a16z Frontier: Andreessen, Horowitz and Dixon on What to Fund Now

Andreessen Horowitz is turning its crypto fund into a16z Frontier, and Marc Andreessen, Ben Horowitz and Chris Dixon explain why AI makes ambitious startups the safer bet. In this hour with host Erik Torenberg, recorded for the a16z channel on October 22, 2026, the three partners cover the widening gap between AI-native founders and incumbents, why open source is the guard against an AI monopoly, how Google’s deal with the open web is coming apart, and what an 18-year-old should study now.

TLDW

a16z crypto becomes a16z Frontier: a broader mandate, led by Chris Dixon, for “outer planet” founders who do not fit the firm’s existing verticals. The partners argue that AI is pulling forward decades of technical progress, so the question for founders has flipped from “is this too ambitious?” to “is this ambitious enough?” They make the case for open source and open-weight models as the alternative to AI that centralizes into one or two companies or a government, describe a coming internet where agents outnumber people, and say the old bargain between search engines and websites is breaking. They close on education: the kids who treat AI as normal will be the ones who build what comes next.

Thoughts

Dixon’s “internet covenant” section is the part of this conversation with the most consequence, and it gets only three minutes. The deal he describes is simple: a website lets Google index it and show a snippet, and Google sends back clicks that the site turns into ad revenue or sales. AI answers break that deal, because the reader gets the answer on Google’s page or in a chatbot and never leaves. Dixon points to Stack Overflow, whose traffic collapsed once AI replacements arrived, and predicts a decade-long “great reconfiguration” of how attention, money and knowledge move online. He is right that this is bigger than mobile, which mostly added users to the existing web. What he leaves open is who pays the people who write the things AI reads. An agent can book the travel and fetch the recipe, but if the travel guide and the recipe site lose their income, the supply of fresh, trustworthy material shrinks. Any startup that answers that question, with payments between agents and publishers or some new kind of licensing, is working on the internet’s economic plumbing, not a feature.

The open source argument is more interesting than the usual “freedom versus safety” framing because Horowitz turns it into a safety argument. He names reward hacking, where a model does something it was never asked to do because it helps it earn its reward, as probably the biggest safety problem in AI, and says the labs have not worked on it very hard because they are busy racing each other. His fix is open weights: let every PhD student, professor and undergraduate inspect the model and try to solve it. Andreessen adds the parent’s version: if the only way your kid can learn how AI works is to get a job at one of two companies, something has gone wrong. There is an obvious counterargument they do not engage, that open weights also give bad actors the same capabilities. But the strongest line here is the structural one. Without open source, Andreessen says, AI “inevitably centralizes” down to a couple of companies or the state, and his test for how you should feel about that is to picture “government sushi” and then “government AI.”

The most practical observation for anyone inside a big company is Andreessen’s claim that many firms are not adopting AI while their employees quietly are, and are not telling management. If that is true, the usual enterprise adoption numbers (Andreessen cites a figure of about 3% of measured enterprise use) undercount what is actually happening, and the people learning fastest are the ones most likely to leave and start something. That is the mechanism behind his prediction that startup formation “is going to skyrocket,” and it explains why the partners think Clayton Christensen’s The Innovator's Dilemma is the wrong lens. Disruption theory pictures companies in one industry fighting each other. The pattern they describe is AI-fluent founders walking into industries that never thought of themselves as tech, which turns the contest into “ten big tech companies and all the startups” against the rest of the Fortune 500.

The hardware section deserves more attention than it got. Andreessen describes an open source tool that points an AI model at any device, reverse engineers it and writes new firmware, so an annoying thermostat can be taken over and reprogrammed. One of the partners says a 30-person team can now credibly build a full-stack device company that would have needed a thousand people a decade ago, and Dixon notes that Claude will design a 3D-printed case for a homebrew gadget. Put together, that means software people can now do things in hardware that were too hard or too expensive before. It also explains the “too early” reframe at the top of the episode. Dixon’s view that every good technology idea eventually works is partly survivorship bias from investors who were early on the winners, but his narrower point holds: if AI pulls engineering timelines forward by years, ideas filed as science projects in 2016 need to be looked at again.

Dixon’s last frontier idea is the cleanest prediction in the episode: AI is still in “single player mode.” Almost everything people do with it is one person typing at a computer and reading the reply, the way the first iPhone apps were flashlights before Uber and Instagram built networks. He expects a multiplayer phase made of agent fleets that solve hard problems together, trust networks between people’s agents, and stablecoins as the natural way for agents to pay each other. Combine that with his earlier claim that agents will become the most numerous participants online, and the internet covenant question comes back: the next big network may be one where most of the users are not human.

