when the whole market was falling apart. That’s where I made my biggest buy.

At $24 the company was worth around $5.7B. About $2.4B of that was cash.

The stock peaked at $287 in June 2026. A lot of my early followers on X made serious money on this one.

The lessons: Management matters a lot. When a company is growing this fast and the market keeps selling, it’s worth asking who’s wrong.

Rocket Lab:

At ~$9 in September 2024, Rocket Lab was a 2021 SPAC that most people had given up on. It had peaked near $21 and bottomed near $3.50 in April 2024.

People saw a small-rocket company. But most of the backlog was space systems. Satellites, parts, defense work.

• Q3 2024: revenue up 55%. Backlog over $1B.

• It had won a $515M satellite contract from the Space Development Agency.

• Neutron, the bigger rocket, was the upside on top.

By Q2 2026, revenue was up 62% and backlog hit $2.36B. The stock peaked at $150 in May 2026.

Today it’s around $68. That’s more than 50% off the high. Great company, still a painful stock at times.

Subscribe now

Sivers

This one is different.

Early this year, the photonics trend was starting. AI data centers need lasers to move data with light. Sivers makes those lasers.

It had a very small market cap. Roughly $250M. Big accounts on X were starting to push it.

I looked into it. Low market cap. Good story. A LiDAR customer upgrade in March 2026 and new AI optics partnerships. The hype was clearly building.

So I jumped in early. I told my followers on X I bought at 8 SEK.

Then it went 20. Then 30. Then 50. It closed near 100 SEK in June.

At some point the move was just too extended. Price had run way ahead of the actual revenue. So I sold at 80 SEK and told my followers and subscribers I was out.

That’s a 10x in a few months.

What happened after matters. The company raised money twice. It restated its financials. The stock fell to about 22 SEK by September. It’s around 30 now. I bought it again after the drop and sold it between 40s and the 30s.

I want to be honest: this is not normal. It’s not something I’d do often. But it helped my returns a lot this year.

The lesson: hype trades can work. But know it’s hype. Keep the size small. Sell into the extension, not after it breaks.

Subscribe now

How I think about risk

Every one of these stocks dropped 40% or more at some point after the entry price. Look at this before you buy your next small cap.

And the pattern doesn’t always work. Plenty of stocks look hated, cheap and “misunderstood” and just keep falling. The 2021 SPAC crowd was full of them. Big story, no revenue growth, endless share sales.

So here’s how to handle it.

Start small 1% to 2%. Small enough that a 50% drop doesn’t wreck your sleep.

One last thing

None of these felt obvious at the time.

Palantir was “just government”. Nebius was “Russian”. Rocket Lab was a “failed SPAC”. Sivers was a tiny Swedish stock most people had never heard of.

The labels were wrong. The businesses were getting better. That’s the whole game.

The same process has worked on smaller wins too. AMD from $100 was one of them.

If you want to follow along, paid subscribers get the names I own right now.👇

The market is strong. But risks remain.

My portfolio is up 116.1% YTD, compared with 13.9% for the S&P 500. Over the last two years, it has gained 222.5%.

The 10-year Treasury yield remains elevated at 5.22–5.24%, creating headwinds for speculative growth stocks, small caps, and companies that depend on external financing.

OpenAI’s revenue story needs context

OpenAI reportedly told investors its annualized revenue was around $50 billion at the end of September. Earlier figures closer to $70 billion included gross partner sales.

The company reportedly still expects to reach around $70 billion by year-end, driven by enterprise growth.

The AI spending opportunity remains huge, but understanding the numbers matters.

Inside this week’s update

• Portfolio moves: What I bought, sold, and why.

• New opportunities: Stocks I’m watching and positions I’ve increased.

• Full portfolio breakdown: Every holding, position weight, and my updated risk map.

A personal thank you ❤️

Your messages and support mean more to me than you know. Knowing my research is helping readers on their investing journey motivates me to keep digging deeper and bringing more value to this community.

Thank you for being part of the journey. 🚀

Portfolio Performance

Last 2 years: +222.5%

YTD: +116.1% vs. SPY +13.9%

Let's Dive Into This Week's Moves…

Read more

This is one of the questions I get most often

Nebius, CoreWeave or IREN? Which has the best setup?

All three are benefiting from one powerful trend: surging demand for GPU infrastructure.

But their balance sheets, customers, growth profiles, and financing needs differ widely.

In this deep dive, I put Nebius, CoreWeave, and IREN side by side to see which offers the strongest risk/reward and which has the most to prove.

Inside the full analysis:

• Which company has the strongest balance sheet

• Who is most exposed to higher rates and near-term maturities

• How their debt, cash and capex compare

• Which company has the best backlog and customer base

• Dilution trends and where refinancing risk can show up

• How the three compare on valuation

  Subscribe now

  The subscription price will increase soon. Existing paid subscribers will keep their current rate. Join now to lock-in.

