Benchmark GP, Victor Lazarte: The 3 Traits All the Best Founders Have

Summary

• Rule-based SaaS investing is dead. Lazarte’s claim: “SaaS was a moment in time where rule-based investing worked really well” — $10M revenue no longer marks a category winner because much AI revenue is “experimental,” thin workflows wrapped around ChatGPT. Benchmark’s replacement test is a single question: “if models get a lot better, is your company worse or is your company better?” — model-better-worse companies are “very hard to touch.”
• The bubble answer is to own the winners anyway. “In the next 3 years I’m sure someone will start a company that’s going to be worth a trillion dollars.” US software spend is ~$1T but labor spend is ~$10T, and AI companies are going after the labor line — so “on average, buying the winners today will work out,” even though “in retrospect” some winners (he cites likely Anthropic at 60 and Perplexity at 15 as the reference points) will prove overpriced.
• Big-company leaders are lying about augmentation. Leaders saying AI augments rather than replaces people — “this is [bleep], it’s fully replacing people.” Lazarte thinks probably ~1% of knowledge work as we know it exists in 10 years, calls replacing knowledge workers “the most exciting opportunity in venture right now,” and still argues it’s net fantastic for society — the hard problem is wealth distribution, not job loss.
• The AI companion is the company he’d most love to build: “Character AI was Friendster” — the precursor, not the winner. His prediction: “five years from now, for the majority of people, the person that will understand you the most is going to be an AI,” attacking the biggest happiness driver — relationship quality — that technology has not attacked.
• Mercor is a bet on the pure-RL era, not a recruiting wrapper. He invested at $1M revenue; it passed $100M in 11 months. On the revenue-quality debate: “it’s revenue for sure” but “no, it’s not” ARR — a run rate. The defensibility: pure RL needs deep experts writing questions and grading rubrics, and “for as long as humans are better than computers at any knowledge task, you’re going to need people to create the evals.”
• Portfolio construction isn’t the constraint. Benchmark’s LPs are the same across funds — “one $3 billion fund or five $600 million funds, for LPs it’s actually the same” — so fund size was never the constraint. His inaugural check was $55M into HeyGen out of a $600M fund (~9% in one shot); the real constraint is context, which is why Benchmark does few deals.
• China is a stabilizing force for the US: the external threat suppresses internal conflict at the moment AI’s wealth concentration plus AI-informed populist voters would otherwise destabilize democracy. Stebbings pushes back hard — “China are absolutely crushing the US,” TikTok is a data weapon — but Lazarte holds: “by all measures the US is winning… ChatGPT still is the best product out there.”

Deep dive

1. Bootstrapped by necessity: $100 in Brazil to a billion users

• Lazarte’s gaming company (with his brother, 2011 Brazil) “makes for a good story, but the reality is we tried really hard to raise money. We didn’t bootstrap by choice.” The best term sheet on offer: $50,000 for half the company. Less sophisticated friends fared worse — one agreed to a 20x liquidation preference he didn’t understand and lost his entire company.
• The founding insight was pure present-tense pattern-reading: mobile app downloads had grown 42x from 2008 to 2009, they loved games, so they made mobile games — eventually growing to “over a billion users” with no employees on the first titles.

2. Don’t predict the future — read the present, and the present is a 100x

• His core heuristic: “it turns out it’s very hard to predict the future. It’s much easier to understand the present.” The Bezos template — seeing the internet growing “23 times a year” at D.E. Shaw — and today’s equivalent: LLM usage has grown ~100x in two years, “from 1x of FLOPs to 100x of FLOPs… not very precise but directionally correct.”
• The method is adjacency: a year and a half ago the things working were ChatGPT and Character AI; now ChatGPT, Character AI, and Cursor. “A lot of times the great opportunities are adjacent to things that are working today” — Facebook followed MySpace and Friendster, and the biggest mobile-era winners (Uber, Instagram) all launched into the same visible wave. His verdict: “it’s the best time in over a decade to start a company.”

3. “Character AI was Friendster” — the agent interface remakes everything

• The company he’d “most love to work on” is an AI companion, and his framing is that Character AI is this cycle’s Friendster — the precursor to something very important. The structural argument: businesses interfaced through branches, then websites, then mobile apps, and now agents — customers talking to something “that feels like a person that knows everything about the business.”
• Every UI shift of this magnitude produces three buckets: incumbents that adapt; incumbents that don’t, with startups recreating today’s successful businesses agent-first; and businesses only possible on the agent interface. The social version, per Stebbings’ blunt gloss: “a 24-hour-a-day constant companion that understands how you feel… your continuous on-demand friend.”