Key Takeaways

a16z crypto is now a16z Frontier, run by Chris Dixon, with a mandate to back “outer planet” founders whose ideas fall outside the firm’s existing verticals. Dixon’s team brings cryptographers, math professors and a policy operation built during the crypto years.
Crypto has blurred into mainstream tech: stablecoins are a normal tool for Stripe and most fintechs, and Wall Street firms are tokenizing stocks and other assets.
Horowitz credits Dixon with the firm’s framework of looking for “good ideas that look like bad ideas,” where the founder holds a secret that makes the bad-looking idea good.
Dixon says Frontier will err on the side of being too early, and that AI may pull forward 20 years of progress, so “science project versus engineering project” judgments need to be redone.
Andreessen says the gap between what AI-native startups and typical incumbents think is possible is the widest of his career, and it is widening week by week.
Many companies are not adopting AI while their employees are, often without telling management. AI is easier for individuals to adopt than for institutions.
Andreessen criticizes “product manager brain”: AI products that only book restaurants and flights for people who look like the product managers who built them. He expects AI to be far more entertaining and creative than that.
Video games are leading the shift, with AI tools that port old games and mash two games into one. Horowitz predicts the AI-native film will be a new medium, not a cheaper movie.
AI makes hardware approachable for software people: reverse engineering devices, writing firmware, designing for manufacturing, and AI-driven home automation.
Dixon says open source models are growing faster than closed ones in token consumption, and that local models running on personal computers are the next step.
Andreessen says the current US administration is fully behind American open source AI and China is pushing open models too, while the EU still wants to restrict them.
Horowitz argues open weights would let universities attack reward hacking, which he calls probably the biggest safety problem in AI.
Dixon says the internet’s covenant (index my site, send me clicks) is breaking as Google and chatbots answer directly, and expects a decade-long reconfiguration of the web’s economics.
For an 18-year-old, Dixon recommends studying “modern alchemy” (math, computer science, maybe economics), because humans just learned to turn sand into intelligence. Andreessen adds classics, economics, biology and even zoology as fields AI will remake.

Chapters

0:49 a16z Crypto Becomes a16z Frontier

Dixon announces the expanded fund and gives two reasons: crypto has merged into fintech and Wall Street, and the firm’s other teams are too busy to cover the strange edges. He traces his own run of “second order” bets from 2013, when cheap smartphone parts made drones and Oculus possible. Horowitz says Frontier fills a hole in the firm’s structure, and Andreessen praises the depth of Dixon’s technical team. Horowitz also tells the story behind the firm’s J.P. Morgan motto about doing first-class business in a first-class way.

15:15 Is Being Too Early Still a Risk?

Dixon says nearly every good technology idea eventually worked, often after a precursor failed ten years too early. VR is his example: 13 years after Oculus, still not mainstream. He argues AI, now improving fast at software engineering and math, may compress decades of technical risk. The partners add that small teams can now build full-stack device companies.

18:18 The Widest Gap Between Early Adopters and Everyone Else

Andreessen describes the gap between AI-native founders and incumbents as the largest of his career. He and Horowitz imagine what they could have built at Netscape in 1994 with today’s AI. The partners say they now ask whether an idea is ambitious enough, because a moderate plan will not stand out, and Andreessen warns incumbents that the clock is ticking.

22:06 Will AI Favor Incumbents or Startups?

Andreessen says the incumbents adapting best are the ones building AI directly, and that employees are adopting AI faster than their employers. The partners reject the Christensen frame of companies fighting within one industry and describe tech-fluent founders moving into the whole economy. Dixon compares AI today to the 1990s web, when companies put old behaviors online before YouTube bet on new ones.

28:33 AI Social Networks and Agents Online

Torenberg asks whether AI will produce a new wave of social networks. Dixon predicts that agents will become the most numerous participants on the internet, needing their own ways to communicate, transact and share knowledge, and the group trades examples of early agent social experiments.

30:08 Beyond Product Manager Brain: Games, Film and Hardware

Andreessen mocks AI products designed by product managers for product managers and predicts a far more fun and lucrative wave. Dixon describes AI tools that port and mash up video games. Andreessen asks what a $180 million film could have earned on a $2 million AI budget, then describes reverse engineering hardware with AI. Horowitz says AI film will become a new medium, and Dixon compares it to the camera making motion pictures possible.

38:22 Is AI Centralizing? The Case for Open Source

Torenberg raises Peter Thiel’s line that crypto is libertarian and AI is communist. Dixon points to open models gaining share, local models, and stablecoins for agent payments. Andreessen praises US and Chinese support for open source, describes wilderness explorers carrying local models on their phones, and calls open source the guarantee against government AI. Horowitz argues open weights are the way to get the world working on reward hacking.

46:37 The Internet’s Broken Covenant

Dixon explains the old deal between distributors like Google and Facebook and the sites that supplied content, and why direct AI answers are breaking it. He expects painful losses for businesses built on the human attention economy, and a chance for founders to redesign how the internet’s economics work.

49:12 What Comes After the iPhone

Dixon expects wearables, a VR comeback driven by AI-made content, brain-computer interfaces and a revival of homebrew hardware hacking. Andreessen describes a building designed by AI and models that can produce design-for-manufacturing specs. Dixon then argues AI is still in single player mode and that a multiplayer phase of agent fleets and trust networks is coming.

53:59 What to Study If You’re 18 Today

Dixon suggests math, computer science and economics. Andreessen lists fields AI is reshaping, from classics (the Vesuvius scrolls) to biology and animal communication, and retells Douglas Adams’s three rules for how each age group receives new technology. Horowitz says the adults, not the kids, are the ones who should worry. Dixon closes by calling the firm’s stance “humble definite optimism.”

Notable Quotes

“I think that that gap is as wide as I’ve ever seen in my career.”
Marc Andreessen, on the gap between AI-native startups and incumbents

“I think many companies are not adopting AI while their employees are.”
Marc Andreessen, on why startup formation is about to surge

“I think the AI film is not going to be a movie. It’s going to be something very, very different.”
Ben Horowitz, on AI as a new medium rather than a cheaper way to make old ones

“One of the really underestimated things about AI is AI will teach you about AI.”
Marc Andreessen, on why open models matter for education

“The one thing that we know has always led to badness in the world is a massive concentration of power.”
Ben Horowitz, making the case for open source AI

“Give Google a snippet, let them index you, and then they’ll send you some clicks.”
Chris Dixon, describing the internet covenant that AI answers are breaking

“Humans just figured out how to turn sand into intelligence.”
Chris Dixon, on what an 18-year-old should study, crediting the framing to Andreessen

Watch the full conversation here.