Yesterday, the Market Took My History Teacher's Test

Every test in Kevin O'Brien's world history class in high school ended the same way. Stay with me, it might just be the worst call I make in my career.

No scan sheet. No lucky guesses. An essay. Whatever the topic was, empires, wars, revolutions, you finished by building an argument in complete sentences and defending it, while Mr. O'Brien watched a room full of teenagers at St. Anthony High School in Wailuku, Maui, learn the difference between knowing a date and understanding a pattern. Although, Lindsay, Peter, Shane and I weren’t seeing it at the time, two of his ideas are more relevant than ever.

The first was recency bias. We trust the present too much because the present is the part we can see. Last year feels permanent. Last quarter feels like a law of nature. The second idea sat right behind it: history repeats itself. Not the events. The patterns repeat, because the people do. The fear. The greed. The certainty that this time the old rules have been retired.

O'Brien had backup at home. My mom had a saying she reached for constantly: you can't see the forest for the trees. As a kid I never understood it, and I made fun of it, repeating it back in a mocking voice every time it showed up, which was often. I say it now. At work. In my own voice. It is the same lesson O'Brien was teaching, hers just came without the essays. Recency bias is the classroom name for it. You get stuck on the tree in front of you, this quarter, this headline, and you lose the forest it belongs to. It only took me thirty years to be reminded.

I have been thinking about Mr. O'Brien a lot this year, because I now sit in rooms where very smart adults talk about AI, and almost every argument I hear would have failed his test.

Yesterday is why. The Nasdaq fell 5.5 percent, its worst day in nineteen months. Chip stocks fell 12 percent. The trigger was not an earnings miss. Democrats swept the midterms, and data centers, now a campaign issue in state after state, suddenly look like something Washington can slow down. By midday my feed had split into two certainties: this has never happened before, and this is exactly how it ends. Both camps were sure, on about six hours of information. Watching it, I kept hearing a world history teacher from Wailuku. Recency bias. History repeats itself. So I ran his test on the whole argument.

The case for repeating

Start with the railroads, because Mr. O'Brien would have. Railroads remade America. They also got wildly overbuilt. When the Panic of 1893 hit, roughly a quarter of the railroad mileage in this country went into receivership within about a year. Northern Pacific. Union Pacific. Santa Fe. The tracks stayed, and the country kept using them for a century. The investors who paid for them got carried out.

The fiber boom ran the same play with better marketing. Between 1996 and 2001, telecom companies spent more than $500 billion laying fiber optic cable, much of it borrowed, on the theory that internet traffic would double every hundred days forever. Traffic grew fast. It did not grow like that. By 2001, only about 2 to 5 percent of the long-haul fiber in the ground was actually lit. Bandwidth prices fell roughly 90 percent. WorldCom and Global Crossing went bankrupt. The Nasdaq fell 78 percent from its March 2000 peak to its October 2002 bottom, and something like $5 trillion of paper wealth went with it.

And then the part nobody quotes at conferences: a decade later, that same dark fiber carried the modern internet. The technology was real the whole time. Being right about the technology did not save the people who financed it at the top.

Now read the current headlines with that essay in front of you. Chip and cloud money is moving in a circle again. In September 2025, Nvidia said it would invest up to $100 billion in OpenAI, whose data centers will be filled with Nvidia chips. A Bernstein analyst, Stacy Rasgon, wrote that the deal will fuel "circular" concerns. Mr. O'Brien had a shorter phrase for arrangements like that. He called it the pattern. Vendor financing flattered the revenue of Lucent and Nortel all the way into the fiber peak, too.

The case for net new

Here is where the essay gets harder, because the differences this time are real, and pretending otherwise is its own kind of recency bias.

The builders are different. The fiber boom was financed by junk-rated newcomers borrowing like the future had already happened. The AI buildout is being paid for by the most profitable companies in history. The big four platforms guided to roughly $725 billion of capital spending in 2026, and every revision since has gone up. That is an enormous number, roughly twice the telecom peak as a share of the economy, but the balance sheets carrying it are not the balance sheets that broke in 2001. These companies can absorb a bad outcome. WorldCom could not absorb a Tuesday.

The demand is also more visible than it was in 1999. Eyeballs were a metric people invented because there was no revenue to count. Cloud contracts are revenue you can count. Nobody has to promise anything this time. The purchase orders exist.

