Dan Sundheim - The Art of Public and Private Market Investing - [Invest Like the Best, EP.460]
What this covers
Dan Sundheim, founder and CIO of D1 Capital Partners, talks with the host about managing over $30B across public and private markets, including stakes in SpaceX, OpenAI and Anthropic. The conversation alternates between his own career and the mechanics of the work. He recounts the Orthodontic Centers of America short that helped him land his first job: the company capitalized expenses that should have been expensed, and unit economics built from management's claims did not reconcile with the filings. He also explains why private markets are a contest for access rather than analysis alone.
The AI material is the longest stretch. Sundheim describes backing Anthropic at roughly 65-35 to 70-30 odds, framing it as a question of skew. He covers the early airline comparison, his expectation that four or five providers will matter, and whether scaling laws and adoption speed keep returns on training capital attractive. He also doubts that companies can pursue many end markets at once. On leadership, he values clarity in writing, says Dario Amodei showed it more than almost any CEO since Jeff Bezos, and disagrees with Buffett that business quality outweighs the people running it.
The last stretch turns to risk and consequence. Sundheim discusses de-risking portfolios after the 2022 drawdown and why a damaged reputation takes years to repair. He recalls the stress of GameStop in early 2021. His contested forecast is that a China–Taiwan collision over semiconductors could produce a Great Depression-scale collapse. He also argues that a society living on AI-funded government checks would be a poor one, and that young people should follow genuine interests rather than fear obsolescence.
Sundheim argues that AI's economic value will concentrate in focused, capital-intensive model businesses whose personalization creates stickiness, that hyperscalers face a worse business model as AI labs eventually insource compute, and that public markets stay inefficient because most participants trade on short-term signals, leaving room for long-horizon intrinsic-value investors.
- Frontier LLM businesses are fixed-asset, capital-intensive models that resemble Netflix in spreading cost over users, and personalization rather than model quality drives durable pricing power, as with Spotify.
- Hyperscalers lose their fragmented-customer advantage as AI labs concentrate demand, and labs will likely insource compute once free cash flow is generated.
- Public market inefficiency persists because passives, retail, and quants focus on short-term moves that exaggerate changes in intrinsic value.
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
When a manager delivers poor returns, capital leaves, and Dan accepts that redemptions were partly deserved, while the asymmetry means loyal investors matter more to him than departing ones.
“I ultimately think that when you screw up in business, capital follows returns. And when you deliver poor returns, capital will leave.”
Once investors look past short-term events to business value, long-term cash flows and Porter's Five Forces attributes, the competitive set becomes thin, so the edge lies in medium-term analysis that short-term-focused players ignore.
“what are the Porter's Five Forces attributes of the company – that's when I think that the competitive set gets pretty thin.”
The core question for LLM businesses is not commoditization but that they are capital-intensive to a degree never seen before, and the return on training spending is unknown, unlike a factory whose output is predictable.
“these are extremely capital-intensive businesses, capital-intensive to a degree that we've never seen before in the history of business”
Taiwan produces roughly 90 percent of the most advanced semiconductors, making the supply chain fragile and easy to destroy, so a collision with China over semiconductors could cause an economic collapse on the scale of the Great Depression if not resolved.
“Taiwan produces 90-something percent of the most advanced semiconductors, and everything we use is semiconductors.”
Dan looks for passion, competitiveness, and deep engagement in leaders, and disagrees with Buffett that business quality outweighs leadership: over a five-to-ten-year horizon, especially in technology, the people running a business are more important than the business itself.
“I kind of disagree with that. I think if you look over 30 years, sure. But over any medium-term period of time, businesses are just people.”
The early skeptical view of LLM businesses compared them to airlines: AI would be huge, but airlines were a poor business because undifferentiated competition pushes returns down to the cost of capital.
“AI will be huge. So was air travel. Airlines were not a good business.”
The best businesses are usually low-cost producers of durable products, where low cost drives volume that drives lower cost, and cost advantages that are substantial and impenetrable, like SpaceX in launch or Costco in groceries, are more powerful than most monopoly positions.
“the best businesses are usually low-cost producers of something that's very durable.”