  Let’s dive in…

Read more

Another big week.

The hike narrative broke.

PCE came in cooler. Payrolls missed badly. Unemployment rose. October hike odds collapsed from ~70% to ~20%. But the 10-year didn’t cooperate.

It briefly dipped below 5.17% before closing around 5.28%.

That divergence matters.

Inside the portfolio

I sold a position completely.

My recent entries are now up 40–60% in just a few months.

In this update, I cover:

• What I changed this week

• My strategy behind the recent entries now up 40–60%

• Full portfolio

• My brand-new portfolio risk map

• My price targets and next entry ideas

• What I’m watching next

Portfolio Performance

Last 2 years: +249.5%

YTD: +135% vs. SPY +12.8%

No options. No leverage.

Let’s dive in…

Read more

A year ago, Micron generated $11.3 billion in fiscal Q4 revenue.

Last night, it generated $54.2 billion.

Adjusted EPS came in at $33.42, comfortably above expectations of roughly $31.60–$31.70. Revenue also cleared a Street estimate that was already above $51 billion.

And gross margin? 87%. A year ago, it was 46%.

Yet the stock barely moved after hours.

At first glance, that looks like a classic “priced in” reaction.

I think there is a more interesting story underneath it.

The real takeaway

The headline numbers were huge, but the most important part of the call was not the quarter.

It was what management said about 2027 and 2028.

Micron expects memory supply and demand to be significantly tighter in fiscal 2027 and 2028 than they were in 2026.

Fiscal 2026 just produced an 87% gross margin quarter, and management is saying the supply-demand environment could get even tighter from here.

Subscribe now

The numbers were insane

Fiscal Q4 revenue reached $54.2 billion, up 379% year over year.

DRAM generated $39.8 billion, representing 73% of sales. Revenue was up 343%, while bits increased only in the mid-single digits. Pricing did most of the work.

NAND revenue reached $14.1 billion, up 526%.

Data-center SSD revenue alone was nearly $10 billion, more than 10x the year-ago quarter.

For the full fiscal year, Micron generated $133.2 billion in revenue, up 256%.

Data-center revenue quadrupled.

Micron generated roughly $44 billion of operating cash flow in Q4 and $90 billion for the full year. It ended the year with $68.3 billion of net cash.

And then there is the number I care about most:

$12.7 billion of customer deposits on the balance sheet.

Most of that cash came in during the quarter.

Customers are not just talking about future demand. They are putting money down to secure supply.

Customers are locking up memory through 2030

This is where the earnings call gets really interesting.

Micron now has 26 strategic customer agreements.

Some contracts now extend into 2031.

Customer financial commitments increased from $22 billion to $32 billion in just one quarter.

Remaining performance obligations increased from roughly $100 billion to $150 billion.

And around 75% of fiscal 2027 shipments are already committed.

Most of Micron’s 2027 HBM capacity is also already under agreement, at prices significantly higher than a year ago.

The market is used to thinking about memory as a commodity business where supply eventually catches up and pricing collapses.

But this time, customers are actively trying to secure supply years in advance.

HBM is still the bottleneck

HBM is becoming increasingly critical because AI accelerators need enormous amounts of high-bandwidth memory.

And Micron is basically telling us that supply is still nowhere near enough.

Micron is also working with Nvidia on NVHBM, which management described as the industry’s first custom HBM4E solution for next-generation GPUs and NVLink Fusion.

The relationship between memory companies and AI chip companies is becoming more strategic.

Subscribe now

Supply is not coming overnight

This is where I think the market may be underestimating the setup.

Yes, Micron and the rest of the industry are spending aggressively on new capacity.

But new fabs do not suddenly create supply the moment construction starts.

There is a long process involving cleanrooms, equipment installation, wafer qualification, yield improvement and eventually meaningful production.

Micron expects first wafer output from its Idaho fab in mid-2027.

ID2 is expected in late 2028.

The first New York fab is a 2030 story.

Japan DRAM expansion is expected in late 2028.

Singapore NAND expansion is expected in the second half of 2028.

Singapore HBM packaging is expected in early 2027.

In other words, much of the capacity investors are pointing to as the end of the shortage is actually a 2028–2030 story.

That is why management saying it still has no clear line of sight to a balanced market matters so much.

Revenue is still compounding

Micron did not guide the year. It guided the next quarter at $61.5 billion, and said revenue rises every quarter. Four flat quarters would already be $246 billion. The Street’s $251 billion just assumes that path holds.

Price did the work, not bits. The orange bars are the tell. Operating income went from $10 billion to $99 billion in a year. The 2027 estimate takes it past $200 billion.

The AI buildout is showing up in the numbers

One argument I hear around AI stocks is that hyperscaler spending is getting ahead of real demand.

Micron gives us a useful way to test that.

The money is showing up downstream.

Customers are putting cash on the table to secure memory.

Server unit growth is still expected to be in the high teens in both calendar 2026 and 2027.