Meta reported on October 21. Revenue rose 26 percent to $64.6 billion. Real money, from real advertisers. Meta also spent $34.9 billion on capital projects in the quarter, which is more than the $32.8 billion its operations generated in cash. Free cash flow was negative $2.1 billion. Both halves of that paragraph are true. That is the whole argument in one company.

Alphabet reported after the close on October 28. Google Cloud revenue more than doubled, up 103 percent to $30.7 billion. Alphabet also spent $51.4 billion on capital projects in the quarter and posted negative free cash flow of $9.1 billion. The market knocked the stock down about 9 percent after hours. The demand was the best evidence the bulls have produced all year. The cash flow was the best evidence the bears have produced.

Amazon reported after the close on October 29, and it was the strongest demand print of the week. Revenue rose 12 percent to $201.9 billion. AWS grew 39 percent to $46 billion in the quarter, roughly a $184 billion annual run rate, at a 38.7 percent operating margin, and management projected AWS growth above 40 percent in both 2027 and 2028. Amazon held its 2026 capital spending plan at about $220 billion. The bear line sits in the same release: over the last twelve months Amazon generated $168.3 billion in operating cash flow and still posted negative free cash flow of $42.4 billion, because the build is that big. The stock fell about 7 percent after hours anyway. AWS also reported a $578 billion backlog, more than double a year earlier. But 39 percent growth at that scale is not a promise. It is a number in a filing.

And then there is the demand document nobody had in 1999. Anthropic filed its S-1 on October 26, the first audited look inside a frontier AI lab. Revenue went from $41 million in 2023 to $4.6 billion in 2025, then $16.3 billion in the first half of 2026 alone. Customers spending more than $100,000 a year went from about 950 to more than 11,900 in twelve months, and existing customers spent roughly nine times what they had spent a year earlier. The filing says plainly that demand for its products currently exceeds the compute it can get its hands on. That is not a metric invented to avoid counting revenue. That is revenue, counted, with auditors watching.

The same filing hands the bears their paragraph, and an honest essay includes it. Most of that revenue is consumption based, and customers can turn it down. Against it sit more than $500 billion of compute commitments, most of them owed whether the usage shows up or not. In 1999 the demand was imaginary and the bills were real. In 2026 the demand is documented and the bills are bigger. Whether that is a difference in kind or only in timing is the whole question. The S-1 does not answer it. It just makes both sides prove their case with numbers.

So the cash-rich argument needs an update. Three of the strongest balance sheets in history just spent past their own cash flow to fund this build. The balance sheets are still strong. The direction is worth circling in red, because spending that outruns cash flow is how the old pattern starts, even when the companies doing it are good ones.

But Mr. O'Brien would circle one more difference, and it cuts against the bulls. Fiber waited. A strand of glass in the ground in 2001 was still state of the art in 2011, which is why the second owners got rich on it. A GPU does not wait. The useful life of the chips filling these data centers is two to three years before the next generation makes them the expensive option. If demand arrives late this time, the asset does not sit patiently in the ground. It depreciates on a rack. The one asset from the last cycle that forgave its investors is the one thing this cycle does not have.

And notice what yesterday's selloff was not about. Nobody reported that demand for AI had fallen. Amazon's $578 billion backlog did not shrink overnight. The market was not repricing the technology. It was repricing the timeline, and the permission. Mr. O'Brien would have underlined that part. The pattern repeats. The trigger never does.

What the pattern actually says

So, is the AI bubble history repeating, or is it net new? Write the essay honestly and you land where Mr. O'Brien always pushed us: the question is wrong. Both things are true at once, and they always have been.

The technology repeats as real. Railroads were real. The internet was real. AI is real. The financing is what repeats as fragile. Every cycle, the build gets ahead of the proven demand, the money starts chasing its own tail, and the people who point at the gap get told they do not understand the future. Sometimes the gap closes in time. Sometimes the first owners pay for a future that the second owners get to enjoy at a discount.

Recency bias runs in both directions, which is the part I have to remind myself. The bulls overweight this quarter's sold-out capacity. The bears overweight a crash that happened 25 years ago and call it a law. Mr. O'Brien was not teaching us to predict the past. He was teaching us to hold two ideas at the same time without flinching: it has happened before, and it has never happened exactly like this.