Humans are wired to create relationships, create value, work, and coordinate, so a society where people live off government checks from an AI-driven boom would not be a good society, even if such checks are offered as a remedy.
“I just don't think humans are wired to just collect a check and go around and play sports all day.”
Dan is more confident that the hyperscalers AWS, Azure, and GCP have a worse business model going forward, though he expects their growth to accelerate for a while because AI labs are growing quickly and becoming a larger share of their revenue.
“I am more confident in the thesis that the hyperscalers are a worse business model going forward.”
Spreading a fixed AI asset across many end markets lowers cost per market, but companies rarely succeed pursuing multiple end markets at once because culture and the A team must focus on one thing, so the trade-off between scale and excellence is unresolved.
“I rarely have seen any company succeed trying to go after multiple end markets at the same time.”
Four or five LLM providers will be relevant in the long term because capital requirements are too large for new entrants, and scale creates a compounding advantage in compute and researcher talent.
“you probably have four or five LLMs that will be relevant in the long term. I don't see that changing.”
If AI scaling laws hold, AI is the ultimate productivity tool, allowing economic growth alongside disinflation, which Dan calls nirvana for markets and which could help cure government deficits and raise living standards broadly.
“what productivity does is it allows you to grow while having disinflation, which is like nirvana for markets”
Public markets have become less efficient because passive funds, retail investors, and short-term-oriented quant and multi-manager funds dominate trading, whereas mutual funds and long-short hedge funds once emphasized long-term intrinsic value.
“Now, it is a lot of passives, a lot of retail investors. People who are making investment decisions that are not based upon long-term considerations of intrinsic value.”
Orthodontic Centers of America's accounting problem was the simplest form of fraud: capitalizing expenses that should have been expensed, and building unit economics from management's claims exposed that the reported numbers did not reconcile with the filings.
“what they were doing was the simplest form of accounting fraud, which is just capitalizing expenses that should have been expensed, in a big way.”
The key questions for LLM economics are whether scaling laws keep returns on training capital attractive enough to attract more capital, and how quickly AI is adopted across the economy, with outcomes ranging from wasted capital to slower-than-expected but eventual returns.
“I think the scaling laws, the returns on capital, and the speed at which these tools and AI is adopted throughout the economy are the questions.”
Like Spotify, which turns a commodity music catalog into a premium personalized service, LLM providers can gain pricing power through personalization: accumulated knowledge of a user's life, health, and preferences makes the product sticky and hard to leave.
“the more that these models know about you, how you live your life, your health, all the things that are important to you – you build up this data and history, and it becomes very sticky.”
Dan's early confidence that LLM businesses would earn returns was roughly 65-35 to 70-30, so the decision was about the skew: if the business models proved moated, the payoff would be huge.
“At that point, it was more about the skew.”
Even when many private investors agree a company is excellent, access is limited because the company itself must want you as an investor, so competition in private markets is about gaining access rather than analysis alone.
“But that company has to want you to be an investor in the company.”
Dan places heavy weight on a CEO's clarity of thought and ability to communicate in writing, because writing forces a leader to work through the full plan and express it so others understand, and Dario Amodei showed this more than almost any CEO since Bezos.
“taking the time to write something down – you actually really have to go through everything you plan to do”
After the 2022 drawdown, Dan told LPs at the semi-annual dinners that D1 would keep the same stock selection but construct portfolios with much less risk, aiming for singles and doubles rather than home runs, even though this slowed recovery to the high-water mark.
“We're gonna hit singles and doubles. It might take us longer to get back to the high watermark”
A manager cannot disprove a negative reputation quickly, because a short burst of outperformance looks volatile and steady performance can cause doubters to give up, so restoring reputation takes years and is achieved by consistency.
“So it's impossible to disprove the negative narrative in the short term. It takes a lot of time. Years.”
Private markets are less competitive because fewer investors look at each situation and less capital is deployed, and private investors all attempt the same task of estimating intrinsic value, whereas public markets have many more competitors who are playing a different game focused on short-term outcomes.
“So there's fewer people competing, but they're all doing the same thing. Whereas in the public markets, there's tons of people competing, but they're all playing a different sport.”
