Bill Gurley & Michael Mauboussin - Putting Theory into Practice - [Invest Like the Best, EP.370]
What this covers
Two longtime friends appear together for the first time: Bill Gurley, a general partner at Benchmark, and Michael Mauboussin of Counterpoint Global. The conversation opens on AI and LLMs, with Mauboussin noting that companies first bolt new technology onto existing workflows and only later rebuild around it, as happened with the Internet. It then settles into Increasing Returns to Scale. Mauboussin separates network effects, learning curves, and old-style scale economies. He cites Wright's law, under which each doubling of cumulative output cuts unit cost by about 20 percent, and says combining demand-side and supply-side advantages produces a "Lollapalooza" effect. Gurley uses Uber's driver-and-rider loop as his example and warns that competition in these games resembles a dollar auction, where price wars become hard to exit.
The range is wide. They cover intangible assets, Paul Romer's partly excludable nonrival goods, and innovation as recombination of existing building blocks. Gurley praises open source and also describes it as a defensive move, as with Android and Kubernetes. On markets, Mauboussin offers data: about 2% of public companies produced roughly 90% of net wealth created since 1926, and 55% to 60% of roughly 30,000 venture investments lose money. He also reports that CFO hurdle rates barely track interest rates, and that firms invested less during the 2009–2021 low-rate era. Gurley adds a hardware view: a roughly $50 million first tape-out and dependence on TSMC's schedule make semiconductor startups far harder to launch than software ones.
The last stretch turns contentious. Drawing on George Stigler, Gurley argues regulation drifts toward serving incumbents, and Mauboussin insists that public-interest regulation and capture are different things that get lumped together. Gurley also makes a case for writing and teaching as ways to sharpen thinking.
Gurley and Mauboussin argue that true increasing returns to scale are rare but dominant value drivers, that AI's gains will be captured mainly by incumbents with scale and capital, and that regulatory capture and bureaucracy are eroding the competitive dynamism that produces new winners.
- Increasing returns (network effects, learning curves, recombination of ideas) are rare, yet they generate outsized value, as in Uber's supply-demand loop and the Magnificent Seven's economic profit share.
- Regulation tends to become a friend of incumbents over time, and the longest-regulated industries (healthcare, finance, telecom) show the most capture.
- Venture and public markets are shaped by power-law outcomes, so missing the rare big winner matters more than avoiding losers.
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.
Since 1926, just under 60% of U.S. public companies failed to earn treasury bill returns, destroying about $9 trillion, while roughly 2% of companies created about $50 trillion of the $55 trillion net wealth created, so 2% delivered about 90% of value.
“2% of the companies were 90% of value.”
Contrary to textbook predictions, companies during the 2009 to 2021 low-rate period invested less, held more cash, and did not increase leverage, with investment down, cash up, and leverage flat or down compared to the prior 13-year period.
“what we found was almost the exact opposite. Investments were actually down... Cash balances actually up... leverage levels actually, on average, went down”
John Graham's survey of CFOs over three decades found companies use a hurdle rate roughly twice their cost of capital, around 15%, which barely changes with interest rates and makes capital structure decisions sticky.
“companies actually don't even use the cost of capital... they use a hurdle rate that's roughly twice as high as our cost capital and it doesn't really change. And it's roughly 15%.”
Using the value-based strategy framework, economic profit is the gap between price and cost, but network effects raise consumers' willingness to pay; a company that does not raise prices captures more consumer surplus, which builds loyalty and future pricing power.
“am I increasing as a company? Am I increasing willingness to pay every single day?”
Under Wright's law, each doubling of cumulative output reduces cost per unit by about 20 percent, a relationship that fits lithium batteries and solar well and that outperformed Moore's Law in a Santa Fe Institute test of 60 technologies.
“for every doubling of cumulative output, your cost per unit goes down by 20%.”
Combining demand-side value creation that raises willingness to pay with old-school economies of scale that lower unit cost creates a Lollapalooza effect, with huge value creation as a consequence.
“when you start to think about like Charlie Munger's idea of a Lollapalooza effect, you're creating value on the demand side”
Data on around 30,000 venture investments shows 55% to 60% lose money, similar to the public market pattern, and the few big winners pull up the whole portfolio, so capturing extreme events is essential.
