Second-Order Effects Shape Gurley’s View of AI, Stablecoins, and Venture Capital

Bill Gurley returns most often to systems thinking: the habit of treating a problem as a “multi-variable non-linear” system rather than a chain of simple, one-variable causes. His reference point is the Santa Fe Institute, where he serves on the board and where complexity theory is central. Weather and stock markets are his examples of systems that can behave one way for a long time, then change sharply when one variable shifts.

The practical value is not that systems thinking lets you predict everything. It is that it keeps you from overtrusting a narrow model. A change can have first-, second-, and third-order consequences. A team can improve one local metric while damaging a more important downstream outcome.

His example comes from a large dating site. The company had a plausible hypothesis: longer user profiles would increase engagement. Testing supported the idea. The company rolled it out. Months later, it learned that the change hurt conversion. When people knew more at that stage, fewer converted.

That is the kind of delayed effect Gurley thinks a systems lens helps expose earlier. The error was not testing. The error was becoming too deterministic about a single metric or single variable. “You just gotta be really conscious of the consequence,” he says, and distinguish what is important from what is not.

That frame organizes much of the rest of his argument. AI regulation is not just a safety question; it can become a moat for incumbents and a constraint on competition with Chinese open-source models. Startup funding is not just a question of whether more capital is available; it changes burn rates, risk, and whether a company can still read its unit economics. Tokenizing private shares is not just a liquidity innovation; it may import public-market volatility into companies that stayed private to avoid it. Payments are not just a consumer convenience problem; slow U.S. infrastructure creates the opening for stablecoins.

The recurring move is to ask what an intervention changes elsewhere in the system, who benefits from the new constraint, and which consequence will show up too late for the original metric to catch it.

Gurley’s investing education began on Wall Street, not in venture capital. The early canon was conventional: Peter Lynch’s One Up on Wall Street, Burton Malkiel’s A Random Walk Down Wall Street, Buffett’s letters, Ben Graham, and Howard Marks. Those sources gave him what he calls a “bedrock of financial understanding.”

That foundation mattered even after he moved into a venture business that did not look like traditional value investing. Gurley’s bridge is Bill Miller, the Legg Mason investor introduced to him by longtime friend Mike Mauboussin. Miller claimed to be a value investor while holding Amazon as a large position for a very long period. His definition, as Gurley recounts it, was that value means an asset is underpriced relative to what it will be worth in the future. If network effects allow a company like Amazon to grow at an unreasonable rate for a very long period, the value framework can still apply.

Gurley thinks many Silicon Valley venture capitalists would benefit from better financial foundations for another reason: Wall Street is ultimately one buyer of the product venture creates. Liquidity comes through M&A or an IPO, and public-market investors help set the terminal price. Even when a company is only “two people in a PowerPoint,” Gurley wants to understand whether the mature version will be something those buyers value.

But bedrock alone is insufficient. Gurley pairs it with obsessive attention to the “bleeding edge.” In almost any career, he argues, the unusually strong person understands both the history of the field and the newest thing disrupting it.

To explain what knowing the history of a field looks like, Gurley tells a story about a dinner at John Lasseter’s house. A Benchmark partner had bought the dinner at a charity auction. Lasseter served a ten-course meal in his viewing room, with each course tied to a classic cartoon he considered essential to understanding animation. He screened the cartoons and explained them. Gurley’s takeaway was that Lasseter knew the history of his craft at a depth that changed how one saw his creative work.

He gives another example from chess: Magnus Carlsen winning a trivia contest about chess history during a break at a world chess tournament. And he points to Picasso, who was already a successful realist painter by age 14, a fact not obvious from looking only at his later Cubist work.

The practical career argument is simple. If a new graduate interviewing for a marketing job at P&G or Pepsi understands the masters of marketing better than the other candidates, and can bring that into the interview, Gurley thinks the contrast would be powerful. It would show not only knowledge but passion. If learning the history of the field feels tedious, he says, that itself may be evidence that the person is not in the right lane.