4. Technology has never attacked the biggest driver of happiness

• Lazarte’s chain: wealth raises happiness only log-linearly, tapering around $75,000/year for a US household, and technology mostly makes us wealthier — while “the biggest predictor of your happiness is the quality and depth of your relationships. And technology has not attacked that.” His existence proof that an unseen entity can be a load-bearing relationship: religious people are happier partly because “God is this friend that you never see… and that makes you that much happier.”
• Stebbings’ counter-thesis, worth keeping verbatim in spirit: the death of religion is the CPG thesis — “we’ve shifted cult-like worshipping from God, Buddha… to Hyrox and CrossFit and Taylor Swift and Lululemon,” communities that consolidate belief the way church did. And his darker stat: young men are now “more likely to die of suicide than cancer,” yet the largest mental-health company is Calm at ~$1.5B.
• Lazarte rejects the dystopia charge: the AI friend won’t replace humans, it will broker them — “I know you really well, and there’s a hundred million people that I know really well, and I’m going to connect you guys,” with reputational enforcement: misbehave and “your AI friend is not going to introduce you to anyone anymore.” On his kids growing up in that world: “I think it’s going to be wonderful.”

5. Open-minded but disagreeable — and the free-time test

• The two traits he hunts for “rarely come together”: founders who are very open-minded but very disagreeable — genuinely curious through your whole argument, then “no, I actually think the opposite is true… to the point that it’s going to upset you.” Both Pedro (Brex) and Brandon (Mercor) had it to a high degree.
• His second filter is stolen from Yuri Milner: walk through the founder’s entire day — “what time you wake up, what is the first thing you do… what do you do in bed.” Pedro’s free time was spent tracking the packets iPhones send to servers, which is how he found one of the iPhone jailbreaks — for no reason and no money. Brandon at 21 had read all of Bill Gurley’s blog posts and podcasts “for no specific reason.” Intrinsic obsession in the young “compounds really well over time.”
• The Brex origin is a bond of “shared distrust of investors”: Pedro built a payments company at 16 in Brazil, signed a deal that cost him control, and Lazarte spent a ton of time with lawyers helping him escape — holding no shares. The payoff: seed check and first board seat at Brex, which went zero to $100M revenue in 18 months; Mercor did $1M to over $100M in 11 months.

6. SaaS rules are dead; one question replaces the spreadsheet

• “SaaS was a moment in time where rule-based investing worked really well… you get to $10 million in revenue, you’re a category winner. This is no longer true.” The reason is revenue nature: it’s now trivially easy to wrap “a very thin workflow around ChatGPT,” sell it to every law firm writing demand letters, and print millions with “just no enterprise value” — because as models get better, the workflow gets less valuable.
• The pitch-room test: revenue growth still earns attention (“it’s not worth what it was worth before”), then the first question — “if models get a lot better, is your company worse or is your company better?” Worse: “very hard to touch it.” Better: “that’s a great place to be in.”

7. Mercor, pure RL, and the lie about augmentation

• The unvarnished version of his favorite theme: big-company leaders claiming “AI is not replacing people, AI is augmenting people’s abilities — this is [bleep]. It’s fully replacing people.” Stebbings’ favorite corporate euphemism: “we’re just not hiring new, but we’re keeping everything as is.” Lazarte calls replacing knowledge workers “the most exciting opportunity in venture right now” — and, he insists, ultimately fantastic for humanity.
• Mercor cleared his models-get-better test on three axes: interviewing is a problem with continuing gains to quality (even Musk and Bezos spend enormous time on it), hiring outcomes provide specific data to improve the model, and the platform is a marketplace with network effects.
• The deeper thesis is the third era of models: pretraining, then RLHF (“that was the big ChatGPT innovation”), now pure RL — a deep expert writes a question the model can’t answer plus a grading rubric, the model generates masses of graded answers, and “if you’re able to create a benchmark for any task, through reinforcement learning and throwing a lot of compute at it, you’re able to create a model that surpasses that benchmark.” Finding those experts is interviewing — Mercor’s actual product for the labs. On the ARR criticism: “it’s revenue for sure,” but “no, it’s not” ARR — a run rate; the hedge-turned-defense: “for as long as humans are better than computers at any knowledge task, you’re going to need people to create the evals.”