“we have data on the returns for 30,000 venture investments over the last quarter century or so. And yes, 55% or 60% lose money.”
Companies behind in an offensive race often launch open source as a defensive strategy, as Google did with Android against the iPhone and with Kubernetes against AWS, to rally carriers, manufacturers, and vendors against a dominant player.
“companies that find themselves behind the eight ball from an offensive standpoint, launch open source as a defensive strategy.”
When several players compete in an increasing returns game, the strategic logic resembles a dollar auction, and once everyone understands the game, price wars such as Uber versus Lyft become bloody and hard to exit, as players keep bidding past sensible limits.
“And the dollar auctions example that they use in these finance classes comes up as an equivalency of that.”
Returns on capital for smaller companies used to match or exceed those of larger companies, but starting around 2000 that flipped, so large companies now grow at decent rates while also earning very high returns on capital.
“the returns on capital for smaller companies used to be bigger or equivalent to larger companies. And starting around the year 2000 that flipped.”
Semiconductor startups now face roughly $50 million for a first tape-out, and if they cannot get onto TSMC's production schedule they are effectively dead, while NVIDIA's power limits options, unlike software where anyone can start building tomorrow.
“it was $50 million for silicon the first tape-out for your alpha chip or whatever. And that's a lot more like some of these biotech plays.”
Uber illustrates increasing returns: more drivers reduce pickup times, which improves the consumer value proposition, attracts more riders, spreads availability to more places, and increases earnings for drivers, creating a win for all sides.
“if you have more drivers, pickup times fall. And if pickup times fall, the consumer value proposition gets better and so more people want to use it.”
Intangible assets scale at near-zero incremental distribution cost, so success can be enormously lucrative, but they face fast obsolescence, are easy to replicate even with patents, and so are hard to sustain excess returns on.
“the cost of incremental distribution asymptotes towards zero. So if it's successful, you can make enormous amounts of money”
At its core, innovation is almost always the recombination of existing building blocks, and today many of those building blocks are digital and intangible, which can be manipulated faster than before, making recombination a source of optimism.
“if you boil innovation down to its core, it's almost always the recombination of existing building blocks”
Open source is among the most compelling innovations for human prosperity because, once shared, anyone can reuse the work at zero marginal cost, and it is the clearest example of ideas being borrowed and built upon.
“open source is one of the most compelling innovations for human prosperity that we've ever had”
Writing and teaching compel clearer thinking, because defending a structured argument reveals gaps, and public distribution builds a thought-leader reputation that helps close deals and win access to top thinkers.
“I just think better when I write... having to defend an argument.”
When new technologies arrive, companies first bolt them onto existing workflows for incremental productivity gains, and only later reorient their whole business around the technology, as happened with the Internet when a wave of Internet-first companies emerged.
“the first generation is it amplifies but kind of in different spots.”
Ideas and recipes are nonrival goods that many people can use at once and are hard to exclude others from, but Romer showed some nonrival goods are partially excludable, letting creators reap part of the economic benefit from intangible assets.
“some nonrival goods can be partially excludable which means you can have the benefits of an intangible assets, but you can reap some of the economic benefits from that.”
If the cost of a complementary good falls, the value of your good rises; Google's Android effectively gave away the complement to mobile, which raised the willingness to pay for its advertising business.
“The idea in economics is if the cost of your complement goes down, the value of your good or service goes up.”
Companies mostly buy back stock to offset dilution from stock-based compensation rather than to signal undervaluation, and during the ZIRP period buybacks were accretive to EPS, but now they are roughly neutral.
“companies buy back their stock largely because it offsets dilution from stock-based comp.”
Since the 1970s, intangible investment has gone from smaller than tangible investment to substantially larger, flipping the relationship between the two categories of corporate spending.
“in the 1970s, tangible investments were bigger than intangible investments. And today, that relationship is almost completely flipped”
Truly increasing marginal utility, where the value proposition to one side rises exponentially as the other side grows, is rare; some people even think it does not exist, though Gurley is certain it does, and it can be scaled as linear or exponential.
“They're so rare in fact, that some people think they don't exist, but I'm pretty certain they do.”
