Business history and operating math on today's AI and markets news. Founder, Operating. Saw Facebook start at Harvard. Newsletter below.
- Tweets 1,560
- Following 401
- Followers 187
- Likes 901
Brett Harrison is building an exchange where AI inference providers keep moving prices to their marginal cost. Next he wants auctions for three-month inference contracts. When compute trades like a commodity, the buyer who can switch wins. @BrettHarrison on @theinformation
1
2
56
Howard Marks on October 2008, when Oaktree was buying and everyone else was selling: "It's by doing the things that other people think are nuts. That's how you make the big money." His catch is that you have to do it and be right. When you're the only buyer, the seller takes your price.
@InvestingCanons
1
24
1/ A24 opened "You Can See Everything" in 855 theaters and grossed $15 million. The film had not been announced before Sept 6. The interesting part is how they launched it.
1
1
66
Munger in 2019 on Mozart. Maybe the greatest talent who ever lived, and he ruined it two ways. He spent more than he made and he was eaten up with resentment. I've seen both sink companies with great products.
(via @TmarketL, Daily Journal 2019)
2
99
Howard Marks on Friday called AI the least knowable thing he's seen in his career. Then the part I'd underline. If you want the upside, own the equity. A lender gets his money back and nothing more.
(via @MikeLongTerm, CNBC)
2
25
Ken Griffin on buying a forced seller's portfolio. Once the market smells trouble, traders front-run the selling and 'transform a possibility into a reality.' Whoever steps in with a full solution gets the reversal when confidence comes back. @MollySOShea, Bloomberg
2
53
Alexandr Wang says most adults have had the ambition they had as kids beaten out of them. His hope for AI is an "escalator of agency": make one thing happen, dream bigger, then make the next thing happen.
That's how a good team compounds too.
via @rohanpaul_ai, from Cheeky Pint with John Collison
65
Charlie Munger's rule: hang around good people, and behave better every year than the year before. The higher you rise, the more it matters. He and Buffett believed good behavior paid. The high road is uncrowded, and there's little competition on it.
@InvestingCanons, CNBC (2016)
1
29
United's Scott Kirby calls Starlink a "truly differentiating" revenue tailwind for 2027. Every airline gets fast Wi-Fi eventually. United's bet is to be first by about two years, win flyers away from other airlines, and keep them once they try the rest of the product. A head start only pays if it turns into retention.
@cb_doge, United Airlines Q3 2026 earnings call
41
Meta is a huge company. Alexandr Wang says the research team building its frontier models is close to 100 people, plus infra, and it's kept intentionally tight.
The hardest work still goes to small teams with clear owners.
via @rohanpaul_ai, from Cheeky Pint with John Collison
1
32
Intel's CTO on why Elon wants his own fab: "We either build the Terafab or we don't have the chips." Tools are the bottleneck, so a new fab doesn't add silicon to the world. It guarantees the owner gets some. When supply is scarce, ownership decides who waits.
@firesidealpha, Dwarkesh Podcast
2
108
IBM's Arvind Krishna on why the mainframe has lasted 65 years: it was vertically integrated. In the '60s IBM went "all the way to dirt," from sand to wafers to fabs to chips. Today it does everything but the fab. His rule: own the layers where you have an edge, and buy the commodity. Nikesh Arora adds the payoff: control the stack and you decide where the profit sits.
@theallinpod, All-In Interview
30
Generations of scientists, engineers, entrepreneurs, and investors built the foundations beneath the laptop on your kitchen table.
What are you going to do with it?
1
2
You didn't have to invent the transistor.
Build the internet.
Launch a satellite.
Or finance a data center.
But you get to build on top of them.
That should stop us in our tracks.
1
2
And somewhere along the way, each breakthrough became the foundation for someone else's ambition.
The semiconductor helped make the personal computer possible.
The internet connected those computers.
The cloud made infrastructure available on demand.
1
5
Computers that filled rooms.
Computers that sat on desks.
Computers that fit in pockets.
Computers we can now ask to help us build almost anything.
1
11
Our grandparents watched computers become companies.
Our children will watch people become companies.
Look at this timeline.
Roughly 75 years of people turning the seemingly impossible into something the next generation takes for granted.
1
1
23
Marc Andreessen says the gap between AI-native startups and incumbents is the widest he's ever seen. Startups plan around what the tools can do this week. Incumbents plan around last year. That's how industries change hands.