At the edge, the relevant trait is obsessive learning. The strongest founders Gurley has worked with are not necessarily historians of their fields, but they constantly learn because technology waves create openings when something dynamic is changing: AI now, mobile before that. When mobile phones emerged, there were no experienced mobile app engineers. A small group learned the edge first and built from there.

Gurley applies the same standard to himself. As a venture capitalist, he says, he was always afraid a new app would appear in the App Store that he had not seen. His response was to play with everything. In AI, he has “like five premium AI accounts” because he does not want to miss something.

For young people, the edge is a way to compete with incumbents. A marketing candidate who knows both the old masters and TikTok has a differentiated combination. A young venture capitalist who understands esports or YouTube creator economics can know more than famous generalist investors because they can spend all their time under that rock. Venture bends toward youth, in Gurley’s view, partly because it is a hustle business and partly because new technologies are often understood first by people closer to them.

Bill Gurley uses AI less as a single tool than as a portfolio of models matched to tasks. The surprising part, he says, is how often users underestimate how much work the model can do earlier in the prompt. Instead of asking for the top 10 of something and then manually studying, comparing, and ranking the list, the user can ask for the top 10, their pros and cons, rankings on one dimension, and rankings on another. Early on, he would ask for numbers and then add them up himself before realizing he could ask the model to do that too.

He likes ChatGPT’s project structure and says he is being “sucked into” its memory because it knows who he is and knows things about him. For restaurants, he uses Gemini because of its access to Google review data, asking not only which restaurants are good but which dishes people rave about and what people warn against. Coding users, he says, swear by Claude. A person he met that morning preferred Perplexity for finance, but Claude for deep research on unfamiliar companies or countries. His conclusion is not that one model has won, but that usage is still a mix.

The larger AI question is whether the world ends up with one dominant general model, a set of niche models, or commodity model layers swapped in and out by applications. Gurley treats the answer as contingent. In coding, users and products already switch models; Cursor lets users choose which model they use. As optimization and price optimization become more important, he says, more swapping may occur.

The investing question follows from that uncertainty. One view is that if general models become near-sentient, there may be no need for vertical AI companies because one model will do everything. Gurley “probably” comes down on the other side, but not as a settled conclusion. Workflows and data moats may matter. Legal AI startups, for example, spend their time ingesting case law and understanding legal processes and principles. Once implemented in a firm, they write on the user’s behalf and build new databases from that work. Gurley is not convinced a user would simply switch that whole workflow back to ChatGPT as the general model climbs the stack.

But he acknowledges the other side. Large model companies have discussed going after verticals in their product groups. Microsoft is the comparison: it began with the operating system while applications like Lotus 1-2-3 and WordPerfect existed elsewhere, then moved up the stack. “That could happen,” Gurley says. “We're gonna see how it goes.”

On training limits, Gurley thinks there is a valid argument that available data may be running out. He calls it “painting in the corners”: the major open areas have been filled in. One current method for improving models is to hire experts for thousands of dollars an hour to fine-tune them by asking hard questions and training the model to solve them. Gurley assumes there must be some limit to that and asks where the edge of human knowledge lies.

That leads to the dispute over asymptotes and superintelligence. Shane Parrish describes the theory that once a model becomes superintelligent, it can make itself slightly better, producing a nonlinear curve. Gurley says some people make that argument, but he is not sure he believes it.

Rather than making himself the authority, Gurley points to Yann LeCun’s criticism: the next version of AI may not be LLMs, and language-based models may hit an asymptote because language cannot capture everything. Gurley connects this to why language models are not specifically great with math and numbers.

The standard counterexample is AlphaGo, where Google’s system found a move shocking to humans. Gurley says that is often cited as proof that AI can innovate beyond what it has been taught. The opposing view is that Go is a constrained game. Computers can search a possibility space far larger than a human can search. The real world is not constrained that way; it contains an effectively infinite number of paths in a complex system. Gurley also notes that AlphaGo was not LLM-based but trained for a specific constrained system.