George Stigler, who won the Nobel Prize at the University of Chicago, identified regulatory capture, and the longer people stay in Washington the more they learn to influence it, so regulation becomes a friend of the incumbent.
“the phrase I use is regulation is a friend of the incumbent.”
Good public-interest regulation and regulatory capture are two distinct things, but people lump them together; the risk is that the country has taken its eye off public-interest regulation and drifted into a regulatory capture environment, with money in Washington distorting outcomes.
“there is good regulation, and you called them public interest regulation”
The top five technology companies spend about twice as much on capital expenditures as the top five energy companies, a level that would have seemed implausible 25 to 30 years ago and that makes it very difficult for new companies to keep up.
“the technology companies are spending 2x the CapEx as the energy companies.”
LLMs are a distinct and narrower category from AI, which has been advancing for about 20 years; LLMs excel at text, notably writing and analyzing code, but are limited because they mainly process text, while AI such as Tesla's full self-driving approach does not involve LLMs.
“you have to separate LLMs from AI. AI is this trend that's been going on for 20 years.”
Hard-tech investments in silicon and energy are harder than software because a first chip tape-out can cost around $50 million, TSMC scheduling and NVIDIA's power can lock out startups, and materials science has followed linear rather than Moore's-law-like improvement.
“it was $50 million for silicon the first tape-out for your alpha chip or whatever... if you can't get in their schedule, you are toast.”
CrowdStrike's security cloud network effect works because threats shared across a network lower each member's exposure, so a customer joining the largest network gets more utility than one joining a smaller network, meaning customer N+1 has better value than N.
“if the threats are shared across a network, then if you belong to the biggest network, then you get the shared learning of everyone in that network”
An academic study of companies seeking to become platforms found that roughly one in six or one in seven achieved it, which suggests people overestimate how common platform-level increasing returns are.
“it was like something like one in six or one in seven that we're able to achieve that.”
Zero-based thinking, asking how one would design a system from scratch with no prior layers, shows that nuclear energy and psychedelic medicine would be obvious beneficial investments if regulatory and emotional baggage were removed.
“If a Martian came down and said, oh, you guys need energy... This is the obvious thing you guys should be doing lots of.”
Venture has low barriers to entry and high barriers to exit, so funds are easy to raise and last about 15 years instead of 10, meaning capital from a boom like ZIRP stays in the system and decimates returns through supply-demand imbalance.
“there's low barriers to entry and high barriers to exit. And so it's very easy to raise a fund. Once you do, these days you're on the field for 15 years.”
Jim Bessen's New Goliaths thesis holds that large firms invested heavily in proprietary software, letting them capture scale benefits and customization, and unlike earlier technologies this software did not diffuse through the economy, creating superstar firms.
“these companies had enormous resources that allowed them to invest substantially in proprietary software”
Large companies spend roughly $200 billion to $250 billion per year on proprietary software that they do not share, which captures economic rents but works against the diffusion of innovation that open source provides.
“a lot of big companies today are investing huge sums of money on proprietary software. When I say huge sums of money, $200 billion, $250 billion per year”
Big banks have no interest in small IPOs, and the requirement that a company reach roughly $1 billion in revenue before going public has made the path to public markets a different game that lacks the American spirit that anyone can start a company that goes public.
“The big banks have no interest in a small IPO.”
Material science and hard tech did not follow Moore's law, so progress in solar and similar fields has been more linear and involves many shots on goal, and regulatory access to government money matters more than R&D results, which is why venture avoided hardware.
“the other problems that have existed material science typically didn't follow the same innovation path the same Moore's law did.”
Psychedelics were considered promising in the 1950s for depression and addiction, then fell out of favor for decades, and nuclear energy would look obvious from a zero-based view, so both show how regulatory and emotional baggage distort outcomes that half-century trends have bent.
“psychedelics, which in the 1950s were demonstrably helpful for certain people for things like depression and addiction”
Paul Romer made technology endogenous to economic models, moving it from the exogenous residual Solow identified into the framework itself, so that recombining ideas to improve the use of labor and capital drives greater output.
“what Romer came along and figured out was, we can make this endogenous, we can build it into the framework.”
