This is a library of people who move large amounts of somebody else's money and have to explain the actual mechanism afterward, so what follows is the discipline that survives across them rather than any one firm's house view. What they converge on is unglamorous: build the seat before you chase the trade, do the primary work yourself, and remove the structures that force you to buy and sell at the wrong moments. Where they contradict each other, which is on whether the current AI buildout is a bubble, on concentration, on who should run an acquired company, on what a moat actually is, and on whether the edge is volume or cost, the contradiction is written out below with names attached.

Do

Reduce every option to one comparable unit before you rank them. Alan Waxman's method, carried over from Goldman's Special Situations Group, strips an investment to three variables: the quality of the business and sector, where your capital sits in the structure, and the documents governing it. That is what lets one team weigh a leveraged consumer buyout against a fifteen-year data-center contract on the same scale rather than in separate silos. [Source: "The Investment Firm That Can 'Do Anything' | Sixth Street CEO Alan Waxman"]

Do the primary work yourself and never outsource the read on a person. Greenoaks nearly passed on backing Elon Musk early at SpaceX because secondhand diligence relayed by mentors made the traits sound disqualifying, and Neil Mehta no longer outsources primary diligence on anyone the firm is seriously considering. Dan Sundheim found what he believed was expense capitalization fraud at twenty-four because he built the model himself and the unit economics would not reconcile. Third Point's FTX loss added direct verification steps, including checking bank balances. [Source: "Finding the Next Figma, Wiz, & Stripe Before It's Obvious | Neil Mehta Interview"] [Source: "Inside Dan Sundheim's Bets on Anthropic, OpenAI, and SpaceX"] [Source: "Legendary Investor Dan Loeb on AI, Credit, & Third Point's $25B Strategy"]

Only enter a position you would want to add to at a higher price. Henry Ellenbogen's underwriting rule is that the firm must be able to write a memo saying it would buy more once the thesis plays out, and if the only bull case is an acquisition, they pass; in practice that meant buying more of unprofitable public names through the 2022 selloff rather than exiting. Greenoaks put just under a billion dollars into Coupang across roughly ten years, led five of eight rounds, and was still buying public shares fifteen years in. [Source: "Finding The 1% of Stocks That Matter | Henry Ellenbogen Interview"] [Source: "Finding the Next Figma, Wiz, & Stripe Before It's Obvious | Neil Mehta Interview"]

Judge a leader by output you can inspect, not by pedigree. Sundheim chose Anthropic over the more established OpenAI almost entirely on reading Dario Amodei's own essays, weighing clarity of written thought the same way he weighs Jeff Bezos's 1997 shareholder letter, a document he says he missed acting on and counts among his costliest misses. Mehta says if he could ask an employee only one question it would be whether their best days feel ahead of them, ranking that above growth rate, margins, or any formal moat framework. Graham Weaver screens young CEOs in a fixed order: will to win, grit, self-awareness, bias for action. [Source: "Inside Dan Sundheim's Bets on Anthropic, OpenAI, and SpaceX"] [Source: "Finding the Next Figma, Wiz, & Stripe Before It's Obvious | Neil Mehta Interview"] [Source: "How Graham Weaver Turns Recent Grads Into CEOs"]

Engineer the seat before you chase the trade. John Arnold's theory of his own edge was not a single insight but leaving Enron in 2001 to build better economics (two and twenty, later three and thirty-five) that then funded top talent, proprietary data sources, and custom trade-entry systems, compounding into a flywheel. Apollo merged with Athene for a permanent balance sheet just over $300 billion and spent nearly $10 billion of its own capital from 2014 to 2022 building origination platforms it owns outright. Sixth Street runs TAO, a roughly $30 billion flexible vehicle, so it can write billion-dollar checks without bloating every specialist fund to match. [Source: "The World's Greatest Energy Trader on Markets, China, and AI"] [Source: "How Apollo Built an $800 Billion Capital Markets Giant | John Zito Interview"] [Source: "The Investment Firm That Can 'Do Anything' | Sixth Street CEO Alan Waxman"]

Remove the structure that forces you to act at the wrong time. Martín Escobari's trick for outperforming in Latin America was that General Atlantic deliberately never raised a dedicated regional fund, because a regional fund forces you to buy when a region is hot and sell when it is out of favor; the firm's hybrid evergreen model means it is never out of dry powder exactly when markets are cheapest. Alpine states 5x MOIC rather than IRR as its objective specifically because chasing IRR pushes firms to sell winners early for fundraising optics. [Source: "Inside General Atlantic: How a $100B Growth Equity Firm Invests"] [Source: "How Graham Weaver Turns Recent Grads Into CEOs"]