@a16z, Frontier
30
Emily Chang asked Jensen Huang if Nvidia would stop selling to a lab doing something too dangerous. "It's a hypothetical, but of course." A conversation comes first. Every supplier needs that line drawn before the call comes. Jensen Huang with @emilychangtv, via @BloombergTV
38
We've seen this before. Global Crossing reported $472 million of "Recurring Adjusted EBITDA" for Q2 2001 and a $607 million recurring net loss. CoreWeave reported $1.51 billion of adjusted EBITDA in Q2 2026 and lost $626 million.
1
7
Useful lives keep moving. Microsoft stretched servers from four years to six in 2022. Amazon and Oracle moved to six in 2024. Meta went to five and a half in 2025, cutting that year's depreciation by $2.92 billion.
1
1
27
Alphabet spent $44.9 billion on capex in Q2 and recorded $7.1 billion of depreciation. Meta spent $34.9 billion in Q3 against $7.2 billion. Oracle spent $28.5 billion against $3.2 billion. That gap is depreciation still to come.
2
1
30
AI compute demand is exploding. But an enormous market cannot rescue every business model inside it.
The companies supplying it can still go bankrupt.
That's the uncomfortable possibility behind Elon Musk's latest warning about the AI infrastructure boom.
1
2
28
Bill Ackman calls Netflix "a high-growth utility." He watches three things: subscriber growth, how long people stay, and pricing power. By his count it trades under 20x earnings, and the company is buying back stock. That's why he says a cheap price in the short run helps patient holders: buybacks retire more shares.
@RealNickMugalli, CNBC Squawk on the Street
56
Jon Gray's bubble test: are you building capacity and hoping demand shows up, or struggling to keep up with demand that's already here? '90s telecom fiber was the first kind. He puts AI in the second: spend on models at Blackstone portfolio companies is up 20x in a year, and power, permits and chips are the bottleneck.
@BloombergTV, Bloomberg Television
2
47
Shou Zi Chew says a video ad that cost a brand $100 to $1,000 to make can now be produced for $1 to $10 with TikTok's Seedance model. He also says most short dramas are already AI-produced.
When making the ad costs almost nothing, the work moves to knowing what to say.
🎥 via @firesidealpha, from DBS Asia Future Forward
2
63
Mark Zuckerberg was ready to write a $100B or $200B check before Meta had seen the models. His question to Alexandr Wang: "Are you sure this is going to work?"
You don't get to meter a frontier bet. You commit first and find out later. That takes a different kind of operator.
🎥 via @guilleflorvs, from Cheeky Pint with John Collison
2
50
Masayoshi Son says SoftBank has more than 30 robotics companies in its portfolio, "very quietly preparing." The target is tiring work like doing dishes, so people can do something more interesting.
Start automating with the work nobody wants to keep doing.
🎥 via @firesidealpha, from Inside Blackstone with Christine Anderson
2
51
Marc Andreessen thinks robotics could be the biggest industry in the history of the planet, with hundreds of billions of robots that someone has to design and build.
The factory floor is where that gets decided.
Marc Andreessen with Joe Lonsdale on American Optimist, via @linasbeliunas
1
2
71
Demis Hassabis had Gemini 4 Argon build a wizard chess game. Watch the pieces move.
A demo you can see beats a benchmark you have to explain.
🎥 via @demishassabis
1
70
Every operator I talk to thinks their data is the moat.
David Vélez's Blockbuster story says otherwise. They had the customers, the stores, the rentals. Netflix had the systems and the habit of using them.
Capture first. Then build the culture that acts on it.
Invest Like the Best, via @linasbeliunas
1
1
50
So the budgets are getting owners. Microsoft: "more impact per token." Meta and Microsoft cut back Claude Code. Accenture: $29 a month, the rest billed to a project code.
Every useful input runs the same arc. Perk, then habit, then a line item with an owner.
1
21
Most companies can't compute the denominator yet. Per the WSJ, the cheaper-per-token model cost more overall in 32% of 6,800+ tasks. Only 11% of 396 organizations forecast AI spend within 10%.
If you can't forecast it, all that's left is a blunt cap.
1
1
4
Ask each side what would change their mind and neither names a token count.
Azhar: enterprise AI spending growth halving two quarters running. Kedrosky: productivity and GDP breaking above trend.
Both tripwires sit on the buyer's side of the ledger.
1
9
The bear's evidence is real. The NYT reports CoreWeave carries $72B in liabilities to run 51 data centers, and investors put better than 60% odds on it missing its obligations.
If you sign multi-year compute contracts, your vendor's credit is part of the price.
1
13
"Tokens times price" is the seller's equation. It describes a lab's revenue.
The buyer's equation has a denominator: tokens times price, divided by the work that actually got done. Call it the cost per finished task.