The read-through for other AI stocks

This is why I think the Micron report matters beyond Micron.

For Nvidia and AMD, the message is that accelerator demand continues to be constrained by memory supply.

For Broadcom and custom silicon, the same logic applies. AI accelerators still need HBM and large amounts of surrounding DRAM.

And for the broader infrastructure trade, it is an interesting sequencing signal.

Micron’s management described memory as the chief constraint compared with logic and data-center power.

The AI buildout still looks capacity-constrained rather than demand-constrained.

Subscribe now

There is another interesting detail: Margins

HBM has historically carried a significant premium to conventional DRAM.

But Micron is now saying that higher pricing in conventional DRAM is narrowing that gap.

That is another sign of how tight the overall market has become.

It is not just HBM that is scarce.

The entire memory environment is benefiting.

And the company’s other businesses are holding up as well.

In Q4:

• Core Data Center: $18.0B revenue, 90% gross margin

• Cloud Memory: $16.3B, 83% gross margin

• Mobile and Client: $13.1B, 90% gross margin

• Automotive and Embedded: $6.8B, 84% gross margin

What to Do With This

Three simple takeaways.

1. The memory bottleneck did not disappear.

2. Don’t confuse a flat after-hours reaction with a weak report. The real news is 2027–2028.

3. If you want memory exposure, scale in.

The biggest takeaway? Customers are already locking up memory through 2030.

That is great news for our other AI positions too. If memory stays constrained, demand for GPUs, networking, power and the rest of the AI stack stays strong.

The AI trade doesn’t look over.

Want to See Exactly What I Own?

I hope you enjoyed this Micron earnings review.

I share my full portfolio, position sizing, average costs, buys, trims, and exits — with complete transparency.

You can see exactly what I own, how much I own, and where I’m putting my money below.

I skipped last week’s write-up, so this update covers the last two weeks.

The market had plenty to worry about: the 10-year Treasury touched 5.23%, the highest level since 2007, while the 30-year reached 5.53%. Brent stayed above $100 as Middle East headlines kept the oil risk premium elevated.

And stocks still moved higher.

The S&P 500 finished around 7,743, up 1.2% on the week. Nasdaq gained about 2%.

Microsoft, Meta’s Muse developments, and another wave of AI infrastructure spending were enough to offset the pressure from higher yields.

That is the key question I’m focused on:

Can AI infrastructure spending keep accelerating while the cost of capital keeps rising?

Goldman Sachs’ baseline is roughly $7.6T of AI capex from 2026–2031, rising from about $765B in 2026 to $1.64T in 2031.

That is the cycle my portfolio is positioned for.

This week in the portfolio

• Started 3 new positions

• Sold 1 position completely

• Trimmed 1 position

I will break down all decisions and where the capital may go next below.

Portfolio Performance

Last 2 years: 247.1%

YTD: 131.7% vs. SPY 12.9%

No options. No leverage.

High-growth investing comes with volatility. Volatility is also where the next entries usually appear.

Access My Portfolio

So What Did I Actually Buy?

Paid members get:

• 💼 Full portfolio transparency

• 📈 Real-time buy & sell signals

• 🎯 My highest-conviction stock ideas

• 🧠 Detailed investment theses and updates

It’s currently $179/year — less than $0.50 a day.

The price will increase soon. Join now and lock in the current rate for as long as you remain subscribed.

Read more

Cerebras opened at $350 on day one in May and reached $386 before falling back to around $212. That is roughly a 45% drawdown from the high.

At the same time, the company has:

• $25.4 billion in remaining performance obligations

• $880–$890 million of 2026 core revenue guidance

• A plan to more than triple revenue by 2027

• A major 750 MW OpenAI agreement

So why is the stock down so much?

The market appears to be weighing three things:

Valuation, customer concentration, and the huge amount of stock becoming available after the IPO lockup.

The One-Sentence Thesis

Cerebras is turning an entire wafer into an inference engine and selling that technology through both hardware and cloud services.

OpenAI has already contracted 750 MW of Cerebras inference capacity.

At around $212 per share, the valuation is:

• 47x 2026 sales

• ~16x 2027 sales, assuming the company more than triples revenue

That is an expensive valuation.

But it is also a business with a very large contracted backlog behind it.

The question is whether the market has already priced in that future growth.

What Does Cerebras Actually Do?

1. Hardware / Systems

Cerebras sells its CS-3 systems, built around its Wafer-Scale Engine.

Instead of connecting many smaller chips together, Cerebras puts an enormous amount of computing power onto a single wafer.

The latest architecture includes:

• 4 trillion transistors

• 900,000 AI cores

• 44 GB of on-chip SRAM

• 21 PB/s of memory bandwidth

The goal is simple: Move data faster and reduce the communication bottlenecks found in large GPU systems.

The downside is that hardware revenue can be very lumpy.

2. Cloud / Inference

This is the part of the business I find particularly interesting.

