Summary

Sundheim’s structural map of the two markets: public equities are “the most competitive in the world” yet less efficient than before, while privates have fewer players who are “all doing the same thing. Whereas the public markets, there’s tons of people competing, but they’re all playing a different sport.” Right now he sees a moment-in-time opportunity in late-stage privates — “some of the largest companies in the world by market cap are private right now,” and D1’s investments in the AI labs give it the technology view every AI-affected public stock now requires.
The LLM debate has moved. The old bear case — “AI will be huge, so is air travel” (airlines’ returns went down to cost of capital) — is “more or less irrelevant”: likely Claude Code and OpenAI’s business are durable, gross margins are high, and only “four or five LLMs will be relevant in the long term.” The real question is capital intensity — “capital intensive to a degree that we’ve never seen before in the history of business” — with unknown returns on training spend and leverage that removes “the luxury of 2 or 3 years of things going slower” than expected.
Hyperscalers are a worse business going forward, a thesis he’s held about a year and grown more confident in — with a twist: growth accelerates (AWS, GCP) even as the model degrades, because the customer base concentrates into a handful of labs that treat clouds as “more of a financing mechanism” and will insource compute once free cash flow arrives in 5-10 years. “The LLMs are actually better at inference than the hyperscalers”; Meta already insourced; neoclouds may persist because they are better at running GPU clusters than traditional hyperscalers, with Nvidia seeking a diversified customer base.
From D1’s year-end letter: there were “basically no shorts” in AI before 2026, but now “there are going to be a lot of shorts, some longs — software is the first one.” His base case (explicitly “fairly low conviction”): software becomes a worse business but evolves like Walmart absorbing e-commerce, with systems of record protected longest — the LLMs told him they’re buying, not building, their ERP. Only a total stop in scaling laws makes AI overblown, “a really low probability assumption.”
The Anthropic bet was a repeat of the lesson he took from missing Amazon: the income statement showed “a sea of red” and the only tell was Bezos’s 1997 letter. Despite smart friends invoking Uber-vs-Lyft against backing a second player, he underwrote Dario’s essays — “Dario just did that better than almost any CEO I’ve seen since Bezos.” Sentiment has since ping-ponged: Anthropic went all-in on enterprise and coding and “is winning… kind of now the Uber,” while OpenAI tries everything at once — and he pushed OpenAI ~18 months ago: “you have to do ads.”
Starship redraws SpaceX’s TAM: full reusability drives launch costs down dramatically, satellite engineering surprised him to the upside, and “the telecom market globally is now the TAM” — within “months, a few years” Starlink will be “dramatically cheaper than any other form of delivering broadband.” The Rivian contrast (a similarly sized check) taught the meta-lesson: “the bad ones tend to be more obvious faster”; great private bets compound slowly.
His biggest worry: “we are on a collision course with China over semiconductors.” Taiwan makes “90-something percent” of leading-edge chips — like one country producing all the world’s oil — and a broken supply chain means an economy “on the order of a depression.” The US needs 10-20 years to replicate the supply chain, and “when dictators say something religiously, you should believe them” — likely Xi emphasizes Taiwan in every important speech.
The GameStop and 2022 drawdown marked a regime change: just after the May 2022 trough he overruled his president and faced four straight LP dinners with the message “we’re going to hit singles and doubles” — de-risked construction, same stock selection — because “emotionally I would not be able to go through this again.” Meanwhile passive, retail, quants, and short-term multi-managers make short-term moves “exaggerate the true change in intrinsic value”: for a fundamental investor with duration, “there’s just endless amounts of shorts.”

Deep dive

1. Two markets, one skill — but publics are a different sport

Sundheim’s 2026 snapshot: “a lot of interesting opportunities in late-stage privates,” a genuine moment in time — “some of the largest companies in the world by market cap are private right now,” and they’re “innovating in a way that’s going to change the world.” The constraint in privates is access, not analysis: investors rarely disagree that a company is excellent, but “that company has to want you to be an investor.”
His structural framing, worth keeping verbatim: publics are “the most competitive in the world” even though less efficient than before, whereas in privates “there’s fewer people competing, but they’re all doing the same thing. Whereas the public markets, there’s tons of people competing, but they’re all playing a different sport.” Privates also lack the economically irrational short-term actor.
The synergy between the books is the highest he’s seen: at D1’s founding ~25% of private work had public-side overlap; now, “because of AI… I’ve never seen greater synergy” — if you own stocks deeply impacted by AI (eventually “almost every public”), you need an opinion on where the technology is going, and investing in the labs supplies it.