Telling young people that AI will make their careers obsolete is self-defeating, so children should pursue genuine interests and curiosity, because interest drives skill and no job is protected unless one is actually interested in it.
“unless someone is really interested in something, they're not gonna be good at it.”
Once AI labs generate large free cash flow, probably within five to ten years, they are likely to bring compute in-house, since AI will become the majority of hyperscaler workloads and it will make economic sense to insource.
“I think that they're likely to insource the compute”
While preparing a Viking-era hedge fund case study on Orthodontic Centers of America, Dan found the company's unit economics did not reconcile with its filings, pointing to capitalizing expenses that should have been expensed, and his anonymous Value Investors Club post helped push the stock down by about half.
“what they were doing was the simplest form of accounting fraud, which is just capitalizing expenses that should have been expensed, in a big way.”
Because LLM businesses are capital-intensive, they carry financial and operating leverage far beyond normal businesses, so they cannot absorb two or three years of slower-than-expected returns or adoption.
“you don't have the luxury of two or three years of things going slower than you otherwise would expect”
Systems of record such as ERP and CRM are difficult to displace because running a whole business on them takes a long time to replace with AI-generated software, and LLM companies themselves buy ERP systems rather than building their own.
“they said, 'No, we're buying a new ERP system from this company.'”
Even if scaling laws stopped and models did not improve, there would still be about three years of people learning to integrate AI into their work and companies, producing profound economic change, though returns for the businesses would be worse.
“even if scaling laws stopped, even if these models got no better, I think you probably have three years of people learning how to incorporate AI”
In private tech investing, bad investments become obvious faster, while great investments are slower to prove themselves because founders compound many decisions over time.
“The bad ones tend to be more obvious faster. The great private tech investments, I think, are sometimes slower to prove how great they are.”
A person with a volatile temperament can train themselves to handle market swings, and some generationally great hedge fund managers started out volatile, but they must prevent emotion from influencing trading decisions.
“But you have to be able to not let that emotion influence your trading.”
About 25% of the time since founding, looking at private companies has produced synergy with public investing, but because of AI and rapid private-market innovation, the synergies are now greater than Dan has ever seen, since understanding the technology trajectory is necessary to judge public companies that AI will affect.
“the synergies are just greater than I've ever seen before”
Anonymous posts on Value Investors Club can move stocks, as Sundheim's short write-up caused the target to fall 20 to 30 percent the next day after a few hours of online discussion, and mutual fund holders then called him about the stock.
“Next day, stock starts to crater. Stock's down like 20, 30%.”
Europe is economically stagnant and likely unattractive except for fundamental stock pickers, while Japan, Korea, and Germany are underappreciated because their companies have strong physical assets and engineering after two decades of digital-company favoritism.
“Japan and Korea were not well positioned for the last 20 years, because it was just digital companies.”
If Starship achieves full reusability and scale, the cost of launching payloads will fall by about 97 percent or more, which makes SpaceX's satellite broadband potentially cheaper than any other broadband delivery and expands its addressable market to global telecom.
“the cost of launching everything goes down dramatically. 97 or whatever percent.”
The software sell-off reflects the first phase of AI's effect on the broader economy, triggered by Claude Code's visibility, and the market tends to swing to extremes, so software will likely become a worse business model but not vanish.
“So there's gonna be phases of this, where the first phase is software.”
Enduring GameStop was the most difficult period of Sundheim's career, and he learned that he would not quit and that rebuilding requires reversing a negative narrative over years rather than in one good quarter.
“I was like, 'Am I really gonna be the guy who quits the first time?'”
Dan's group chat for founders keeps CEOs connected to his firm, because founders are lonely and value being around others who understand them, and it lets companies share insights, so even after months without contact, CEOs feel close to Dan.
“even if I haven't spoken to a CEO in like three months and I call them, it's almost like they feel like they talk to me every day.”
Amazon's early income statement showed only losses, and the only telltale sign of its potential was Jeff Bezos's 1997 shareholder letter, which showed clarity of thought about value creation that Dan says would have been a very profitable signal if read.
“The only telltale sign was reading Jeff Bezos' 1997 shareholder letter”
LLM companies resemble Netflix: they spend heavily up front to train models, then sell access at very high incremental margins, and must sell as widely as possible to fund the next model, creating a flywheel that makes competition hard for early movers.