1
14
The AI bubble debate is arguing over the wrong number.
Bears count tokens times price. Bulls count demand. Buyers count what a finished task costs, and companies are only now finding out.
1
28
Shou Zi Chew on how TikTok checks what a French-speaking 14-year-old in Nigeria actually sees. It used to be one person setting up an account, watching a hundred videos and writing a report. Now it's a thousand agents overnight, with the report ready by morning.
The audit used to be a sample. Now it can cover the whole shelf.
🎥 via @firesidealpha, from DBS Asia Future Forward
1
36
TheFacebook dropped "The" in 2005 and became the name for everyone, not just Harvard. I was on campus when it started. MySpace's parent sold to News Corp that same year for $580M, and MySpace never caught up.
Peter Thiel on why names predict outcomes. Via @jawwwn_
33
Zuckerberg says an H100 is worth more today than it was a year ago.
Own the asset that outlasts its depreciation schedule.
🎥 @firesidealpha on Cheeky Pint
1
1
31
Ben Horowitz says Andy Jassy told him Amazon is sold out until the end of 2028, maybe longer. In 2001 the debt showed up years before the customers did. This time the customers are lined up, and Horowitz still keeps a16z out of the neoclouds. Ben Horowitz on All-In, via @dnapway
1
48
Alexandr Wang says Meta's huge offers were roughly net neutral for top researchers.
Pay gets you in the room. The mission closes the hire.
🎥 @rohanpaul_ai via David Senra's podcast
26
The X bot also runs my video: find the clip, download it, transcribe it, crop it to vertical, write the copy, wait for my yes, schedule it in Buffer.
It started October 11 and reached four platforms by October 13. 174 posts sit in Buffer for October 13 to 24.
1
10
Every public post still waits for my approval.
That rule has protected the brand. It has also made me the slowest part of the system.
Once I say yes, the work leaves through Buffer, the browser's X session, or the browser's Substack session.
1
1
8
Routines are what make the team run while you sleep.
About 30 scheduled runs across five bots every weekday. The News Bot opens at 6:47 AM. The X bot closes at 9:53 PM with a voice review comparing what I posted against what it drafted.
1
1
11
The team has four groups:
Personal ops: Grok Bot as the front desk, plus a Chief of Staff
Content desk: the X, Substack, and LinkedIn bots
Research: a News Bot and an Analyzer
Client work: a pricing research bot walled off in its own database
ALT Org chart of John Brewton's AI operating team in Grok Bot: eight working bots in four groups, a meta bot, and six group rooms.
1
1
16
On September 1 I hired my first AI bot to keep me out of calendar noise.
Six weeks later I run eight working bots and six group rooms.
On a typical weekday they send me about 24 items and publish about 28 posts across five platforms.
Here's how it's built.
ALT Title card: How to Build an AI Operating Team in Grok Bot. Eight working bots, six group rooms, one approval gate.
1
1
258
Nvidia and Elon Musk's teams are designing a server for data centers in space. Jensen Huang says it might be just as good on Earth, and Musk says it ships in volume next year. Build for the harshest customer and everyone gets a better product. Huang on @BG2Pod, via @firesidealpha
3
41
Zuckerberg talked to the heads of every AI lab, then decided "Alex is the guy."
Diligence wide. Commitment narrow.
🎥 @firesidealpha on Cheeky Pint
2
27
The coding data, the people, and a sales team already inside thousands of companies.
Buyers pay for distribution as much as product.
🎥 @muskonomy
1
38
AI makes execution faster. Your time understanding the problem decides whether that speed creates value.
1
8
Vitalik Buterin says that without AI, parts of the Ethereum roadmap might have stretched to 2040 or 2050.
Worth rereading your own someday list with that in mind.
🎥 @coinbureau
31
One intern table at Hudson River Trading: Scale, Cognition, Hyperliquid, Decagon.
Talent density compounds.
🎥 @GoshawkTrades
77
"They had us in checkmate. We just invented a new piece."
The best operators change the board when the math says they've lost.
🎥 @ArkkDaily
2
15
Qualcomm's CEO says AI labs want phones running 100B-parameter models all day by 2028.
Your most demanding customers are handing you your roadmap. Ask them early.
🎥 @firesidealpha
1
15
Alexandr Wang says Meta's CFO kept pulling him aside to tell him the model had to work.
Every big bet needs someone who says the downside out loud. Then you build anyway.
🎥 @firesidealpha on Cheeky Pint
29
Zuckerberg says that in five years everyone will have an AI agent that "really intimately understands your goals and everything around you."
The winning agent will be the one people trust with that much context.