2. The Bezos tell: backing Dario over the Uber-vs-Lyft objection

OpenAI at the $125 billion round “wasn’t contrarian at all” — if you believed in LLMs as a business model, “OpenAI was the one”; the debated question was whether to believe at all. Anthropic drew the classic pushback from people he considers very smart: Uber versus Lyft — “investing in the second player is not the path to glory.” His counter: at that stage it was “incredibly difficult to say who was going to be first and who was going to be second” among maybe five, six, seven players that could ultimately be important.
The deciding evidence was written, not modeled. His filter for what he missed in early Amazon: the income statement was “a sea of red” — the only tell was Bezos’s 1997 shareholder letter, its clarity “greater than almost any public CEO I dealt with.” Reading Dario’s essays: “Dario just did that better than almost any CEO I’ve seen since Bezos.” He weights “clarity of thought and the ability to communicate,” especially in writing — “rightly or wrongly.”
The same lens applies to leadership generally: “real passion,” a strong competitive streak, command of detail, “somebody who people want to work for” (Musk in the factory “is not all giggles,” but people want to learn from him). And he dissents from “the business matters more than the leader”: over 30 years, maybe — but “businesses are just people,” and on his 5-to-10-year horizon, people are more important, especially in technology.

3. The LLM debate has moved: commoditization is dead, capital intensity is the question

The original bear case was the airline analogy — “AI will be huge, so is air travel,” and undifferentiated airlines earned their cost of capital. D1 took the other side at “65/35, 70/30” confidence; the bet was skew — if the models proved moated, the outcome “would be huge.”
That debate is now “more or less irrelevant”: with likely Claude Code and even OpenAI’s business, “these are durable businesses.” Yes you can switch — “the same way you could switch AWS or Azure” — but it’s not worth it for most, and “the gross margins are quite high… not the margins that you see in a commoditized industry.” The field is set: “four or five LLMs that will be relevant in the long term,” locked in not by talent scarcity but by capital requirements and the snowball of capital → compute → better researchers.
The live debate: these are “capital intensive to a degree that we’ve never seen before in the history of business,” and unlike a factory, you’re training models without knowing what the asset will sell. That intensity “introduces financial leverage and operating leverage” — “you don’t have the luxury of 2 or 3 years of things going slower than you otherwise would expect.” His scenario weighting is telling: “equally likely if not more likely” than a genuine bust is that the return comes, just slower — enterprise adoption lagging the spend. Scaling laws, return on capital, and adoption speed are the questions.

4. His pitch to the labs: you’re Netflix crossed with Spotify — so focus, and do ads

The framing he gave LLM executives directly: “your business is some kind of combination between Netflix and Spotify.” The Netflix half — massive upfront spend on a fixed asset, then sales at extremely high incremental margins, feeding a flywheel of revenue → more content → more revenue until competing is nearly impossible. Key difference: Netflix’s content was differentiated, whereas “the models are more similar than they are different.”
Hence the Spotify half: music is theoretically a pure commodity, yet personalization gives Spotify pricing power. Same for models — “the more that these models know about you, how you live your life, your health… it becomes very sticky.”
On focus versus do-everything: TAM “is certainly not the problem” — the trade-off is spreading the fixed asset across more markets versus being best at nothing. “I rarely have seen any company succeed trying to go after multiple end markets at the same time” — even Amazon entered enterprise about 7 years later, after it had gone public. OpenAI is doing everything (Apple hardware, robotics, enterprise, consumer, science); Anthropic went all-in on enterprise after consumer traction faded, took a market-leading position in coding, and “generally sentiment is that Anthropic is winning… kind of now the Uber.” He’d “err on the side of focus” — and expects sentiment to keep swinging both ways.
The ads push: about a year and a half ago he told OpenAI, “you have to do ads,” knowing Silicon Valley is “allergic” to it. Even Netflix — where “Reed” would have been like, “you are out of your mind” — ultimately caved. His logic: “you can’t really compete against companies that are using ads if you’re not… if you’re ultimately going to do it, you might as well start earlier,” because the culture takes time to build.