Instead of buying a Cerebras system, customers can simply rent inference capacity and pay for usage.

Q2 core cloud revenue reached $127.7 million — up 287% YoY.

Cloud now represents approximately 61% of the company’s core revenue mix.

That is a major shift. Cerebras is gradually moving from:

“Sell the machine.” to: “Sell the tokens.”

Subscribe now

The Financial Picture

Management guided to: Q3 core revenue: $214–$216 million

2026 core revenue: $880–$890 million

Core gross margin: 41–43%

Core operating margin: –19% to –17%

So this is not yet a profitable business.

But the company is growing rapidly while reducing operating losses relative to revenue.

It also has a significant balance sheet.

Following the IPO, Cerebras had approximately:

• $8.6 billion in cash, restricted cash and investments

• $850 million of unused revolver capacity

That gives the company substantial financial resources to expand infrastructure.

Why Wafer-Scale Computing Matters

The easiest way to understand Cerebras is to compare the architectures.

Nvidia: General-purpose AI infrastructure with enormous CUDA software ecosystem.

Groq: Specialized inference hardware using many smaller chips.

Cerebras: One enormous wafer-scale processor designed to minimize chip-to-chip communication.

The pitch is primarily about inference speed.

Published benchmarks for Llama-class models have shown Cerebras reaching roughly 2,100–2,500 tokens per second in some configurations.

Groq and GPU systems can be significantly lower depending on the model, hardware and benchmark.

Cerebras also claims up to 15x faster performance than leading GPU systems on some open models.

Cerebras Doesn’t Need to Replace Nvidia

This is an important distinction. I don’t think the Cerebras thesis requires Nvidia to lose.

Hyperscalers and AI companies will likely operate mixed infrastructure fleets.

Nvidia remains extremely important for:

• Training

• General-purpose workloads

• Software compatibility

• Large-scale AI infrastructure

Cerebras can instead capture a specific part of the market: high-speed inference where latency matters.

If a customer will pay more for extremely fast responses, Cerebras has a differentiated product.

So Why Did the Stock Fall?

1. Customer Concentration

OpenAI is becoming extremely important to Cerebras.

Before OpenAI, the company already had significant concentration through G42 and MBZUAI. Those customers represented roughly 86% of FY2025 revenue.

Management expects OpenAI to remain a major customer in 2027, but its percentage of revenue should decline if AWS and other cloud customers scale.

We still need to see that diversification show up in actual revenue.

Subscribe now

AWS

Cerebras is targeting Amazon Bedrock general availability in Q1 2027.

This matters because AWS could provide Cerebras with another major distribution channel. If AWS adoption becomes meaningful, it could help reduce the company’s dependence on OpenAI.

Lockups

There is another reason the stock has been under pressure: Supply!

A large amount of previously locked-up stock is becoming available.

• ~19.4 million shares on September 30

• ~19.4 million shares on October 14

• ~19.4 million shares on October 28

• Final release around the Q3 results / November 9

Lockups can create significant selling pressure even when the underlying business continues to perform.

Cloud Business Could Change the Economics

This may be the most interesting part of the story. Cerebras can sell a $2–3 million system once. Or it can potentially use that infrastructure to generate years of inference revenue.

The cloud business is:

• Growing much faster

• Usage-based

• More recurring, potentially higher quality revenue

But there is an important caveat. Cerebras currently has low-40% gross margins.

This is still fundamentally a hardware and infrastructure business that is learning how to monetize its infrastructure through recurring usage.

Risks

🔴 Customer Concentration

OpenAI is a major growth driver. Any delay or change in deployment could materially impact expectations.

🔴 Complex Economics

The OpenAI deal includes warrants and a working-capital loan repaid through service credits, making the economics more complex.

🔴 Lockup Supply

September and October unlocks could create near-term selling pressure.

🔴 Profitability

Core operating margin is still guided at –19% to –17%, while scaling requires significant capital.

🔴 Power & Capacity

Cerebras has ~600 MW contracted through 2027, while OpenAI alone represents 750 MW—making execution and capacity expansion critical.

Valuation, Targets & My Plan

Upgrade to Paid

My portfolio is currently up 130% YTD.

As a paid member, you get:

• Full access to my portfolio

• Real-time buy & sell alerts

• Exclusive deep-dive research

• Access to a great community of investors

The current price is $179/year — just $0.49 per day.

The price will increase. Join now and lock in $179/year forever as long as you stay subscribed.

Upgrade

Read more

The stock was $495 in June is sitting near $340. That’s a 31% drop on a company that just told you AI semiconductor revenue is going to $58 billion this year, $115 billion next year, and $230 billion in 2028.

This article is not about “Broadcom is an AI stock.” Everyone already knows that.

The interesting question is whether $230 billion of AI silicon in two years is already in the price or not.