His test was never really about the dates. It was about whether you could make the argument, in writing, with the evidence on the table. All of it. Including the evidence that hurt your side.

So here is my call, since the opener promised one. AI is not going anywhere, and neither is the demand. It sits in filings now, with auditors on it, and a $578 billion backlog does not unwind over one bad Thursday. I think the builders spend through the fear, the usage keeps compounding, and the cash flow flips. The risk is not that the demand turns out to be imaginary. The risk is the timeline, the permits, the politics. The stuff that never shows up on a capex chart.

One thing I never did was go back and tell him any of this. Teachers hand out lessons like his and mostly never hear where they landed. So here is the last essay, Mr. O'Brien. Thank you. The class was called world history. It turned out to be about everything since.

Embedded AI Is Beating Standalone Chatbots

Meta AI hit 1.2 billion monthly users by living inside apps people already use. Google's AI Overviews now appear in 43% of searches, and Alexa for Shopping drives 6x more product page views on Amazon than a year ago (Similarweb 2026 Generative AI Landscape).

Direct indexing, one year later: the long/short upgrade

A year ago I moved most of my ETF index money into direct indexing on Frec and wrote up why. The short version was tax-loss harvesting: own the stocks in the index instead of the fund, sell the losers all year, and use those losses against gains somewhere else. The ETF gives you the return and keeps the losses trapped inside the fund. Owning the stocks gives you both.

Same disclosure as last time, because it still applies: this is not a sponsored post. I like the product enough that I once asked the founder about investing in the company. I am a customer. That is the whole relationship.

What a year of the classic version taught me

The harvesting is real, and year one is the easy year to love. Your cost basis is fresh, the market always hands you some losers even when the index finishes up, and the losses pile up faster than most people expect.

Then the math starts working against you. Every harvest steps your basis down. In a market that mostly goes up (latelly), a mature direct index portfolio runs low on losers to sell, and the tax benefit thins out year by year. Frec's classic product charges 0.09% to 0.35% with a $20,000 minimum, and for a young portfolio that trade is easy. The question I kept coming back to was what happens in year four, or year seven, when the portfolio is sitting on gains everywhere you look.

Long/short is the answer to that problem

A long/short direct index holds your index portfolio and adds two extensions on top: extra long positions in stocks the model favors, and short positions in the ones it does not. The shorts do the tax work. Short positions are always churning, which means fresh losses keep showing up on both sides of the book instead of drying up as your basis falls.

The leverage comes in tiers. Frec starts at 140/40, meaning 140% long and 40% short against your capital, and goes up to 250/150. Frec says it is currently the only platform where a US investor can open a 250/150 account directly. The versions of this that existed before, from firms like AQR and Aperio, came through a financial advisor with minimums around $1 million and a second fee stacked on top. Frec's long/short starts at $100,000.

The costs are honest ones, and they are higher than classic. The base fee is 0.50%, plus financing on the borrowed side, roughly 0.23% after tax at the starting tier. Fees and financing both rise with the leverage, up to 1.30% on the fee side at the top end. You are paying roughly five times the classic fee for a harvest that keeps producing. Whether that trade wins depends entirely on your tax bracket and whether you have gains to offset. Mine, in New York City, made the math easy.

The parts the marketing leaves out

Harvested losses defer taxes. They do not erase them. Your basis steps down with every sale, so the bill waits for you at the end unless you die holding the portfolio or donate the shares. Anyone who tells you otherwise is selling something.

Shorting adds risks the classic product does not have. There is a margin account underneath this, borrowed money, and financing costs that move with interest rates. Frec keeps a 35% excess margin floor and rebalances inside the account if a margin call ever comes, and the structure is built to make that unlikely. Unlikely is a word worth reading twice before you add leverage to the most boring part of your portfolio.

And treat the big numbers on Frec's site the way you would treat any vendor's numbers. The claim that long/short can harvest up to 337% of a portfolio in losses is cumulative math over many years, not an annual figure. The direction is right. The headline needs that footnote.

Who should actually do this

You need a taxable account, because none of this does anything inside an IRA. You need gains somewhere for the losses to work against, now or in the next few years. And the higher your marginal rate, the more each harvested dollar is worth, which is why this product makes more sense in New York or California than it does in Texas.