Lock every assumption before anyone is allowed to see the answer. Bending Spoons debates and finalizes each input to a deal model before the team looks at the resulting profit and loss, explicitly forbidding the peek so nobody rationalizes backward toward a nicer number, then runs a Monte Carlo across the assumption distributions to produce the ranges that guide the offer. [Source: "The Playbook on Buying and Running Companies Forever"]

Institutionalize the look-back, because slow drift is invisible quarter to quarter. Durable runs quarterly operating reviews position by position with the whole investment team, plus a formal three-year look-back on every holding comparing what was underwritten to what happened, specifically to catch the case where twelve straight quarters of small deviation have become undeniable. After the Qwikster split cost Netflix roughly 75 percent of its stock price and the post-mortem found executives had privately doubted it and stayed quiet, Netflix built a shared document where every leader scores major decisions from minus ten to plus ten before they ship. [Source: "Finding The 1% of Stocks That Matter | Henry Ellenbogen Interview"] [Source: "The Netflix Culture Code That Changed Entertainment Forever | Reed Hastings Interview"]

Name which of your two failure modes is actually the expensive one, then lean that way on purpose. Krishna Rao spends thirty to forty percent of his time on compute because buying too little means losing the frontier and the customers with it, while buying too much means going out of business. Anthropic models a wide cone of uncertainty one to two years out and deliberately buys toward the top end rather than the median, accepting visible overcapacity as the cheaper mistake. [Source: "Inside Anthropic's $100 Billion AI Compute Commitment | CFO Krishna Rao"]

Don't

Don't let a valuation stand because nobody is paid to move it. Gurley estimates roughly a thousand private zombie unicorns, each having raised over a billion dollars in the zero-rate era, sitting on limited-partner books at around three trillion dollars in aggregate and still marked near 2021 peaks, with the people setting those marks sometimes bonused on paper valuations. The liquidation preference mechanic locks it in: a company that raised $300 million and is now worth $400 million could see preferences claim roughly 75 percent of a sale, so nobody wants to acknowledge the markdown. [Source: "The Gift and The Curse of Staying Private with Bill Gurley"]

Don't ride a theme after everyone else has copied it. Waxman gives a good theme a shelf life of roughly twelve to thirty-six months before it gets crowded and overleveraged, and migrates capital away rather than riding it down. Gurley's version is Yale, whose endowment pioneered heavy illiquid allocation and compounded around thirteen percent over thirty-five years, now itself shopping six billion dollars of private-equity secondaries once every major institution has adopted the same model. [Source: "The Investment Firm That Can 'Do Anything' | Sixth Street CEO Alan Waxman"] [Source: "The Gift and The Curse of Staying Private with Bill Gurley"]

Don't defend a margin your new cost structure will not support. Gavin Baker's argument is that traditional software carries seventy to ninety percent gross margins because it is written once and distributed cheaply, while an AI product recomputes an answer on every query, so AI-native competitors run profitably around thirty-five to forty percent; a company defending eighty percent while building an agent strategy is guaranteeing it loses that business. Alex Sacerdote acted on the same read, selling down almost all of Whale Rock's application-layer software and going net short entering the year. [Source: "GPUs, TPUs, & The Economics of AI Explained | Gavin Baker Interview"] [Source: "Why the AI Boom Is Just Getting Started"]

Don't average opinions on a decision that only pays if it is non-consensus. Netflix's informed captain model deliberately avoids committees: one person gathers as much outside opinion as possible and then makes the call alone, on the reasoning that averaging destroys exactly the contrarian thinking that created most of the company's value. [Source: "The Netflix Culture Code That Changed Entertainment Forever | Reed Hastings Interview"]

Don't keep someone you would not fight to keep. The keeper test asks a manager one question: if this person resigned, would you fight to keep them. Netflix paired it with four to nine month severance packages for the specific purpose of making managers actually willing to act, and ran roughly twenty percent first-year attrition as the cost of the model. [Source: "The Netflix Culture Code That Changed Entertainment Forever | Reed Hastings Interview"]