“you invest heavily, you get the capital to invest heavily, you get the revenues, you spread it out over an increasing number of people”
Unlike Netflix's differentiated content, frontier LLMs are similar to each other because expertise and innovation diffuse quickly, so the leader changes over time and no model is decisively better.
“The difference – if you were to say like, what is an important difference of Netflix versus these models – is Netflix's content was differentiated.”
Hyperscalers historically benefited from a fragmented customer base of every corporation and economies of scale, but LLM demand is likely concentrated in four or five companies, which changes that dynamic.
“The problem going forward is that I think that economically it's highly unlikely that LLMs are not very concentrated in the hands of four or five companies.”
Dan's Rivian investment rested on EVs being software-defined products like the iPhone, but autos are a bad, capital-intensive business, and manufacturing delays while burning cash meant Rivian failed to scale fast enough to earn the planned return.
“Ultimately, autos are a bad business.”
Dan considers his hedge fund business horrible in enterprise-value terms because it generates large cash flows but has no terminal value, whereas startups he invests in have no cash flow but large terminal value, so the two partner well.
“Objectively, our business is horrible. It's amazing cash flows but has no terminal value.”
Publishing a well-researched short idea anonymously on Value Investors Club moved the stock sharply and led hedge funds and institutions to contact him, which helped him get his first job, illustrating the power of early online research posting.
“That's how I got my job.”
Any view of AI's economic impact is inherently low conviction because everyone underestimates how much models will improve, so investors must think like science fiction writers rather than conventional investors.
“you have to almost not think like an investor, you have to think like somebody who's into science fiction.”
Dan pushed OpenAI about 18 months ago to adopt advertising, because companies that refuse ads eventually face competitive disadvantage and need time to build an advertising culture, and he believes it would have been better to start sooner.
“If you're ultimately going to do it, you might as well start earlier because you have to build a culture around it.”
The first AI disruption will hit software, but moats in robotics and traditional industrial companies are harder to assess, and countries that arbitrage labor will face new questions relative to developed economies, so investors should map which companies have moats protected from digital LLMs.
“which companies do you think will have a moat in most circumstances”
Dan bought SpaceX at a multiple of revenue with very little cash burn, because the company had achieved what he regarded as the most impressive engineering feat he had seen, and the skew was very favorable even without knowing future outcomes.
“if I could buy a company that had achieved the most amazing engineering feat I'd ever seen, at some multiple of revenue with very little cash burn”
Most of the time, short-term stock price moves exaggerate the real change in a company's intrinsic value, creating mispricing that a medium-term investor can exploit, even though Dan does not consider himself a five-to-ten-year buy-and-hold investor.
“The majority of the time, the moves you see in the short term exaggerate the true change in intrinsic value of the company”
Sundheim's calm temperament comes from confidence in his analysis: when he feels confident in the research he is balanced, and volatile personalities can train themselves to manage emotion, as some generationally great hedge fund managers did.
“when I feel confident in the analysis, I generally am pretty balanced.”
When Dan first invested in Anthropic, many smart people argued it was the Lyft to OpenAI's Uber and that investing in the second player was a poor path, but at that stage it was very hard to know which company would be first and which second.
“it was incredibly difficult at that stage to say who was going to be first and who was going to be second”
Building GPU clusters differs from building CPU clusters, and running inference on GPUs differs from CPU workloads, so LLM labs are better than traditional hyperscalers at inference operations.
“I think the LLMs are actually better at inference than the hyperscalers.”
Neoclouds are not just overflow capacity that will vanish once hyperscalers get GPUs, because they are better at operating GPU clusters and Nvidia wants a diversified customer base, though they are not fantastic businesses.
“I don't think they're going away like people thought, because I think they're better at running GPU clusters”
Replicating the semiconductor supply chain in the US would take 10 to 20 years, and during that window China must avoid aggressive action against Taiwan, after which a path could exist for China to reintegrate Taiwan, though US defense commitment would weaken.
“We need 10 to 20 years to replicate this supply chain.”