A year in, I have not moved my index money back to ETFs, and I do not plan to. The classic product sold me on owning the stocks. The long/short version is me admitting the first product has an expiration date on its best feature, and paying up to remove it.

Not financial advice. Run your own numbers, or pay someone who will.

Prime Day Shift Muddies Amazon Earnings As AWS Cloud Ramps

Amazon's Q3 earnings are out and it was a doozy with the Prime Day shift.

In summary, Amazon managed the business to a higher operating margin, with AWS driving growth and inventory roughly in line with sales on a YoY basis.

Three points for brands selling and advertising on the platform:

1. Prime Day timing muddies Q3, but the slowdown is real. Service sales growth fell from 24% in Q2 to 19% in Q3. Ad revenue growth fell from 26% in Q2 to 12% before currency in Q3, and Amazon's Q4 guide for high-teens ad growth still sits below the low-20s pace Amazon had been running. Online Stores grew 2% before currency, versus 8% a year ago. That matches what happened during Prime: shoppers showed up, spent carefully, and a better Fall Prime did not change the combined quarterly trend.

2. Q4 planning should protect winners and starve the tail. Paid units grew 12% and third-party seller services rose 10% before currency to $46.5B, so demand is still moving through the marketplace. Product sales grew only 2%, while Amazon inventories rose 11.6% from Q2 into the holiday quarter. Keep hero ASINs in stock. Be strict on slow movers, weak bundles and bulk inventory buys that can age into January.

3. Ad budgets should be defended with proof, not habit. Amazon guided Q4 ad revenue growth to the high teens, so spend will come back into the auction and CPC pressure will follow. Still, Q2 plus Q3 plus the Q4 guide points to a slower business than the one brands planned against earlier this year. Use ASIN priorities, margin floors and warm audiences built in DSP from custom combo audiences, not off-the-shelf audiences, and feed the system better signals through Conversions API to ensure you aren't buying a customer that's already yours. Judge the spend on incremental sales and contribution margin. Platform ROAS alone is too easy to game in a quarter this messy.

ChatGPT's Share Fell From 85% to 54% - Here's the Data

Similarweb's 2026 Generative AI Landscape: ChatGPT's visits more than doubled to 5.6B a month, yet its share of chatbot traffic dropped from 85% to 54%. Gemini holds 28%, Claude 9%. Gen AI overall: 9.5B monthly visits, up 70% year over year.

The #1 Use for AI Agents & Assistants Is Shopping - Zuckerberg Just Proved It

The number one use for AI agents and assistants is shopping - and Zuckerberg just proved it. A short take on what happens when the assistant becomes the storefront.

BREAKING: Instacart Is for Sale - and Only the Ads Are Worth Buying

Instacart is exploring a sale to Uber or DoorDash. The picking model is archaic and the growth has stalled against the market - the only asset worth the premium is the advertising business. My take, and why Amazon only bids if the inventory joins its main ads roster.

ep06-preview

Show notes

  • The news: Instacart exploring a sale, preliminary talks with Uber and DoorDash (The Information, Oct 23, 2026). Stock +9% Friday; ~$12B market value; $16-17B at a typical takeover premium.

  • What a buyer is actually buying: users and frequency, not a growth engine - Instacart is growing slower than online grocery overall.

  • Why the picking model is archaic: human shoppers in other retailers' stores vs Walmart first-party, Amazon Now, and bundled platforms.

  • The real prize: Instacart Ads - the toll on every basket, whoever picks it.

  • The Amazon question: a bid only makes sense as an ads-inventory play.

  • For brands: fewer, bigger retail media networks; grocers lose their neutral partner.

Source: The Information report via Reuters, October 23, 2026. Talks are preliminary; no deal is guaranteed.

Transcript

Daniel: Breaking news on Digital Shelf Life, with Ryan Craver. I'm Daniel.

Maya: And I'm Maya. Instacart is for sale.

Daniel: Exploring a sale, officially. The Information reported Friday that Instacart has held preliminary talks with Uber and DoorDash. No deal is guaranteed, and Instacart declined to comment.

Maya: The market commented for them.

Daniel: It did. Instacart stock closed up about nine percent Friday. Uber fell about three, DoorDash about four. Instacart's market value was roughly twelve billion dollars at Thursday's close.

Maya: And with a normal takeover premium?