Tesla’s full self-driving is another constrained system: visual inputs produce outputs through brake, steering, and accelerator. Parrish says he would now be comfortable sitting in the back seat while his Tesla drives. Gurley would not. The issue is corner cases. In a world where all cars in a geographic area were autonomous, he says, it would be easier to accept. In the real world, humans introduce randomness, including people who think it is fun to test the cars by jumping in front of them.

The useful boundary in Gurley’s view is not “AI works” or “AI fails.” It is whether the environment is constrained enough for the system to search, learn, and act safely — and whether the business model around it depends on capabilities that may still be unresolved.

Regulation is one of the forces that could shape AI market structure. Bill Gurley says that if compliance becomes difficult, mundane, and expensive, it could lead to more oligopoly. Parrish suggests that some incumbent players know this and are “begging for regulation.” Gurley agrees with the incentive: regulation can become a protective moat, especially against Chinese open-source models.

On global AI regulation, Gurley is explicitly uncertain. Copyright is his example. If American models must follow special rules, while Chinese open-source models do not, that could have an effect. He does not know how the EU would rule in such a situation.

His systems analysis of China’s AI ecosystem is sharper, though still framed as his assessment. Parrish mentions several strong Chinese open-source models; Gurley says there are “like 10.” Because competition in China is more intense, he says, many players have chosen open source. That creates a system he believes is capable of innovating faster than the competitive system in the United States because models can learn from each other, train each other, and test each other.

His metaphor is agricultural. Imagine two farming societies. In one, farmers come to market, sell each other goods, and return home. In the other, they are forced to share best practices with all the other farmers. The second society should evolve faster.

That openness also helps Western startups, though with irony. Gurley says many startups are forking Chinese open-source models, including in Silicon Valley. He calls it a “quiet secret” because he has not seen it on the front page of the Journal, but says that from a breadth or volume standpoint, companies are using these models throughout Silicon Valley. Whether regulation later tries to stop that remains, in his view, an open question.

The non-consensus backdrop is Gurley’s discomfort with reflexive vilification of China. Having spent a great deal of time there over the past 20 years, he says it is hard for him to adopt the mindset now common in Washington and parts of Silicon Valley. The United States, he notes, is only a small share of global population. When people say “American exceptionalism,” he wonders what the other 95% of the planet thinks when they hear it.

The strategic tension is that regulation meant to constrain risk can also constrain learning. If one system shares more of its improvements and another raises the cost of participation, Gurley thinks the learning rate itself becomes a competitive variable.

Gurley says he would not have believed, five years ago, that the largest technology companies would become worth trillions and then take free cash flow that had been $50 billion to $100 billion a year “down near zero” by spending it all on capital expenditure. The size of the AI buildout shocks him.

He connects that willingness to spend with a broader change in investor psychology. Venture investors and the wider investor community have absorbed the lesson of increasing returns and power laws. Companies such as Google, Amazon, and Meta became the reference cases because their scale, footprint, and users appeared to reinforce further growth, and because they ended up worth far more than many expected. As more investors believe this pattern, they become more willing to invest ahead of evidence and take more risk.

The losses before cash-flow positivity keep growing. Gurley cites a chart someone sent him showing the leading company in a field before it went cash-flow positive: Amazon lost roughly $2 billion or $3 billion, Uber around $15 billion, and current AI companies may lose far more. His conclusion is that the venture community as a whole has become more risk-seeking because of what it learned from prior winners.

A correction remains possible; Gurley does not claim to know whether or when. He says there has not really been one yet. The dot-com crash is his cautionary analogy: a three- or four-year lull, almost a “nuclear winter,” before companies like Amazon climbed out again. AI today is still surrounded by optimism. He does not know whether the market reaches a point of very little optimism.