The one-sentence thesis

Broadcom is the only company that co-designs the custom accelerators and sells the Ethernet fabric that glues those accelerators into clusters. They have a $35 billion high-margin software annuity sitting underneath

At ~$340 you are paying roughly 11–12x the 2028 earnings.

What the business actually is (not the press-release version)

Two engines. Different economics.

Semiconductor solutions is now the growth engine. In Q3 it did $20.8 billion, +127% YoY. Inside that, AI semiconductors were $16.7 billion +221% YoY

That AI line is custom XPUs (Broadcom’s term for customer-specific accelerators) plus the networking silicon that moves tokens between them.

The six named XPU customers: Google (TPU family, Ironwood v7 in volume, v8i starting), Anthropic (same TPU family, ramping hard), OpenAI (Jalapeño, first-gen inference ASIC), Meta (MTIA, three generations through FY28), plus ByteDance and Fujitsu.

Networking is not a small part. In earlier quarters it was ~40% of AI semiconductor revenue. Tomahawk-class Ethernet and optics scale with every extra gigawatt of compute. You cannot fill a 5 GW cluster without the switch silicon.

Hock Tan has spent fifteen years buying high-ROIC franchises and stripping them for cash. VMware was the biggest of those bets. The “customers will flee” narrative after the 2023 deal never fully showed up in the P&L. What showed up was pricing power on VCF and a private-AI-cloud pitch that now sits next to the XPU story.

Custom silicon plus the fabric that glues the racks together, with a high-margin software annuity underneath. That is the business.

Google TPU family. Broadcom has been inside this partnership since 2014. That is not a design win you have every year.

The number that should have moved the stock

Q4 guide: ~$34.8B revenue (+93%)

• Anthropic: 1 GW Ironwood in FY26 → 5 GW TPU v8i in FY27 → incremental 10 GW in FY28

• OpenAI: 1.3 GW Jalapeño in FY27 → >5 GW across Jalapeño + next-gen in FY28

• Meta: 3 GW across three MTIA generations through FY28

• Google continues to grow on top of that

The Street wanted a beat-and-raise on the next ninety days. Broadcom gave them a two-year doubling of an already enormous AI line and a path to ~$30 of non-GAAP EPS in FY28. The stock went down.

Why custom ASICs?

Nvidia sells a general-purpose training/inference platform with CUDA gravity. Broadcom sells a cost and power solution for the handful of labs that already know their model architecture and want tokens per dollar, not flexibility.

Management has said OpenAI’s Jalapeño can deliver inference at less than half the cost of a comparable GPU stack. That is the whole pitch to a lab that is burning cash on inference at scale.

Two things follow:

1. This is not a winner-take-all versus Nvidia. Hyperscalers and labs will run mixed fleets. Nvidia still owns training and the long tail. Broadcom owns the customized inference and networking slice for the six customers who can afford to co-design.

2. The switching cost is brutal once a generation is taped out and the cluster is live. You do not rip a 5 GW Anthropic deployment because Marvell offered a 4% cheaper.

   

   Subscribe now

   The software line nobody models correctly

   VMware was supposed to be dead. Instead:

   • Software is still ~30% of revenue in a quarter where AI exploded

   • Q3 software +29%, ARR +15%

   • Segment operating margins remain in a different universe from semiconductors

   If AI silicon is the call option, software is the bond portfolio that pays you while you wait. It also gives a cash-flow machine to fund buybacks when the multiple compresses.

   Dividend is $0.65/quarter ($2.60 annualized), 16th consecutive annual increase, yield ~0.75% at $340. I am not interested in dividends.

The risks that actually matter (not the generic ones)

• Customer concentration. Six XPU names. Anthropic is about to be #1 and is still private. If Anthropic’s IPO or fundraising stalls, or if one hyperscaler pauses a generation, the FY28 $230B number has a hole in it. This is the real risk.

• Mix and margins. XPUs carry lower gross margin than the old Broadcom mix.

• Supply. HBM and advanced packaging are the bottleneck for the entire industry.

• Competition and insourcing. Marvell is the only other serious custom-ASIC house. MediaTek took a slice of a Google training part.

• VMware / EU. Licensing changes are being looked at in Europe. Not existential.

• AI capex pause. If the labs decide tokens-per-dollar no longer justify 5–10 GW orders, the whole custom-ASIC complex reprices.

Disclaimer: This content is provided for educational and informational purposes only and should not be considered financial, investment, tax, or legal advice. Nothing in this article constitutes a recommendation or solicitation to buy, sell, or hold any security. Investing involves risk, including the potential loss of capital. Always conduct your own research and consider your personal financial circumstances before making any investment decisions. Not Financial Advice.

Paid members get the full portfolio, every position, entry price, position size, weekly changes, and my latest buy/sell decisions.

Hundreds of members are already in the paid community.

Upgrade

For less than the price of a coffee each week, you get access to the portfolio behind the performance.

Read more

It was a pretty messy week in the market.