5. Hyperscalers: growth will accelerate — and the business is getting worse

The thesis, held for about a year and firming: “I am more confident in the thesis that the hyperscalers are a worse business model going forward.” The twist is that it comes with accelerating growth — AWS and GCP (maybe not Azure) speed up as their lab customers grow enormously. The rot: the customer base is shifting from every corporation in the world (fragmentation plus unmatchable economies of scale) to concentration in four or five LLM companies.
The insourcing logic: today the cash-burning labs treat hyperscalers “as more of a financing mechanism” — big balance sheets, not superior builders. “Building CPU clusters is different than building GPU clusters,” and “the LLMs are actually better at inference than the hyperscalers.” When the labs turn massively free-cash-flow positive in the next 5-10 years, “they are likely to insource the compute” — precedent: “Meta is not a hyperscaler, but they insourced all their compute.” Patrick adds the news that Anthropic is considering securing 10 gigawatts of its own power.
On neoclouds, consensus said pure overflow capacity, “dead as soon as Microsoft got their” GPUs. His view: “I certainly would not make the case that they are fantastic businesses, but I don’t think they’re going away” — they run GPU clusters better than traditional hyperscalers, and Nvidia’s balance sheet wants “the customer base diversified.” Ten-year net: fast growth, but margins “my guess is will be challenged” — more capital-intensive workloads, more concentrated customers.

6. Phase two: “there are going to be a lot of shorts” — software first

From D1’s year-end letter: before 2026 “there was basically no shorts” in AI — shorting on an AI thesis made little money. Now: “there are going to be a lot of shorts, some longs, because of AI — and software is the first one.” The trigger was likely Claude Code entering “the zeitgeist” — people see “I created a CRM system in a day” on Twitter and conclude “this isn’t good.” He notes the market “tends to swing to extremes.”
His base case — flagged as “fairly low conviction,” because “everything about AI’s impact on the economy is inherently low conviction” — is the Walmart analogy: software becomes “a worse business model going forward” but evolves, the way Walmart absorbed e-commerce through painful investment and margin hits. Systems of record hold longest: he asked the LLMs “are you designing your own ERP system?” — answer: “no, we’re buying a new ERP system from this company.” “At least you’re protected for a few years.” But no vendor can “just sit back and say we’re a system of record, we’ll be fine.”
Could it all be overblown? Only if “scaling laws just totally stopped” — and even then he’d expect ~3 years of adoption-driven change in the real economy. “Betting that scaling laws are going to stop is a really low probability assumption… everything suggests the opposite.” The closing frame: “everyone is likely underestimating how much these models are going to improve,” and to reason about it “you have to almost not think like an investor. You have to think like somebody who’s into science fiction.”

7. GameStop to the trough: singles and doubles, and why he keeps score

“As an investor, that was about as bad as it gets” — from “everyone thinks we walk on water” to “everyone thinks we’re going out of business” (“that was just nonsense,” he says; they never came close). It was lonely — maybe “one or two other people” going through the same thing; he reread Ken Griffin’s 2008 interviews. The hard lesson: “it’s impossible to disprove the negative narrative in the short term” — even hitting “the ball out of the park for three months” just reads as volatile and crazy. Likely Bill Ackman’s counsel stuck: every day, “try to do something that makes things a little bit better.”
The pivotal moment: semiannual LP dinners scheduled for June 3, 2022 — days after the drawdown’s trough. Jeremy, D1’s president: “We can’t do these dinners. This is going to be a bloodbath.” Sundheim overruled him — “this is the most important time to go out there and speak to our investors.” The message: same stock selection, de-risked portfolio construction — “we’re going to hit singles and doubles,” even though “usually the best time to take a ton of risk is when you’ve lost a lot of money.” The honest reason: “emotionally I would not be able to go through this again.”
On redeemers: “I don’t harbor any ill will… when you deliver poor returns, capital will leave. Capital follows returns.” His gratitude to those who stayed outweighs any resentment — “it’s pretty asymmetric.”
Why keep going: “money to me is a scorecard, and I want to have the best score.” He doesn’t want D1 to have enterprise value: “our business is horrible. It cash flows really well. It has no terminal value.” His line to portfolio CEOs: “You have no cash flows and tons of terminal value. I have tons of cash flows, no terminal value. So we’re good together.”