He is especially concerned about “circular deals.” He refers to Dario Amodei being asked about them at the DealBook conference and explaining, in Gurley’s paraphrase, that a cloud service provider might fund a model company because the model company wants to spend $5 billion developing a model but does not have the money. The provider gives the company money, which the company spends back on the provider’s services.

Gurley’s objection is straightforward: if the provider did not give the money, the company would not spend it. These arrangements enhance growth by funding customers to spend on the funder’s service. In his phrasing, they may both “enhance the probability” of a correction and “extend the time” before one arrives.

The same capital abundance changes startup behavior. If a company is successful, someone will try to preemptively fund it. If it takes $300 million, the only way to use that money is to raise burn.

Ten years ago, Gurley says, burning $1 million a month was very risky. Today, some companies burn $5 billion a year, or more than $100 million a month. At that level of financial aggression, he argues, it becomes difficult to know what the unit economics really are.

The concern is not simply “too much money.” It is that the memory of power-law winners can make larger and larger losses feel rational, while the scale of the spend makes the underlying economics harder to read.

Bill Gurley sees no shortage of funding for startups right now, so retail access to private-company assets is not solving a capital bottleneck. But tokenization could still change pricing, access, and market behavior.

His caution is that tokenizing assets without financial disclosure regulation could produce heavy speculation and manipulation. Private companies also deliberately avoid the daily market-price dynamic. If a legally tokenized share of a company such as Stripe fluctuated wildly in price, Gurley says it would affect the company and its employees. One reason companies stay private is to avoid exactly that.

When private companies run employee liquidity events, they negotiate a price with a handful of trusted investors. The underlying asset may move around significantly, Gurley says, but because it is not constantly recorded, operators do not have to manage the consequences. Public company CEOs, by contrast, know volatile stock prices create chaos for employee-owners trying to interpret what the moves mean.

His more enthusiastic case for tokenization concerns the IPO process. He calls the current process “insanely unfair” to companies because bankers pick the price and the shareholders. In his view, if a freshman computer science student and a freshman finance student were asked to design how a company should go public, they would match supply and demand anonymously in an auction, similar to how an ICO works.

Gurley says he and others pushed direct listings because they used a more auction-like mechanism. Wall Street could have embraced that, in his view, but returned to a controlled oligopoly. Tokenization, even if it only improved how shares are allocated, could be disruptive.

Stablecoins are the sharper payments example. Gurley argues that stablecoins are compelling in the United States because the domestic payment system remains, in his view, unnecessarily slow and expensive. He says most of the developed world has government-enabled instant transfer from bank account to bank account and from bank account to merchant. He cites the UK’s Faster Payments system as doing this 20 years ago, and says Argentina did it “with PIX” in the past six years and that it quickly became 60% to 70% of transactions.

The United States has not done the same, Gurley argues, because of regulatory capture by banks. The government has FedNow, but he says there is massive pushback in the finance committee in Washington. The result, as he frames it, is a credit card ecosystem charging roughly 2% to 2.5%, with a set of companies built underneath that umbrella.

A stablecoin, as Gurley explains it, is a cryptocurrency that is backed dollar-for-dollar by assets such as U.S. Treasuries if the issuer follows the regulation. He says he believes USDC is doing that. Because stablecoins run on crypto rails, which he describes as proven, fast, global, and immediate, they allow dollars to move quickly between people or companies.

In the U.S., sending $50 digitally through ACH involves three-day settlement. Gurley rejects the idea that this is technically necessary. Same-day wires are possible, but they can cost $25 and require forms and sometimes a verbal bank confirmation. Stablecoins route around that system. Gurley thinks they may get there faster than the government.

That is why he sees Visa and Mastercard as heavily threatened. He says the two companies have among the highest operating margins in business, around 60%, operate as duopolies, and were created by banks that retain a stake in the current structure. He sees “zero reason” payments should cost 2% or 3%.