Don't assume the standard process will deliver. Shyam Sankar's claim is that nothing which goes through the ordinary bureaucratic process ever actually ships, which is why progress comes from insiders who fight their own institution: Hyman Rickover built the first nuclear submarine in seven years after being told it would fail, with his first office on the project a women's restroom assigned as a humiliation, and Billy Mitchell was court-martialed and died penniless for advocating an air force. [Source: "How AI Is Changing Warfare | Palantir CTO"]

Don't leave the cheapest input in the plan unchecked. During Operation Warp Speed the team discovered mid-effort that the country held only 1.2 million testing swabs, enough for thirty-seven sites, and had to airlift swabs from an Italian factory. Kushner's reading of it is that the most overlooked, lowest-cost item is what becomes the actual bottleneck under stress. [Source: "Jared Kushner - The Mechanic"]

Don't sell the position that was going to pay for all the others. Ellenbogen's founding insight came from studying fifty years of T. Rowe's New Horizons archives and finding that Walmart had been sold shortly after its IPO; held, that one position alone would have been worth more than the entire roughly eight billion dollar fund he later ran, mathematically outweighing every other good decision made across five decades. [Source: "Finding The 1% of Stocks That Matter | Henry Ellenbogen Interview"]

Where they disagree

Is the AI buildout a bubble, or the correctly priced trade? Paul Tudor Jones reads stock market capitalization to GDP at roughly 252 percent today against about 170 percent at the 2000 peak, 85 to 90 percent in 1987 and 65 percent in 1929, and warns that a wave of IPOs and post-lockup unlocks could reverse a decade of buyback-driven share shrinkage. Bill Gurley adds that some reported AI revenue may be resold compute counted several times across wrapper, model and hosting layers, several of them running at negative gross margin, which makes the ten to twenty times revenue multiples less trustworthy than they look. Against them, Dan Loeb argues the current semiconductor and hyperscaler rally is fundamentally unlike the dot-com bubble because the capex is funded largely off real balance-sheet cash flow rather than speculative external financing, and Alex Sacerdote calls enterprise adoption an L-curve rather than an S-curve, starting from under one percent penetration with no visible ceiling. What decides it is who funds the marginal gigawatt, and on that Dylan Patel's arithmetic cuts against Loeb: a gigawatt costs roughly fifty billion dollars, Nvidia's pledge covers about ten billion of it, leaving OpenAI to source the other forty billion through debt or infrastructure funds, while roughly half of Nvidia's own gross profit on the deal cycles back as equity. That is precisely the external and circular financing Loeb says is not present. [Source: "Legendary Trader Paul Tudor Jones on AI Risk, Bubbles and Buffett"] [Source: "The Gift and The Curse of Staying Private with Bill Gurley"] [Source: "Legendary Investor Dan Loeb on AI, Credit, & Third Point's $25B Strategy"] [Source: "Why the AI Boom Is Just Getting Started"] [Source: "Inside the Trillion-Dollar AI Buildout | Dylan Patel Interview"]

Concentrate everything on a handful, or build a portfolio that cannot lose? Neil Mehta says only ten to fifteen founders a year worldwide meet his bar, so Greenoaks stays deliberately small and Mehta personally takes first meetings rather than running a sector-and-stage coverage matrix. Ellenbogen's version comes from the arithmetic that over any rolling ten-year period only about forty of roughly four thousand US stocks compound at twenty percent or better, and roughly eighty percent of those started small, so the entire firm is built to catch one of the forty. Against them, Escobari's General Atlantic runs a loss ratio near four percent against a typical twenty to forty percent for venture and growth equity by refusing binary risk outright, its worst-case underwriting being that a company merely grows into the price already paid. Weaver underwrites each Alpine deal to about 3x over five years on debt and operational growth rather than multiple expansion, relying on portfolio asymmetry to reach the 5x fund target. The interesting case is Dan Sundheim, who sits on both sides across time: he made concentrated private bets on OpenAI, Anthropic and SpaceX, and then at the trough of the 2022 drawdown told his limited partners he was shifting the whole portfolio construction toward singles and doubles instead of concentrated risk. What actually separates the two camps is whether you can survive being early, not whether you are right. [Source: "Finding the Next Figma, Wiz, & Stripe Before It's Obvious | Neil Mehta Interview"] [Source: "Finding The 1% of Stocks That Matter | Henry Ellenbogen Interview"] [Source: "Inside General Atlantic: How a $100B Growth Equity Firm Invests"] [Source: "How Graham Weaver Turns Recent Grads Into CEOs"] [Source: "Inside Dan Sundheim's Bets on Anthropic, OpenAI, and SpaceX"]