8. Rivian and SpaceX: the bad ones show fast, the great ones compound slowly

The story of two similarly sized checks. Rivian’s thesis: EVs would dominate autos, and a software-defined car was “the equivalent of the iPhone versus Motorola and Nokia” — incumbents couldn’t make the jump. What broke it: “ultimately, autos are a bad business” — software or hardware, capital-intensive and brutal to scale — and manufacturing delays while burning cash kept Rivian from the scale that decides EVs (one reason Tesla won). The return “wasn’t what we planned for.”
The meta-lesson: “the bad ones tend to be more obvious faster. The great private tech investments are sometimes slower to prove how great they are” — great founders’ decisions compound quietly over years.
SpaceX was underwritten on skew: launch alone was clearly a very good business, the engineering “insane,” cash burn small at entry — “I just knew the skew was very good.” Now Starship (“you caught a skyscraper with chopsticks,” as Patrick puts it) means the cost of launching everything drops dramatically, and the satellite engineering “surprised me to the upside.” Consequence: “the telecom market globally is now the TAM” — boats, planes, homes without cable — and within “months, a few years” Starlink will be “dramatically cheaper than any other form of delivering broadband.”
His best-business aesthetic runs straight through it: durable low-cost producers with a cost-volume flywheel — SpaceX in launch, Costco in groceries, Amazon in e-commerce. Very few monopolies exist, and when they do “they tend to get lazy.”

9. A born short-seller in a market getting less efficient

“My wife begs me all the time to stop shorting stocks… it’s a bad business.” But almost nobody does it anymore — most participants aren’t fundamental, and story stocks proliferate via social media and Robinhood: “there’s just endless amounts of shorts if you have duration and if you take a fundamental view.”
Why efficiency fell: mutual funds and long/short funds gave way to passives, retail, quants, and multi-managers who are fundamental but “by necessity short-term oriented” — so short-term moves “exaggerate the true change in intrinsic value of the company.” Geographically, Europe is the most inefficient (though “economically stagnated”), Japan and Korea “pretty inefficient” too — and with Japan likely re-emerging as a military power, he sees hard-asset engineering companies in Germany, Korea, and Japan that were not well positioned during the last 20 years of digital companies.
The origin story, 2002: a private-equity analyst at likely Bear Stearns posting on Value Investors Club ($5,000 weekly for the best idea). A hedge fund interview case study — Orthodontic Centers of America — became a fraud discovery: “nothing reconciled,” until it hit him they were “capitalizing expenses that should have been expensed — the simplest form of accounting fraud.” He posted the six-page writeup anonymously before his follow-up interview; the stock cratered 20-30% and people at likely T. Rowe Price and Fidelity started calling him at Bear Stearns. He skipped that fund (healthcare wasn’t his interest) — but the writeup circulating among hedge funds got him hired.
The Viking arc: from banks analyst under Tom Purcell to CIO, by 2016 running ~55% of Viking’s capital — “an abnormal percentage” that he concluded served neither Andreas nor LPs. At 40, “relatively late in life,” with little left to achieve there and energy that wouldn’t last forever, he founded D1.

10. The tail risk: a collision course with China over semiconductors

“The thing that troubles me the most”: Taiwan produces “90-something percent” of the most advanced semiconductors, “and everything we use is semiconductors.” His analogy: as if 50 years ago a single country produced all the oil — “we went to war over oil even though you could get it all over the world.” If that fragile supply chain is destroyed or disintermediated, “we would have an incredibly bad economy on the order of a depression.”
There is no happy equilibrium: “there is no scenario I can think of where everybody’s happy… somebody’s going to be unhappy — either because the economy collapses or because their sovereignty is handed over.” His hoped path: replicate the supply chain in the US over 10-20 years while China sees an eventual route to reintegrating Taiwan — with the uncomfortable corollary that US self-sufficiency makes America “probably less likely to defend Taiwan.”
Take dictators literally: “when dictators say something religiously, you should believe them” — Putin acted on Soviet-glory rhetoric “as soon as he had the capabilities,” and likely Xi emphasizes Taiwan in every important speech. AI “just raises the stakes so much.”
The offset is his macro optimism: AI is “the ultimate productivity tool,” enabling growth with disinflation — “nirvana for markets” — and possibly curing deficits. His unease is civilizational, not economic: “we’re no longer going to be the most intelligent beings on the planet,” and against Dario’s UBI sketch, “I just don’t think humans are wired to just collect a check” — humans need work, relationships, and achievement. To his son’s generation he channels Musk: “better to go through life being an optimist and be proved wrong than a pessimist and be proved right.”

The 2028 Global Intelligence Crisis: Will AI Lead To A Market Crash?Insights from Coatue's Growth Investor Lucas Swisher