Public interest regulation, which aims to benefit the broad public, is a distinct concept from regulatory capture, which serves special interests, and people wrongly lump them together in one bucket.
“there is good regulation... public interest regulation... And then there's regulatory capture... they are actually two very distinct things.”
The Magnificent Seven generated about 45% of the economic profit of the U.S. stock market, while their market caps were only about 25% to 30%, so their profit contribution exceeds their valuation weight.
“the Magnificent Seven specifically generated about 45% of the economic profit for the U.S. stock market.”
Predictions of universal basic income and productivity gains forever from AI optimists are impossible, because competition is always present, so any gain cannot be held onto indefinitely.
“just because you get a gain doesn't mean you get to hold on to it forever.”
In a low-rate environment investors had to pursue returns and moved out on the risk spectrum into private equity, with buyouts using leverage and venture buying younger, higher-failure companies to generate excess returns.
“in a low-rate environment, a lot of investors understandably had to pursue returns. And so they moved out on the risk spectrum”
Tesla's per-unit cost advantage in electric vehicles comes from cumulative output, and being ahead keeps widening the lead, which is why learning-by-doing is an increasing-returns mechanism.
“Tesla's cost per unit advantage. Well, they're just so far out ahead of everybody”
The number of U.S. public companies has shrunk by about 46 percent, a decline Gurley finds unhealthy because it pushes companies to stay private longer and drives regulatory attempts to open private markets to ordinary investors.
“the number of public companies has shrunk dramatically... Yes, down 46%.”
Engineering leaders report that AI-generated code can make debugging take longer, because engineers must fully understand code they did not write, so the expectations around AI have outrun a middle-ground reality.
“The debugging is taking longer than it did before because writing code, I don't understand, so now I got to go fully understand the code.”
A 17-year drug patent is too long, and drug companies make small changes to reapply patents to draw out protection, whereas software patents written like drug patents would bring the software industry to a halt.
“I don't know that 17 years makes any sense whatsoever and why that industry gets protected down to a very minute detail of innovation”
China's more autocratic government can clear regulatory obstacles for every project at once, while layered US and California regulation makes projects like renewables and infrastructure slow, giving China a massive advantage.
“This is where I think China has a massive advantage, they could just clear the decks for good for every project going forward”
Venture is one of the few asset classes with high persistence, meaning past winners tend to remain winners, which makes access to top managers very valuable and difficult to obtain.
“it's one of the few asset classes with high persistence”
Energy usage is one of the big unsolved problems that creates an opening for innovation, and nuclear is the most efficient way to create energy, though the US stepped off the learning curve and red tape must be removed to return.
“we know that the most efficient way to create energy is nuclear, and we stepped off of the learning curve that Michael talked about.”
The claim that no one will invest in R&D without hard patent protection has been proven false, because several companies have invested billions of dollars in Linux, which has no such protection, often for defensive reasons.
“the number of companies that invested billions of dollars in R&D in Linux, there's several.”
Large technology companies, unlike AT&T in its era, are now outperforming smaller firms in growth and value creation, which is unusual regardless of AI.
“in 2023, the Magnificent Seven or whatever, we have the largest companies on the American Stock Exchange at least outperforming most of the smaller companies”
Health care is ripe for change because of its poor management and regulatory hurdles, and hospitals are not complying with the price transparency that Congress demanded, paying fines instead, which shows how hard it is to get the industry to act.
“Congress demanded that the hospital systems disclose their prices and several of them just sit on their hands and paid the fine.”
The industries regulated the longest, such as healthcare, finance, and telecom, are the most captured or trapped, with healthcare being a massive problem and Boeing's move to Washington, D.C. illustrating growing proximity to regulators.
“the industries that have been regulated the longest are captured or trapped.”
Total U.S. venture assets under management are roughly $1 trillion to $1.5 trillion compared to about $50 trillion in public equity market cap, and early-stage capacity is small, with Benchmark's recent funds at $450 to $500 million versus multi-billion-dollar funds elsewhere.
“the market cap of public equity is probably at year-end, $50 trillion-ish. The assets under management for U.S. venture, a little over $1 trillion, maybe $1.5 trillion”
After major legislation like Dodd-Frank, competition in the affected industries went down rather than up, which Gurley cites as evidence of regulatory capture.
