Victor Lazarte on Founders, AI and Investing Through a Technology Shift

Benchmark general partner Victor Lazarte speaks with Harry Stebbings about building Wildlife Studios, backing founders and the effects he expects from AI. Lazarte invested in Brex and says he backed Mercor and HeyGen; his assessments of these companies and the market should be read with those investment interests in mind. The interview's figures and forecasts are not independently verified.

Lazarte says he and his brother bootstrapped their mobile games business in Brazil, but not by choice: they tried hard to raise money. Around 2011, he says, the local venture scene was less developed and he lacked business sophistication. The best offer he recalls was $50,000 for half the company, which they rejected.

The brothers loved games and looked for evidence of a growing market. Lazarte recalls that mobile app downloads had grown approximately 42 times from 2008 to 2009; they decided to make mobile games and built the first titles themselves. The figure is his recollection, not a verified market series.

His lesson on timing is to study what is working now rather than rely exclusively on predicting the future. He points to earlier internet and mobile adoption and estimates that LLM usage, measured in compute, grew around 100-fold in two years. He calls that estimate imprecise but directionally useful.

Lazarte sees the spread of LLM inference as a major opportunity to start companies. He points to ChatGPT, Character.AI and Cursor as examples of products with substantial usage, and suggests looking for new businesses adjacent to products people already love rather than requiring an entirely unprecedented category.

His favorite hypothetical company would build an AI companion. He predicts a future company may relate to Character.AI as Facebook did to earlier social networks: an important successor to a pioneering product. That analogy is a forecast, not a judgment that any particular successor will emerge.

He divides the agent transition into incumbents that adapt, incumbents displaced by agent-first challengers, and businesses made possible by the new interface. In his view, an agent could allow customers to interact with a business as though speaking with an owner who knows its products and can solve problems.

Stebbings asks whether an AI companion would effectively remove other humans from social life and raises concern about mental health. Lazarte sees a potential digital friend that is available continuously. He argues that wealth has diminishing effects on happiness and that the depth of relationships matters, but the interview does not establish that an AI companion improves mental health.

Asked whether social media already deepens relationships, Lazarte says Instagram helped him meet his wife and is positive, but he thinks it only scratches the surface. He argues friends help people shape a narrative about themselves and feel understood. He predicts that in five years an AI may understand most people better than any person does.

Stebbings asks whether that is a world Lazarte wants for children. Lazarte believes AI companions could reduce isolation and help people connect to others, rather than replace in-person friends. His proposed system would know users well and introduce compatible people; those benefits and safeguards are a vision, not demonstrated outcomes.

Lazarte recalls meeting Pedro Franceschi while building his own business in Brazil. He says Pedro had lost control of an earlier payments company after an investor agreement, and Lazarte helped him work with lawyers despite holding no shares. He encouraged Pedro to build in the US and later invested in Brex's seed round and joined its board.

He says knowing Brex influenced his decision to back Mercor: its founder Brandon reminded him of the Brex founders. Lazarte looks for two traits he finds unusual together, open-mindedness and disagreeableness: a founder listens with curiosity but can still confidently disagree. These are his selection criteria, not a tested formula for success.

Another question he borrows from Yuri Milner is how founders spend an entire day, especially their free time. He recalls Pedro examining iPhone network traffic and finding a jailbreak out of technical curiosity, and Brandon reading extensively about business and Bill Gurley. Lazarte believes such intrinsic interest can compound over time.

He says Brex reached $100 million in revenue 18 months after launch and Mercor grew from $1 million in annual revenue when he invested to over $100 million in 11 months. These are his descriptions of growth; the conversation does not provide complete accounting periods or comparable revenue definitions.

Stebbings asks whether younger founders require a different kind of investor or board member. Lazarte says yes and emphasizes access to experience: he connects Mercor's young founder with Pedro, who had faced similar challenges running a fast-growing company in his early twenties. A board member needs both relevant experience and enough company context to apply it well.

Lazarte attributes Silicon Valley's strength to accumulated company-building knowledge and networks. He suggests that a founder on the US West Coast is roughly a thousand times more likely to build a tech giant, using a rough population and company comparison. The interview does not establish a rigorous probability or control for other factors.

He says he meets roughly five to ten new companies a week, with only around one arriving through raw inbound contact. Most opportunities he pursues come from conversations with knowledgeable people and sector research. Of four companies he had worked on since joining Benchmark, he says three came through that process and one was a company he co-founded that remained in stealth.

Stebbings asks whether exceptionally fast growth has broken the old SaaS playbook. Lazarte says simple rules that once treated reaching a revenue threshold as evidence of category leadership have become less reliable: much AI revenue can be experimental, and a thin workflow around a model may lose value as the underlying model improves.

His first test for an AI company is whether better models make the business stronger or weaker. Revenue growth still deserves attention, he says, but he wants a product that gains value from model progress. This is an investment heuristic, not proof that any company's moat will endure.

Lazarte calls replacing knowledge workers an exciting opportunity and predicts AI could interview candidates better than humans. As an investor in Mercor, he sees potential advantages in improving interviews, learning from subsequent job performance and building a marketplace. These are his commercial judgments, not independently established measures of hiring accuracy or fairness.

Stebbings raises criticism of Mercor's revenue claims. Lazarte says the activity generates revenue but is not annual recurring revenue backed by annual contracts; the figure is a run rate. He argues demand from AI labs for expert evaluators may persist while humans remain better than models at some knowledge tasks, though that durability is uncertain.

He describes a training process in which experts create difficult questions and grading rubrics, models generate responses, and the resulting evaluations support reinforcement learning. In his account, finding qualified experts is difficult and Mercor helps labs do that. His description is a simplified account of one approach, not a complete characterization of all model training.