8. The bubble question: are you paid for the risk?

• Stebbings presses the buyside worry directly: with run-rate revenue replacing enterprise ARR and “experimental” spend behind names like HeyGen, are investors paid for the risk at AI’s premiums? Lazarte: case by case — HeyGen and Mercor “were very cheap rounds” (Stebbings: “You think two billion was cheap?” — Lazarte clarifies he means the rounds he did).
• The asset-class answer: AI is “much bigger than mobile, maybe the order of magnitude of the internet.” Smart investors see $1T of US software spend versus $10T of labor spend now in scope, so outcomes will be “insanely larger than they were before.” Conclusion, hedged exactly as delivered: “I think on average buying the winners today will work out. Are there some winners that are overpriced? In retrospect, the answer will be yes. Absolutely.” Stebbings’ memento mori: Gurley passing on Google on price.

9. Probably 1% of knowledge work survives — and China keeps America honest

• Asked what share of behind-a-computer work exists in 10 years: “probably like 1%.” His worry is the college graduate — “if you’re coming out of law school, what are the things you can do that a model won’t be able to do in three years?” — while owners of shares get “way richer” and trillion-dollar companies run on tiny teams: “a very destabilizing force.”
• Still net-positive: “companies are humanity’s best invention,” and the difference between growing up in Brazil and living in the Bay Area “is the quality of the companies.” Stebbings rejects UBI outright — “humanity requires purpose… we will not paint and write poetry,” pointing to addiction, gambling, prostitution as how psychology actually breaks. Lazarte’s own destabilization mechanism: everyone will carry a powerful AI that tells them whether politicians serve them, so populist redistribution agendas “will get votes.”
• His resolution is geopolitical: “China is actually a very stabilizing force for the US” — the external threat means “we cannot afford to waste time on internal conflict” while someone gets to powerful AI first, “and it needs to be the US.” Stebbings disagrees on the scoreboard: China is “absolutely crushing” it on education, talent depth, work ethic (“Europe’s not even in the [bleep] race”), TikTok is “a weapon of consumer data aggregation.” Lazarte concedes more people working harder, but: “the true network effect is a city” — Silicon Valley’s knowledge density is why a West Coast founder is roughly a thousand times more likely to build a tech giant (70% of the largest tech companies from ~0.1% of world population), and “by all measures the US is winning.”

10. Inside Benchmark: no process, no portfolio construction, unconditional belief

• “I think rigidity is nonsense” on must-lead-the-A, and portfolio construction gets the same treatment: LPs are identical across funds, so “one $3 billion fund or five $600 million funds — for LPs it’s actually the same.” His first check: $55M into HeyGen from a $600M fund. The true constraint is context — Benchmark does few deals “because getting context on the companies takes a bunch of time,” and “if you write a small check and you’re one of many, you don’t make any difference. It doesn’t matter how smart you are.”
• Stebbings’ best pushback of the episode: every VC claims first-call status — “your mom should be your first call. I want to be your worst call — when the [bleep] really hits the fan, call me.” Lazarte’s version of commitment, learned from Peter: on a struggling company, “completely uneconomical,” the move was “we should double the frequency of the board meetings” — “we’re here for as long as the founder wants to try.”
• Mechanics worth stealing: any partner can do a deal regardless of the vote — partners vote 1 to 10 and “you can’t vote 5,” purely as a temperature gauge that forces a number. The board role has flipped from governance to “amplifying the founder’s ambition”: “my fiduciary duty comes in at the moment I write the check — I’m underwriting that person.” Brett Taylor’s Sierra round was “more like a conversation than a pitch” — dinner with the partners, a founder who could self-fund choosing Peter anyway.
• Quickfire signal: he’s done four deals in ~18 months at Benchmark, three sourced outbound, and the fourth is a stealth company he co-founded. Ten-year buy-and-hold: Duolingo — “sneaky in a good way,” a language app becoming “your AI friend that teaches languages,” with the founder “a total genius” on track to build “a free AI tutor for everyone.” Funds: Green Oaks for growth, Conviction for early. And the belief most people reject: soon “we’re going to wake up and do whatever the app tells us to do… we’re going to be obedient to machines and we’re going to love it.”