The thesis that LLM APIs would be commoditized into a race to the bottom is largely irrelevant, because LLM businesses are durable: switching is possible but not worth the cost for most customers, similar to switching cloud providers, and models are sufficiently differentiated.
“that debate is more or less irrelevant, because as you've just seen with Claude Code and even OpenAI's API business, these are durable businesses.”
Software companies cannot assume a system-of-record position will protect them; they must integrate AI into their business model, as Walmart did with e-commerce, even though the transition is painful for years.
“You're gonna have to integrate AI and find ways the same way Walmart integrated e-commerce into their business model.”
When dictators speak religiously about a goal, such as Xi's emphasis on Taiwan or Putin's Soviet nostalgia, one should believe them, because Putin acted once capable, so Taiwan rhetoric should be taken seriously.
“Usually when dictators say something and they say it religiously, you should believe them.”
Very few investors actively short stocks on fundamentals, since most market participants invest on stories or social media, so fundamental shorting faces endless opportunity for those with duration and a fundamental view.
“Very few people are doing it.”
OpenAI is pursuing everything at once, including Apple hardware, robotics, enterprise, consumer, and science, while Anthropic dropped consumer after lack of traction and went all-in on enterprise and coding, and Dan would err on the side of focus.
“I would probably err on the side of focus”
Netflix publicly rejected advertising for years, even 15 years ago, but ultimately adopted it, which suggests that companies that vow never to use ads often change course.
“Even Reed would've been like, 'You are out of your mind. We would never do that.' Ultimately, they did it.”
Studying history of dictators and failed systems is useful because patterns repeat: communism without dictatorship tends to fail as people reject it, so it persists only with a dictator benefiting from it, and history contains more things going wrong than right.
“communism without dictatorship doesn't work, because eventually people realize it's not good and so then they want to change”
Economic growth benefits everyone from hedge fund managers to lower-wage workers, and if a country does not grow, living standards cannot improve, so Dan is bullish on growth driven by AI productivity.
“If a country's not growing, it's hard to have a better standard of living.”
Most investors who love investing start with a deep value bent, as early pioneers like Ben Graham and Warren Buffett were deep value investors, and Sundheim's own initial bent was shaped by working under Tom Purcell as a bank analyst.
“I think most investors who love investing start out with a deep value bent.”
Even Amazon, a consumer company that entered enterprise, did so about seven years later, even after going public, which illustrates how hard it is to pursue multiple end markets simultaneously.
“Even Amazon, which you'd say is an example of a consumer company that got into enterprise. They got into it like, seven years later, even after they went public.”
Starting as a banks analyst and moving sector to sector at Viking, learning different value drivers in TMT, healthcare, and industrials, gave Dan exposure to nearly every industry and taught him that value creation differs by sector.
“what created value in TMT was different than what might create value in industrials or healthcare”
Art made by people reflects emotion and the moment of its creation, so as AI automates and produces abundant content, human-made art may become more valuable as a record of human experience.
“art created by people reflects emotion, and oftentimes what's happening in that moment in the world”
The kindest thing anyone did for Dan was his then-girlfriend's tearful joy when he told her he had a job at Bear Stearns, which made him realize how much she cared and motivated him to become a better partner.
“And immediately I walked out and I was like, 'I'm gonna marry that girl and I'm gonna be a better boyfriend slash husband going forward.'”
Humanity has never encountered a change like the coming one, in which humans will no longer be the most intelligent beings on the planet, and the implications are unclear and potentially negative.
“We are no longer going to be the most intelligent beings on the planet.”
Bill Ackman advised making things a little better each day, and Sundheim notes that a manager who hits home runs for three months looks volatile, while slow methodical improvement can still lose investors who give up before results appear.
“Bill Ackman was like, 'Look, every day try to do something that makes things a little bit better.'”
Public markets are the most competitive in the world for generating returns from company analysis, even though they have become less efficient than in the past, because more people in more places examine information about companies.
“I'd say the public markets are the most competitive in the world. Even though they are less efficient than they were before”
Sundheim hires long-time friends only if they are highly competent, because the job's bar is extremely high, and the value of relationships formed before he had money is that such friends are not nice to him for what he can provide.