Inflation came in hot, rate expectations moved quickly and suddenly the Fed meeting became the story again. By Friday, markets were pricing roughly an 80% chance of a September hike.

I expect the Fed to hold the current range next week. When positioning becomes this one-sided, the surprise can come from the other side. I am less interested in the prediction-market contract itself and more interested in what it would mean for growth multiples.

Spending on compute, power, memory, optics and launch infrastructure continues. Some of the higher-beta names got messy this week, but the infrastructure buildout did not.

This week in the portfolio

• Sold 2 positions completely

• Trimmed 1 position

I will break down all decisions and where the capital may go next below.

Portfolio Performance

Last 2 years: +226.1%

Last 1 year: +135.2%

YTD: +113.7% vs. SPY +11.8%

No options. No leverage.

Want the full portfolio update?

Paid members get the entire book: every position, average cost, weight, weekly changes, and the reasoning behind what I hold, add, trim, or sell.

Get Full Portfolio Access

It is currently $179/year — $0.49 a day.

The price will go up. For less than a coffee a week, you see the portfolio behind the performance.

Join now and you lock in the current rate for as long as you stay subscribed.

Read more

That is the question I keep coming back to.

Most investors focus on the most visible part of AI. GPUs. Chatbots. Big tech.

But the biggest opportunity may sit somewhere else.

The scarce part of the stack.

The company that owns a bottleneck has leverage. When demand keeps rising and supply cannot catch up, pricing power follows.

That is the trade I have been looking for.

First, understand the stack

Every AI data-center dollar has to move through this stack.

If one critical layer is missing, the whole system slows down.

That is why I do not think “AI stocks” is a useful portfolio category.

A 9/10 scarcity business and a 3/10 story stock can both be called AI.

They are not remotely the same investment.

The rent equation

Before looking at individual stocks, I use four questions.

Rent Score = Pricing Power × Scarcity Duration × Incremental Margin × Multiple Room

In plain English:

1. Pricing power
Can the company raise prices and still sell everything it makes?

2. Duration
Will the bottleneck last two years or eight?

3. Marginal profit
Does the next dollar of revenue produce huge incremental profits?

4. Multiple room
Is the stock already priced for perfection?

That last one matters more than ever.

Nvidia still scores extremely well on quality, pricing power and margins.

But it is no longer an automatic winner on valuation.

That is the debate.

Layer 1 — Power

The input every other layer needs

This was one of the most overlooked parts of the AI trade.

It is not anymore.

U.S. data-center electricity demand is expected to rise dramatically this decade. Hyperscalers are already signing major nuclear and power agreements. Turbine and grid equipment lead times are measured in years.

That changes the investment case.

This is not just an energy story.

It is a supply constraint.

The names I watch

CEG — Constellation Energy

The largest U.S. nuclear fleet gives CEG a unique position.

Long-term power contracts with investment-grade customers create unusually visible cash flows.

The downside is simple.

You are still buying a utility-like business. The upside may come more slowly than with a semiconductor.

TLN — Talen Energy

A more aggressive way to play the same theme.

Nuclear plus gas. Significant exposure to data-center demand.

The catch is concentration.

GEV — GE Vernova

This may be one of the most interesting picks in the entire power chain.

Turbines. Grid equipment. Electrification.

The key point is simple: data centers cannot come online without physical infrastructure.

GEV is selling equipment years into the future.

VST — Vistra

A merchant power angle on the same problem.

When electricity gets tight, power prices matter.

That creates torque.

Why I like this layer

Duration.

You cannot build a reactor or a major turbine overnight.

AI models can improve in months.

Power infrastructure takes years.

That mismatch creates scarcity.

Scarcity score: 9/10

Layer 2 — Packaging

The quiet bottleneck

You can have the GPU. You can have the HBM.

But if you cannot package them together, you still cannot ship the system.

That is why advanced packaging is so important.

TSMC remains the dominant player in leading-edge 2.5D packaging, including CoWoS.

Capacity is expanding.

Demand is expanding even faster.

That is the key.

This is one of the least exciting parts of the AI trade.

It may also be one of the most important.

Customers cannot simply switch suppliers overnight.

That creates pricing power.

A next-generation AI accelerator is not just a chip.

It needs memory. Packaging. Networking.

Everything has to arrive at the same time.

Miss one piece and the system does not ship.

TSMC sits directly in the middle of that bottleneck.

Amkor and other OSAT companies matter too.

Scarcity score: 9/10

Subscribe now

Layer 3 — Memory

The new oil

Every accelerator eventually becomes a memory problem.

HBM demand has exploded.

HBM3E was effectively sold forward.

HBM4 is now moving into the next generation of AI systems.

SK hynix, Micron, Samsung and SanDisk are the leaders.

For U.S. investors, Micron is the obvious listed vehicle.

The most important thing here is that HBM is not simply another DRAM cycle.

It is becoming strategically important to the AI system itself.

Two things stand out to me.