“you look at major pieces of legislation like Dodd-Frank and others and you see competition goes down, not up, after they happen.”
Investment patterns decay over time because the world changes, so a venture capitalist's accumulated pattern recognition from 30 years has diminishing edge, and Gurley is not convinced long tenure gives massive competitive advantages.
“I think the patterns have a half-life in a decay”
Venture firms should frame partner discussions around how big an opportunity could be rather than nitpicking failure, because cynicism spreads easily in large groups and missing the big winner matters far more than avoiding a loss.
“make the primary part of the discussion, how big could this be rather than trying to nitpick whether or not it might fail.”
Daniel Kahneman was willing to be proven wrong and actively sought views different from his own, reasoning that abandoning a false belief moves one closer to truth, which Mauboussin considers an inspirational way to live.
“if there's a truth out there, and I now no longer harbor a false belief, I'm stepping closer to the truth.”
The most valuable thinkers are truth seekers who are open to discussion and willing to be proven wrong, and engaging them with specific mental-model conflicts makes relationships powerful.
“these people that are just kind of truth seekers that are just always open for discussion and always trying to get to the next place.”
Unlike past technology windows that opened for startups unexpectedly, the current AI wave is highly choreographed, with platform players like OpenAI and others publishing APIs and leaning in, which reduces the openings for new entrants and may favor incumbents.
“This appears highly choreographed. The windows that open up huge doors for innovators and start-ups are usually not choreographed.”
Initial foundation models are structurally flawed at providing personal memory and at letting people become dependent on them, and approaches like RAG do not seem to solve this; a model that retunes for each individual or a local model with phone context could change the game.
“the initial versions of these models are structurally flawed around providing personal memory”
Founders who succeed at building a product are not necessarily good at allocating capital, and in increasing returns games players who know the outcome may still be trapped into burning capital, as with foundation model companies spending $100 to $200 million a year.
“There's no way possible from my standpoint, that the companies at the foundation models... are burning, $100 million a year, $200 million a year.”
Foundation model companies are burning $100 million to $200 million a year, and Bill sees no chance of high-quality capital allocation there; they may be stuck in a ridesharing-style price war because they lack alternatives, similar to a dollar auction.
“There's no way that high-quality capital allocation from my point of view.”
Celebrating exceptions that clear bureaucracy, like the I-95 bridge rebuilt in 12 days or the TSMC Arizona plant getting regulatory exceptions, reveals that the system has painted itself into a corner and needs permanent zero-based reform.
“Now when we start celebrating people moving bureaucracy out of the way, in some ways, we're admitting that we painted ourselves into a corner.”
The key investor question about AI is whether it is a sustaining innovation that helps incumbents get stronger, as big cloud players like Microsoft, Amazon, Google and Meta already appear to be, or a disruptive innovation that lets new companies break in.
“does this just help the strong get stronger? ... versus does this allow new businesses to start up and merge and disrupt the traditional guys?”
A rule set in software that only Twilio succeeded with small business led venture investors to develop an anti-small-business mindset, causing them to pass on HubSpot, Shopify and Twilio, which shows how a pattern can cause large misses.
“in the software world, Twilio was the only company that succeeded with small business and everyone else failed. And so, you develop this anti-small business mindset”
In the 1930s and 1940s, automakers' consortium-style sharing and easy reverse engineering let General Motors' 1940 automatic transmission be offered by every manufacturer within a decade, showing fast diffusion of technology.
“within a decade, every manufacturer had offered an automatic transmission, and this is 10 years with a war in the middle”
The foundation model companies have moved to a scale similar to Uber, Lyft and DoorDash, where so much money is flowing that they are no longer really startup markets; the conversation should have happened four or five years earlier to be nonconsensus.
“It's already evolved to a point that's very similar to where Uber, Lyft and DoorDash ended up”
John Holland's genetic algorithms solve problems by generating many candidate programs, scoring them with a fitness function, and breeding the top performers over generations, which is a form of recombining ideas toward a goal.
“he developed an idea called genetic algorithms. And the idea is pretty simple, you have a problem you're trying to solve, and you have some sort of fitness function”
If you are paying $90 billion post-money for Stripe or OpenAI, you are unlikely to achieve the venture-style returns implied by the skewed-return data, because late-stage rounds are often pure jump balls with heavy competition.