China is his illustration of a different path. Because money transfer became easy, Alibaba and Tencent built digital wallets that consumers use for everything from street vendors to car purchases. Restaurants can put a QR code on the table, and customers pay with WeChat Pay or AliPay in one click. Gurley’s conclusion is not that the alternative must be stablecoins. It is that where governments made instant transfer easy, payment systems innovated far beyond the U.S. model. Because the U.S. waited so long, he thinks the threat may now come from stablecoins, especially given crypto momentum in Washington.

The pattern is the same as in his IPO critique: when an old intermediary controls allocation or settlement, the opening for disruption is not abstract technology. It is the accumulated cost of a bottleneck.

AI’s effect on established information intermediaries is unsettled but broad. Bill Gurley takes Moody’s as an example: a company that sells analysis on debt and has strong margins. In theory, AI could perform equal or better analysis. Gurley responds that Moody’s power comes from being a trusted standard. Even if it uses AI on the back end, the Moody’s name remains the watermark.

Proxy advisory firms are more vulnerable in his telling. Gurley thinks the U.S. has reached a bad place because passive index funds hold large blocks of shares but do not have time to evaluate every vote. They rely on services such as ISS that score companies through a black box. Gurley says the way companies can learn more about the score is by hiring the same firms, meaning the firms get paid on both sides.

He calls the arrangement closer to a “heist” than a service. In his view, the advisors drifted away from shareholder interest and toward rule-based corporate governance risk mitigation. The Tesla compensation package for Elon Musk is his example. Gurley says he would agree to that kind of package for every company he has worked with, and most CEOs would not take it: the executive makes money only if the stock goes way up, and if it does, earns an enormous amount. Proxy evaluators opposed it, in his telling, because they focused on the headline number and rule deviation rather than what shareholders would receive if the conditions were met.

The rise of passive ownership creates second-order effects beyond proxy advisors. Gurley suggests one improvement would be for index funds not to vote, allowing active shareholders to have more say. Another proposal is that index funds could vote in the same proportion as direct holders. Gurley says non-voting would naturally produce something like that outcome.

He also sees passive ownership changing active management. At first, active public investors became scared because they were measured against indexes. If the Mag 7 rose and they did not own them, they had a bad year, pushing them toward “closet indexing.” But with so few active investors left, Gurley says some now argue that the ability to gain an edge may have increased. He does not claim certainty. Beating the S&P remains hard, and he notes that QQQ has likely outperformed 80% or 90% of venture funds.

The deeper point is that a trusted standard is harder to displace than analysis alone. AI may make the work cheaper or better, but the institution that decides what counts can remain powerful until the trust mechanism itself changes.

Bill Gurley puts storytelling among the top traits of successful founders. He came to that view partly through reading. Before business school he did not read much, then began with well-known business books, moved into personal development books such as Dale Carnegie and The 7 Habits of Highly Effective People, then biographies, and eventually long-form nonfiction journalism that reads with the force of fiction.

He names Malcolm Gladwell, Michael Lewis, and Jon Krakauer as writers in that mode, and says he studied books about “New Journalism” and “New New Journalism.” The attraction was the power of 20 pages of nonfiction to affect the reader deeply. He also studied Buffett and Howard Marks as investors who wrote publicly and clarified their thinking.

Writing, for Gurley, is not only communication after the fact. It is a way to think. His most successful investments often fell into what later became known as marketplaces. Before there was a knowledge base for marketplace companies, he and others had to craft one, codify it, and write it down. That process helped them reason through corner cases.

He connects this to Jeff Bezos’s six-page memo practice at Amazon. If an idea has to stand alone in writing and remain cogent, the writer has to think through loose ends and make the argument more cohesive. In venture, writing also becomes a calling card. If founders who do not know an investor see that the investor understands their problem, they may reach out.