First, the amount of revenue generated per gigawatt of AI infrastructure keeps rising.

Second, memory and compute now represent a much larger share of data-center costs than they did earlier in the decade.

There are only a few qualified HBM suppliers. The products are tied directly to leading AI accelerators.

And customers cannot easily replace a supplier once a product is qualified.

Scarcity score: 8/10

Layer 4 — Compute

Nvidia is still the king

The company continues to post numbers that would have sounded absurd a few years ago.

Data-center revenue is enormous. Guidance remains strong. The Rubin roadmap is already becoming a major part of the next cycle.

CUDA remains one of the strongest software moats in hardware.

I am not betting against that.

But I do think the nature of the opportunity has changed.

Nvidia is no longer a cheap way to express the AI trade.

The easy multiple expansion is probably behind us.

From here, earnings growth has to do the heavy lifting.

And there are two risks I take seriously.

1. Custom silicon

Broadcom is becoming increasingly important in custom AI chips.

Google has TPUs. Amazon has Trainium. Meta is building more of its own silicon.

2. Nvidia depends on the rest of the stack

This is often overlooked.

Nvidia cannot ship GPUs without memory.

It cannot ship systems without advanced packaging.

It cannot power a data center.

CUDA is still the best software moat in AI hardware.

I just do not think it is the only moat anymore.

Scarcity score: 7/10

Subscribe now

Layer 5 — Networking

A giant GPU cluster without a strong network is useless.

The more GPUs you connect, the more important the fabric becomes.

Broadcom is playing here through custom silicon and networking.

Arista is one of the cleanest public-market ways to own AI Ethernet infrastructure.

The numbers are impressive. Growth remains strong. Margins are extremely high.

And the customer base is expanding rapidly.

This is what I want to see in an AI infrastructure stock:

High incremental margins.

Strong demand and a product that becomes more important as the system gets larger.

Scarcity score: 7/10

Layer 6 — Cloud

This is where things get interesting.

Microsoft, Google, Amazon and Oracle are not simply buying AI infrastructure.

They are monetizing it. They turn electricity, GPUs and networking into rented compute.

That gives them something many smaller AI companies do not have: Balance sheets.

The spending numbers are enormous. Hyperscaler capex is now in the hundreds of billions.

Cloud growth remains strong.

AI demand is flowing through the entire ecosystem.

And then there is CoreWeave, Nebius Group, Iren, CIFR, WULF etc…

These companies are no longer simply small GPU landlords.

That makes it a real infrastructure business. But there is a risk.

All this spending has to earn a return.

At some point, investors will ask a very simple question:

Are AI revenues growing fast enough to justify the capex?

Google trading at a relatively modest forward multiple is a very different setup from buying a high-growth company at an extreme valuation.

Scarcity score: 6/10

Balance-sheet score: 9/10

Subscribe now

Layer 7 — Models

Huge strategic value. Limited public-market exposure.

This is where most of the excitement lives.

OpenAI. Anthropic. xAI. Gemini. Open-source models.

Alphabet and SpaceX are the cleanest public-market exposure to frontier models.

Microsoft gives you exposure to OpenAI alongside Azure.

But I do not want to build a portfolio around models I cannot value properly.

I would rather own the companies every model needs.

The infrastructure.

The compute.

The cloud.

The distribution.

Strategic importance: 10/10

Listed scarcity: 4/10

Layer 8 — Applications

Where I think the next big winners will emerge

Infrastructure dominated the first phase of the AI trade.

Now I want to see what happens further up the stack.

This is where the next 5–10x stocks could come from.

But there is a catch. You cannot pay an infrastructure multiple for an application company.

Palantir

Palantir is one of the most impressive growth stories in the market.

My first 10-bagger investment.

Revenue growth is enormous. U.S. commercial growth is even stronger.

The company is becoming increasingly important to large enterprises and governments.

The problem? Valuation.

At an extreme earnings and sales multiple, you need a lot to go right.

I like the company. I do not like the current price

Asymmetry means the upside is large relative to the downside.

AppLovin

APP is a completely different setup.

Margins are exceptional. Revenue growth remains strong.

And the market has become more cautious about the story.

That can create an interesting setup.

When financial performance stays strong while sentiment weakens, I start paying attention.

That is often where the biggest opportunities appear.

CrowdStrike and other vertical software companies belong here too.

But dispersion is huge.

Some will become dominant platforms.

Scarcity score: 5/10

Dispersion: extremely high

Subscribe now

The scorecard

Follow the dollar

Take $750 billion of annual hyperscaler capex as a rough example.

The exact split will change.

Now the interesting part.

The average investor owns the largest and most obvious slice: Compute.

But the smaller slices can control whether the compute actually gets shipped.

That is the opportunity.

Power. Memory. Packaging.

Those layers can become a constraint on the entire system.

And when a constraint becomes binding, its economic value increases.

The bottom line

The rent goes to the bottleneck that can say no.