Does the founder scale, or does the trained operator? Brian Chesky's position is that founders are born but CEOs have to be trained, and that the training only works through hands-on immersion in detail: he came out of losing eighty percent of Airbnb's business in eight weeks convinced that delegation had disconnected him from his own company, and he still personally co-hiring-manages the top two hundred hires and spends two to three hours a day recruiting. Ben Horowitz built a16z on the same premise, requiring every investor at the firm to have founded or run a company. Against them, Graham Weaver made a deliberate and, in his word, brutal break around 2010: Alpine stopped backing existing founders entirely and now installs its own young, first-time CEO on essentially one hundred percent of platform deals, rebuilding its sourcing engine around that promise. Luca Ferrari goes furthest, saying he saw no correlation between the most talented founders he knew and which ones succeeded, so zero to one is mostly luck while operating a digital business exceptionally well is a repeatable skill, which is why Bending Spoons buys proven products outright rather than founding new ones. What decides it is whether the value still to be created requires another zero-to-one leap or the disciplined operation of a machine that already works. [Source: "How Brian Chesky Is Redesigning Airbnb for the AI Era"] [Source: "Why The Laws of Startup Physics Have Changed | Ben Horowitz Interview"] [Source: "How Graham Weaver Turns Recent Grads Into CEOs"] [Source: "The Playbook on Buying and Running Companies Forever"]

Is a moat a structure, or an experience? Gokul Rajaram names five structural moats and treats them as the whole question: ownership of a scarce asset, a control point over money or data flowing through the product, hardware that is costly to replace, membership in an essential workflow, and network effects. Without one of those, he argues, a startup gets outbuilt by whichever foundation-model company or in-house team decides to copy the workflow in a weekend. Against that, Neil Mehta ranks the emotional evidence above any formal moat framework: he coined jaw-dropping customer experience while investing in Coupang, where same-day delivery lifted cohort retention from the industry's thirties into the sixties and customer interviews captured mothers crying on camera because the diapers arrived before they ran out. Reed Hastings runs Netflix's content budget as a venture portfolio precisely because hit-picking never became a repeatable formula, with the studio's biggest recent hit reportedly its roughly thirtieth animated film. Sundheim splits the difference on the AI labs, framing them as Netflix-like in their fixed-cost model and Spotify-like in that the differentiation is personalization rather than raw quality. The corpus does not settle it; what it does show is that Rajaram's moats decide who survives a copy, and Mehta's experience decides who is worth copying in the first place. [Source: "He Built The Revenue Engines for Google, Facebook & Square"] [Source: "Finding the Next Figma, Wiz, & Stripe Before It's Obvious | Neil Mehta Interview"] [Source: "The Netflix Culture Code That Changed Entertainment Forever | Reed Hastings Interview"] [Source: "Inside Dan Sundheim's Bets on Anthropic, OpenAI, and SpaceX"]

Is the edge volume, or cost per unit? Sam Altman's answer to the worry that someone will distill and undercut OpenAI is volume rather than margin: so much future compute goes to selling inference that even a modest margin on what he calls trillions of dollars of revenue funds the expensive training runs, and he says the durable advantages are compute fleet size, network effects and workflow integration rather than any one model. Dara Khosrowshahi ran the same play operationally, blowing through Uber's entire annual AI compute budget in a single quarter and leaning further in rather than cutting. Against that, Gavin Baker reads the whole buildout as a cost-of-tokens war: Google trained Gemini 3 on chips he considers a generation behind Blackwell and still became the lowest-cost producer of tokens, which he believes let it run inference near a negative thirty percent margin specifically to starve capital-constrained rivals. Dylan Patel's evidence points the same way, that OpenAI deliberately kept GPT-5 close to GPT-4o in size and serving cost because the industry bottleneck is serving capacity and cost rather than model quality. If Baker is right, Altman's modest margin is not a choice he gets to keep making. [Source: "Sam Altman on AGI, Compute, and Human Agency"] [Source: "Uber CEO on AI, Autonomous Vehicles, and the Future of Transportation"] [Source: "GPUs, TPUs, & The Economics of AI Explained | Gavin Baker Interview"] [Source: "Inside the Trillion-Dollar AI Buildout | Dylan Patel Interview"]

The one line

Build the structure that lets you act well before you need to act, do the primary work yourself, and be honest about which of your two possible mistakes is the one you cannot come back from.