“if you're paying $90 billion post for Stripe or Open AI, you're probably not looking at the type of returns Michael was talking about.”
Big companies using their own balance sheets to reinforce AI company valuations is remarkably dangerous and unhealthy, and high market caps reflect discounted future expectations that set up disappointment.
“Part of it ties into the big guys being interested in what they're doing with their own balance sheet to reinforce this, which I consider to be remarkably dangerous and unhealthy.”
Distinguished people such as Bob Dylan, Bobby Knight, and Danny Meyer built their craft through very large amounts of hard work and drew on diverse influences, so the visible output hides the years of effort behind it.
“these guys worked incredibly hard at mastering their craft and putting lots and lots of hours, lots of passion, lots of sweat.”
Kenneth Arrow, a Nobel laureate in general equilibrium theory, wrote his most-cited paper in 1962 on learning by doing, where productivity improves with experience, and he encouraged complex-systems work at the Santa Fe Institute.
“I think his most cited paper is actually about learning by doing, which is 1962.”
Murray Gell-Mann reacted sharply when Gurley confused the observer effect in quantum measurement with the Heisenberg uncertainty principle, showing that these concepts are often conflated and that precision matters to top scientists.
“the thing he's describing is the observer effect... people that confuse the observer effect with the uncertainty principle, and he let me know it.”
Economic profit equals the spread between return on capital and cost of capital, multiplied by the invested capital, so it is a dollar measure of value created above the cost of funding the business.
“it's return on capital, less cost of capital spread. So you're earning above your cost of capital times the invested capital in your business.”
Gurley and Mauboussin read Brian Arthur's work on increasing returns and network effects in the Santa Fe context, which profoundly shaped Gurley's venture career by directing him to find companies where increasing returns are true.
“it outlined Brian Arthur's original thesis on increasing returns and network effects. And probably no other kind of theory affected my venture career more”
A research report from the prior year found that venture capitalists are the investor group most enthusiastic about pattern recognition, which is intuitive expertise built from seeing many similar cases.
“the group of investors who seem most enthusiastic about pattern recognition are venture capitalists.”
Intangible investment now dominates tangible investment, which means strategic analysis must account for scalable digital building blocks and their obsolescence risk.
“So there's sort of the intangible investment, first, documenting that it's much bigger than it used to be.”
Gurley looks for increasing-returns potential in most investments because it is an unfair competitive advantage, though he acknowledges he is probably overly optimistic about being able to create it, and entrepreneurs can be provoked to think about it differently.
“I think it's such a unfair competitive advantage. I think it's just such a great way to win.”
The big unlock is an AI that holds all of a person's prior knowledge, emails and conversations, like the movie Her, whereas today's LLMs are an advanced Wikipedia that has ingested the world's information but not the user's own information.
“the LLMs today are really more like a advanced Wikipedia that can also code, but it has sucked in the world's information. But it hasn't really sucked in your information”
An AI that has all of a user's emails, conversations and prior knowledge, able to recall past interactions, would be a game-changing product that is currently cost-prohibitive but that foundation model companies see as an opportunity.
“And I do think it will be the game changer if someone can get it right.”
Companies owning creative productivity software such as Microsoft and Adobe quickly realized AI would be a required feature because everything is cloud-based and API-based, and they moved fast to integrate it.
“the companies that owned the creative productivity apps very quickly realized that this was going to be a required feature.”
Many people are wrapping unrealistic expectations around LLMs and AI that will outpace what is reasonable in the short term, such as the claim that an idle computer will find a cure for cancer, which is not going to happen soon.
“that's not going to happen anytime soon.”
When Bill evaluates venture investments, he looks for increasing returns in most cases, though other success paths exist such as early go-to-market advantage or a strong founder, and he admits he is probably overly optimistic about being able to create them.
“I look for them in most cases. I'm probably overly optimistic about being able to create them.”
Compensation for experts hired by AI companies, including expected large secondary payments, has reached pro-baseball-level money, which is beyond what would be considered the startup world.
“this is pro baseball level money at this point”