Stebbings asks whether investors are adequately compensated for less durable revenue and high AI valuations. Lazarte says pricing must be assessed case by case and calls the particular HeyGen and Mercor rounds he invested in cheap. When challenged on a later $2 billion valuation, he clarifies that he means his own earlier round, not necessarily that valuation.

Lazarte predicts that a company founded within the next three years will eventually be worth $1 trillion and believes a basket of today's AI winners will work on average. He also concedes that some apparent winners will prove overpriced. His reasoning is that AI could target labor spending as well as software budgets; the outcome and returns are speculative.

Stebbings worries that small teams building hugely valuable companies could replace labor spending and widen inequality. Lazarte agrees the transition could be destabilizing and estimates that only about 1% of current computer-based knowledge work may remain in ten years. This is an extreme personal forecast, not an established employment estimate.

Lazarte disputes claims that AI will only augment workers, arguing it will replace people. He nevertheless expects society overall to benefit from more productive companies, while acknowledging a serious problem in distributing the wealth created. Neither the speed of displacement nor the distribution of gains is established by the interview.

Stebbings argues that universal basic income would not resolve people's need for purpose. Lazarte counters that extreme concentration of AI wealth could force redistribution through democratic politics; he imagines citizens using AI assistants to assess politicians. Both arguments are speculative, including his assumption that such systems will reliably identify voters' interests.

Lazarte views competition with China as a force that could focus the US amid internal disagreements, saying he wants powerful AI controlled by the US. Stebbings argues China has advantages in education, infrastructure and talent; Lazarte responds that US networks and current AI products still give it a lead. Their geopolitical assessments are opinions, not verified rankings or evidence of hidden capabilities.

Asked about adoption timelines, Lazarte accepts the general warning that people can overestimate one-year change and underestimate ten-year change. Even so, he predicts meaningful changes in daily life within five years, citing the rapid growth of ChatGPT as evidence of adoption. This is a forecast rather than a guaranteed timetable.

He imagines an app that gradually earns enough trust to tell people what to do each day, starting with routine support such as exercise or medication reminders. Eventually, he predicts, users may follow more consequential recommendations. The scenario is his vision, not evidence that blindly following an AI system is safe or beneficial.

Stebbings suggests students choose hands-on professions such as dentistry or surgery for job security. Lazarte instead recommends computer science, even to people who may not become developers: he believes understanding the technology, logical reasoning and breaking large problems into smaller ones will remain useful. Neither career prediction guarantees future employment.

Lazarte says Benchmark participated in a Cursor round. Asked whether the firm must lead a particular early round, he rejects rigid stage rules and says the aim is to become a close partner to important founders. Stebbings observes that many investors make similar claims and describes wanting to be the person a founder calls when things go badly.

Lazarte argues Benchmark makes few investments partly so partners can spend more time with each company. He says an outside investor rarely changes a company's trajectory easily, but can be a thought partner and connect founders to expertise. Years of board involvement can build the context needed to help with important hires and decisions.

Small participations often lack that context, he says, though they may sometimes begin a relationship that grows. At Benchmark he was told to help build the firm rather than preserve every tradition. He believes a small team and light process can make adaptation easier as venture investing changes.

Lazarte describes a shift from board oversight focused on avoiding losses toward supporting ambitious founders and maximizing upside. Stebbings challenges him: a director's fiduciary duty to shareholders is not always identical to a founder's ambition. Lazarte replies that he underwrites the founder when investing and generally tries to enable the founder even when he disagrees; that is his philosophy, not a statement of law.

Asked whether he has lost faith in a founder, Lazarte says some of his angel investments have gone badly but describes learning from Peter Fenton to remain committed while a founder wants to keep trying. He recalls a struggling company where the response was to increase board meeting frequency rather than abandon it.

He describes Benchmark's investment process as lightweight: partners must present an idea and seek colleagues' opinions, but do not need their approval. A vote from one to ten, without five, acts as a temperature gauge rather than a veto. The partner leading an investment can proceed despite disagreement.

Looking ten years ahead, Lazarte wants Benchmark to be involved in many of the companies he predicts could reach trillion-dollar value. He thinks the firm should widen its view of what a Benchmark round looks like: a capable founder pursuing a large market might raise $5 million or $200 million. Those outcomes are projections, not a portfolio guarantee.

His first Benchmark investment, he says, was a $55 million check into HeyGen from a $600 million fund. He argues the nominal fund size is less constraining than the firm's ability to help founders early, and says larger checks can sometimes use special-purpose vehicles. This is his view of portfolio construction and a disclosure of the firm's interest in HeyGen.

Discussing Sierra founder Bret Taylor, Lazarte says Taylor valued a relationship with Peter Fenton and met the partners in a conversational setting. He believes strong founders benefit from assembling the right people around them early. The example illustrates his view of partnership rather than proving a particular investment outcome.

In the quick-fire round, Lazarte praises Fenton's ability to understand people and encourage them to open up. He recalls a dinner after which he felt he knew a longtime acquaintance much better, and says such conversations help build founder relationships.

Asked to pick one public stock for ten years, Lazarte chooses Duolingo. He predicts its language product could become a free AI tutor for more subjects and a valuable company. That is his personal investment thesis and forecast, not a recommendation or a verified product roadmap.

For funds, he says he is an investor in GreenOaks and chooses Conviction for early-stage investing. He closes with the ambition to be an important part of companies working on education, security and companionship, whether as a board member or co-founder. His claims about AI solving major problems are hopes for the future, not established results.