“they have to be really competent do to the job. This is a very intense job. The bar is extremely high”
For Dan, money is a scorecard that measures whether he is best at investing; he does not seek to build enterprise value because that would require becoming a CEO of hundreds of employees, which would reduce time spent investing.
“Money to me is the scorecard. I want to have the best score.”
Monopolies are rare, and when they exist they tend to become complacent, which lowers their returns compared with cost-advantaged businesses that keep serving customers well.
“Very few monopolies, and when they are a monopoly, usually what happens is they tend to get lazy and the returns aren't as good.”
Dan views investing as more art than science because the core financial mechanics such as DCF are simple and learnable, while judgment about businesses, humanities, and aesthetics drives results.
“That is why I perhaps look at my job as more art than science. The science is very simple.”
OpenAI is pursuing a strategy of doing everything at once across Apple hardware, robotics, enterprise, consumer and science, which is hard to pull off, though the smartest people in the world work there so it may succeed.
“I think OpenAI is taking the strategy of, 'Let's do everything.'”
The GameStop period was about as bad as an investor's experience can get, moving D1 from top performer to a firm seen as going out of business, and Dan says his resilience was tested because he had not faced severe adversity in his career before.
“as an investor, I'd say that was about as bad as it gets.”
The gross margins of LLM companies are quite high, which is inconsistent with commoditized industries, so these businesses are not commodities.
“The gross margins are quite high.”
Hyperscaler margins will face pressure over the next decade because AI is more capital-intensive than traditional cloud workloads and the customer base is becoming more concentrated, though growth may remain fast.
“I think the margins, my guess is, will be challenged”
Shifts in global politics, including Japan probably becoming a military power again, create investable themes in defense, and Asia contains more interesting developments than Europe.
“for the first time Japan's probably going to become a military power at some point in the future again.”
Since AI will eventually affect almost every public company, investors should form a view on where the technology is and where it is going, and owning private AI companies provides that perspective.
“if you're going to take a view on public companies that are deeply impacted by AI – which eventually will be almost every public company – you should have an opinion”
The total addressable market is not the problem for LLM companies, because the models can be applied across consumer assistance, physics, drug discovery and enterprise efficiency.
“The TAM is certainly not the problem.”
Some of the largest companies in the world by market cap are private right now, and they are innovating in ways that will change the world, which makes late-stage private investing especially interesting at this moment.
“Some of the largest companies in the world by market cap are private right now.”
Before 2026 there were essentially no profitable shorts tied to AI, so D1's year-end letter predicted there would be many shorts as well as some longs, a view that became one of the firm's main public-market theses.
“There haven't been any shorts in AI. There were like, basically no shorts prior to 2026, really.”
At Viking, Dan moved from banks into healthcare, industrials, and TMT, and by 2016 managed over half of firm capital, but he left to start D1 at age 40 because Viking's risk structure capped his growth and he had the energy then.
“By 2016, I was managing just over half of Viking's capital.”
Money is the scorecard for Sundheim, and his competitive drive comes from wanting to deliver measurable, tangible results in the arena rather than running a family office, which would leave him bored.
“Money to me is the scorecard. I want to have the best score.”
At the time of Anthropic's early investment, the LLM companies were considered one of roughly five, six, or seven players that could ultimately be important, so the models were not highly differentiated at that point.
“probably at that point – five, six, seven players that could ultimately be important.”
Anthropic tried consumer early, found it lacked traction, and went all in on enterprise, where it has succeeded in coding and taken a market-leading position, so sentiment now treats it as the Uber.
“They tried consumer early on, but it came clear they didn't have traction, so then they just went all in on enterprise.”
The LLM companies have taken different lanes: OpenAI has excelled at consumer with good enterprise traction, while Anthropic has been very successful in coding, which Dan sees as evidence of real differentiation.
“OpenAI has been great at consumer and has had good traction in enterprise. Anthropic has been incredibly successful at coding.”
The initial investment in OpenAI at a $125 billion round was not especially contrarian, because if you wanted exposure to LLMs as a business, OpenAI was the leading choice, though whether LLMs were a viable business at all was heavily debated.
“We invested originally in the $125 billion round.”