🎥 @linasbeliunas
1
21
Deliberate thoroughly, then once you're sure, empty the clip.
Most companies lose their time after the decision is made.
🎥 @davidsenra
1
1
24
Aravind Srinivas on a one-person company spending about $1M a year on Perplexity and making about $3M.
One person. $1M a year on AI. $3M in revenue. The org chart of the future fits on an invoice.
🎥 @MollySOShea on Sourcery
20
"I believe so much in AI that it's going to cure cancer… my chips are all in."
Shyam Sankar calls it a mass psychosis. Judge AI by what ships and what it changes for the people doing the work.
🎥 @firesidealpha
31
Zuckerberg on Musk making companies more technical.
Every layer between the person doing the work and the person deciding costs a week.
🎥 @rdominguezibar
1
2
107
Pessimism is getting loud. The opportunity is getting bigger.
I'm a techno-optimist, because optimism changes what we do next.
1
26
Emad Mostaque: labor and capital are about to divorce.
For most of history, capital needed people to put it to work. The owners who win next are the ones who still know what the work is and what the potential hurdles are that will become obstructions and need humans in the loop to solve once the work leaves the agentic environment.
🎥 @rohanpaul_ai
1
19
Your next competitor might be a parent with a smartphone.
Parents are using ChatGPT to make their own school portraits: a photo against a plain wall, then a request for the familiar school-picture background.
Packages commonly run $25 to $60.
1
26
Been building this over the last week. @bot is undoubtedly the most important new tool I've added to my stack all year.
Here is the org chart I've built out initially.
Complete how-to article and deep dive on what I've been building, coming Wednesday for my paid subs at Operating by John Brewton on Substack.
3
21
This is not really a workbook you read. Upload the PDF into Claude or ChatGPT. Paste one sentence describing what you do. The workbook takes over from there.
You give it one capability and answer a short interview. It scores six business models against your constraints, shows you the grid before it sorts, lets you correct anything you disagree with, and then produces a one-page PDF with the result.
The workbook estimates the conversation at about seven minutes.
Go get to work.
25
The AI jobs debate is not about jobs created from the buildout of infrastructure. It is about the jobs that will be created from the industries that are built on top of that infrastructure.
1
10
A great deal does three things.
Improves the economics.
Protects your ability to actually deliver.
Raises the odds both sides want to do it again.
ALT A great negotiation does three things. 01 Improves the economics: the deal makes financial sense. 02 Protects delivery: you can actually fulfill what you sold. 03 Creates the next deal: both sides want to return to the table. Winning the contract is not the same as winning the deal.
1
13
For years I had no trouble defending the economics of a company.
Then my own name replaced the logo on the proposal.
The hardest person I ever negotiated for was myself.
ALT I negotiated more than $250,000,000 for companies. My first coaching offer: 4 hours of my time, $99 total price. The hardest person I ever negotiated for was myself. When your name replaces the company logo, pricing suddenly feels different.
7
Proof. Market. Value. Price. Terms. Walk.
Most people start at price. It is the fourth question, not the first.
ALT The Operating Negotiation Framework. Six stages in order: 01 Proof, what can you demonstrate. 02 Market, what alternatives exist. 03 Value, why does solving this matter. 04 Price, what does the evidence support. 05 Terms, what can move besides price. 06 Walk, when is no deal better. Price is step four. Most people start there.
6
A customer asks for 10% off. That is a position, not a reason.
Maybe procurement needs to show savings. Maybe cash flow is the real issue. Maybe delivery speed matters more than price.
ALT The price request isn't always the problem. A position reads: we need the price reduced by ten percent. Ask why. It branches into eight possible interests: savings target, cash flow, competitor quote, budget constraint, delivery risk, service level, volume, internal approval. Understand the interest and one variable becomes eight.
5
Revenue can make very intelligent people do very stupid things.
A large contract appears. Everyone sees the top-line number. Fewer people ask what operating it actually requires.
ALT $5,000,000. A five million dollar contract can still be a terrible deal. Revenue does not equal good revenue. Six questions: can you deliver it profitably, can you finance it, do you have the capacity, what risks are hidden inside it, what does it consume, what better opportunity does it prevent. Never win a deal you cannot afford to deliver.
1
8
Three prices to know before any important negotiation.
Market price: what this problem normally costs to solve.
Strategic price: what this relationship justifies.
Walk-away price: where the work stops making sense.
ALT Know these three prices before you negotiate. 01 Market price: what does this problem normally cost to solve? 02 Strategic price: does this relationship justify different economics? 03 Walk-away price: below what number does the work stop making sense? Decide before emotion enters the room.
1
7