Founders need storytelling because they are always selling: recruiting employees, recruiting executives, raising money, closing customers, closing partnerships. Gurley names Bezos, Shopify’s Tobi Lütke, and Spotify’s Daniel Ek as examples of founders gifted at describing what they are trying to do in a way that makes others want to follow.

Product instinct is another founder advantage. Gurley says it took most of his career to fully understand how hard it is to hire someone who is not “product first” and make them good at product. He allows that examples exist, but estimates the success rate at 5% or less.

The third quality he emphasizes is determination. He recalls asking Jeff Bezos how he built such a successful angel portfolio despite having no free time. Bezos told him he asks one question when he meets an entrepreneur: is this person going to do this no matter what?

Gurley says that level of determination is present in all great founders. The skill stack is not merely charisma plus persistence. It is the ability to understand the product, explain the mission, recruit belief, and keep going when the market has not yet agreed.

The real-world lesson from Uber, for Gurley, was that there are situations for which no Harvard Business School case study exists. Investors in ride sharing understood the winner-take-all dynamics and network effects. That created a logic of continuous funding. If Lyft received $1 billion, Uber might receive $3 billion. The only way to compete in that environment was to spend.

The burn rates became larger than any public company would spend to pursue a new category. Gurley remembers realizing there was no board of Walmart, Costco, GM, or General Electric to call for guidance. The best board members at the best companies would not have faced that situation before. There was no mentor to find.

He calls that recognition harrowing. Uber was, in his phrase, an early case in “mega burn,” even if Amazon had burned heavily before. AI companies now face the same kind of situation, only with “a zero” added.

Benchmark offers the internal version of the same systems lesson: structure changes behavior. Gurley attributes part of Benchmark’s culture to its founding reaction against hierarchical venture firms. The founders had worked in firms where senior patriarchs took too much money and credit relative to the work needed for firm success. Benchmark instead made the partnership equal: no lead partner, no king, no president, just five equal partners.

The second- and third-order effects matter more than the formal structure. First, Gurley says, it made recruiting exceptional investors from other firms easier. A hierarchical firm could offer equality only after someone threatened to leave; Benchmark’s structure already worked that way.

Second, it encouraged partners to develop new people. If Gurley shares equally in a new partner’s success, he wants that partner to succeed. The structure reduces internal competition. If one partner’s company needs a CFO and another partner knows a candidate, the incentive is to share immediately because one company’s success is economically equivalent to another’s. The annual political overhead of compensation review and recutting the pie disappears.

The major negative is that without a CEO, scaling and new initiatives become difficult. No one clearly owns cross-firm projects. The website became the running example. When Matt Cohler joined, he took on the website, built a complicated version with founders connected to partners, and then faced complaints when details were wrong. Eventually he removed it and replaced it with a splash page. Gurley says Benchmark still has a single page, in part because of the organizational issue he describes.

He does not present Benchmark’s model as the only successful one. Many highly successful venture firms are structured differently. But he sees clear trade-offs: less internal politics and more shared incentives on one side; less centralized capacity for initiatives and scale on the other.

In a world full of capital, founders still choose investors for more than price. Gurley says successful venture capitalists benefit from reputation: Mike Moritz and John Doerr had track records that made their stamp of approval valuable in itself. Some people call venture the only investing category with network effects because reputation improves deal flow. Beneath that, founders want to work with people who understand what they are doing and are excited about it. That again gives younger investors an opening when the new field is one they understand more deeply than the established generalists.

Bill Gurley says his definition of success has changed. Looking back on his venture career, he made a specific decision to stop because he felt there was no work left to do. Venture had been his dream job. He says he loved it so much that if society required everyone to work for free, or for the same salary, he would still have chosen that job.

But that chapter is done. He was moved by Arthur Brooks’s Strength to Strength, which addresses the next stage of life. Gurley now wants to take the techniques that worked for him in venture — blogging, understanding problems, synthesizing — and apply them to larger social problems. His phrase is to “dent the universe a little bit” in that next arena.