In 2023, that bottleneck was GPUs.

In 2026, it is much more complicated.

Power matters. CoWoS matters. HBM matters. Networking matters.

Cloud balance sheets matter.

And eventually, the software that monetizes all of this will matter even more.

The goal is to figure out where the pricing power is still underappreciated.

Full portfolio

I own some of the names discussed in this article.

My current portfolio, including weights, cost basis and recent changes, is available in my latest weekly portfolio update.

I update it regularly so you can see how I am expressing these ideas with real position sizes, not just theory.

A hotter jobs report pushed rate-hike expectations higher. Inflation is still sticky. Then, Friday night, S&P Dow Jones Indices confirmed what the market had been anticipating:

Bloom Energy is joining the S&P 500.

That matters for the portfolio.

$BE: From $175 Buy Alert to $266

I sent the $BE buy alert to paid subscribers on July 29 at $175.

Just over a month later, Bloom reached $266.14 after Friday’s close. +52% in a very short period of time.

This week I made three portfolio moves:

• Added a new position

• Sold one position completely

• Increased one existing position

I’ll break down each move below, including my reasoning, entry prices and position sizing.

Portfolio Performance

Last 2 years: +231.3%

Last 1 year: +138.6%

YTD: +116.1% vs. SPY +12.7%

No options. No leverage.

🔒 Want the full portfolio?

Paid members get the full portfolio, every position, entry price, position size, weekly changes, and my latest buy/sell decisions.

Hundreds of members are already in the paid community.

It’s currently $179/year — $0.49 per day.

For less than the price of a coffee each week, you get access to the portfolio behind the performance.

Read more

That position became my third 10-bagger.

At the June high near $287, it was up more than 1,000%. Around $215 today, it’s still roughly 8.8x.

A big part of my high portfolio return came from this one position.

But when I bought it, nobody was excited about Nebius.

Quite the opposite.

“It’s Russian.”

“They just rent GPUs.”

“This is a sanctions leftover.”

Those were the arguments I kept hearing.

And honestly?

That was exactly what made me interested.

In October 2024, Nasdaq brought the former Yandex holding company back to the market under a new name and ticker.

The market saw the baggage.

I was looking at the infrastructure.

Nebius had something that was becoming increasingly scarce:

AI compute capacity.

Not another AI app.

Not another chatbot.

Power + GPUs + data centers + customers willing to pay for capacity.

What I saw that others were missing

I only needed three things to be true.

1. The market was looking at the wrong story.

Investors were still focused on the company’s Russian history.

I was focused on what the business was becoming.

2. Compute was becoming a scarce resource.

You can build an AI application relatively quickly.

You cannot create gigawatts of power and massive GPU infrastructure overnight.

3. Serious customers were willing to put real money behind it.

Then Nvidia invested in Nebius at roughly $21 per share in December 2024.

That got my attention.

When the company at the center of the GPU shortage invests in your infrastructure business, the “just a GPU renter” argument becomes a lot harder to make.

Then the numbers started catching up

The company stopped looking like a leftover.

It started looking like an infrastructure company positioned directly inside the AI buildout.

Q2 revenue reached $582 million, up 454%.

ARR reached $3 billion.

Contracted power was guided toward 5 GW.

And major customers started signing meaningful capacity deals.

Microsoft and Meta. About $40 billion in commitments.

Suddenly, the market wasn’t debating whether Nebius was a leftover from Yandex.

It was debating how big the AI infrastructure opportunity could become.

The important part

I didn’t buy Nebius because I knew it would reach $287.

I bought it because the market was pricing in a version of the company I didn’t believe.

At $24, the market was essentially saying:

This probably doesn’t work.

Today, at around $215, the market is saying something very different:

A lot of this probably works.

My filter for the next 10-bagger

This was my third 10-bagger.

And the first two didn’t come from consensus names either.

I’m looking for companies where:

1. The market still hates them for a lazy reason.

2. They control something genuinely scarce.

3. Customers are willing to write serious checks.

4. The company can fund the build without destroying shareholders.

5. The stock price still assumes failure.

That last one is the most important.

At $24, NBIS assumed failure.

At $215, it doesn’t.

So I’m looking for the next company where the market is making the same kind of mistake.

Not necessarily another Nebius.

But the same setup.

Want to see exactly what I’m buying, selling, and holding?

Join my paid Substack and get access to my real-time portfolio, every buy and sell, high-conviction stock ideas, in-depth research, valuation work and the reasoning behind my decisions.

I share my portfolio transparently so you can see exactly where I put my own money — not just hear me talk about stocks after they go up.

You’ll also get access to the private community where I share buy alerts, portfolio updates, and new investment ideas.

Join hundreds of paid members already inside.

The subscription price is going up soon.

If you join now, you’ll lock in the current $150/year price forever for as long as you remain subscribed.

That’s less than $0.42/day for full access.

Upgrade Now