Daniel: Thirty to forty percent gets you to sixteen or seventeen billion. That is the price tag on the table.

Maya: For what, exactly? That is the question Ryan keeps asking. What is the buyer actually buying?

Daniel: His answer is blunt. One asset is worth the money: the advertising business. Everything else is a user list attached to an archaic model.

Maya: Archaic is a strong word.

Daniel: Defend it, then. Instacart's core model is a human being walking the aisles of somebody else's grocery store, picking bananas off a retail shelf, and handing them to a driver.

Maya: That was genius in twenty fifteen.

Daniel: In twenty fifteen it was the only way to do online grocery without building warehouses. But the world moved. Walmart built pickup and delivery into its own stores and its own app. Amazon is rolling out Amazon Now, groceries in thirty minutes or less. Kroger and Albertsons push their own channels.

Maya: So the pickers are competing with the store itself.

Daniel: With the store, the retailer's own app, and platforms that bundle grocery with rides and restaurant delivery. The middle layer gets squeezed from both sides.

Maya: And the growth story?

Daniel: That is the part the stock chart hides. Online grocery keeps growing, but Instacart is not growing at the rate online grocery is growing.

Maya: Growing slower than your own market means one thing.

Daniel: Losing share. Slowly, politely, but losing it. So a buyer is not purchasing a growth engine. A buyer is purchasing installed users, an Instacart Plus membership base, and frequency it does not have to build.

Maya: Buying users is not automatically dumb.

Daniel: No. Buying users at the right price is a strategy. Uber bought Postmates for users. Uber agreed in July to buy Delivery Hero for nearly fifteen billion. Consolidation is Uber's playbook right now.

Maya: But you said the ads are the prize. Why?

Daniel: Because the ads business is the one part of Instacart with real margin and real leverage. Brands pay to show up at the digital shelf at the exact moment a shopper is building a basket. Closed loop: the ad and the purchase happen in the same place.

Maya: Retail media, grocery edition.

Daniel: Exactly. And it travels. If Uber owns that business, Instacart's advertiser relationships and retailer inventory plug straight into Uber's ads ambitions. If DoorDash owns it, same story, bigger grocery footprint.

Maya: The delivery operation does not travel the same way?

Daniel: The delivery operation is people picking from stores the buyer does not own. The ads business is a toll on every basket, whoever picks it. Ryan's rule from episode one applies: own the toll, not the carrying.

Maya: Okay, the question everyone in Seattle is asking. Why not Amazon?

Daniel: Ryan's answer: potentially interested, but not for the reasons people assume.

Maya: Amazon does not need grocery users?

Daniel: Amazon has Prime. Amazon does not need a picking network; it is building its own rapid delivery. Bolting Instacart's service onto Amazon's service adds almost nothing operationally.

Maya: So what would Amazon actually want?

Daniel: The shelf space. If Instacart's retailer inventory, hundreds of grocery and retail storefronts, got folded into Amazon's main roster of advertising inventory, inside the Amazon ads machine and the DSP, that is real. Amazon selling ads against baskets it does not fulfill.

Maya: Amazon as the ad layer over everyone else's stores.

Daniel: Which is the toll booth endgame, and it is the only version of an Amazon bid that makes strategic sense. Absent that, Amazon watches Uber and DoorDash bid and smiles either way.

Maya: What about the grocers in all this? Instacart was supposed to be their neutral partner. The Switzerland of grocery.

Daniel: Switzerland just got annexed. If your digital storefront and your ad platform are owned by Uber or DoorDash, your partner is now also your biggest competitor's owner. Expect the big retailers to accelerate their own channels even faster.

Maya: And for brands, the CPG advertisers listening?

Daniel: Consolidation of the shelf. Fewer, bigger retail media networks, more bundled buying, and one less independent place to spend. The negotiation table gets smaller for everyone except the platforms.

Maya: Bottom line, then. Does a deal happen?

Daniel: Talks are preliminary, and antitrust will have opinions, especially about DoorDash. But the logic is not subtle. Instacart proved Americans will buy groceries online.

Maya: It just never owned the store.

Daniel: Or the shelf. Or the road. Now the road is doing the buying, and the only part of the company truly worth the toll is the ads.

Maya: Digital Shelf Life. Instacart's expiration date may have just been printed.

Daniel: We will be watching the filings. See you next episode.

Personal Assistants (Muse and Instinct) Trump Chatbots

Part 4, and the point of the whole series: personal assistants bring AI to the masses.

Chatbots haven't truly been adopted by everyday people, and almost nobody will pay for one. A chat window is a destination you have to remember to visit. An assistant is different - it lives in your messages, your calendar, your day, and it does things instead of just answering things.

That's where Muse and Instinct come in: AI built for everyday, every-hour usage, distributed through the apps people already live in. The leaderboard can freeze all it wants. The next hundred million users won't arrive through a chatbot. They'll arrive through an assistant that was already in their pocket.

AI Spend Doesn't Equal Traffic

Part 3: spend and traffic are telling two different stories.

In a16z's latest AI leaderboard, only one app dominated across all three measures - web traffic, mobile traffic, and consumer spend. Everyone else is strong in one lane and weak in the others. Big traffic with no revenue. Real revenue with modest traffic.

For anyone allocating budget or building in this space, that's the tell. The consumer AI market isn't one market. It's a traffic game, a subscription game, and an attention game being played by different players - and almost nobody is winning all three at once.

Who's Actually Paying for AI?

Part 2 of the a16z leaderboard story: the spenders.

a16z's latest data shows only the builders and the productivity crowd are actually paying for AI subscriptions. The masses just aren't interested in paying. Free tiers get the traffic; a thin slice of power users carries the revenue.

That's the uncomfortable math under the whole consumer AI boom. Hundreds of millions of users, a small fraction of payers. The models keep getting more expensive to run while the willingness to pay stays flat for everyone outside the builder class.

Which raises the real question: where does the money come from? Advertising, distribution bundles, or assistants people actually live inside. Part 3 looks at why spend and traffic don't move together.

The AI Leaderboard Is Freezing Over

Three years into tracking consumer AI, the leaderboard is barely moving.

In the latest a16z Top 100 Gen AI apps, only 11 products are making their first-ever appearance - the smallest number of debuts across all seven editions of the list. ChatGPT is still #1 by traffic. But Claude has quietly become the clear #3, passing DeepSeek and Perplexity, and going neck-and-neck with Gemini on paid subscribers.

The real story isn't who's winning. It's how frozen the top has become - and what that means for new apps trying to break in. The window for a new consumer AI app to crack the top tier is closing. Distribution and habit beat novelty now.

Part 1 of 4 on the state of consumer AI.

Amazon = Toll Booth of E-commerce

Amazon's Accelerate conference wrapped in Seattle this week with three announcements every seller should know: Seller Assistant going agentic inside Claude and Amazon Quick, Walmart/Shopify/eBay/TikTok orders managed from Seller Central, and free 24-hour payouts.

The pattern across all three is unmistakable — Amazon is positioning itself as the toll booth of e-commerce, taking a cut of every order on every channel.

Here's the 60-second breakdown and the play for sellers.

OpenAI $100 billion in annual advertising sales by 2030?

OpenAI's staggering advertising ambitions — a reported $100 billion a year by 2030 — and what it means for Google, Meta, and the future of the ad industry. A 78-second explainer on the race to monetize AI attention, and his most-watched video to date.

Phase 3 of AI: Employees vs Tokens vs Model Routing

Phase 1 of AI was magic, Phase 2 was adoption — Phase 3 is economics. Companies are now forced to choose between employees, tokens, and model routing, and how that trade-off is reshaping budgets and strategy. The great AI cost debate.

Instinct vs Meta's Muse for the AI Assistant Race #instinct #muse #ai

With no unfamiliar interface to learn, calling or texting your AI like a person leads to more usage and longer conversations, with simplicity driving adoption. Instinct has the Silicon Valley buzz, while Meta’s Muse has the distribution potential of 4 billion users globally — and the question now is who wins, and how soon Google, ChatGPT, and Apple launch something similar.

Why would Stripe acquire OpenRouter?

Stripe acquiring OpenRouter sounds like an odd pairing — until you think about where AI payments are headed. The strategic logic: who gets the toll-booth position on AI agent transactions, and why owning the model gateway could be Stripe's smartest move. A short, sharp take on one of fintech's most interesting what-ifs.

Space race. AI chip race. Taxi race. Drones race?

Amazon vs Walmart vs DoorDash vs Uber. Everyone wants a piece of final mile delivery via air.