The Frontier Ideas a16z Wants to Fund Now | Marc Andreessen, Ben Horowitz & Chris Dixon

Summary

• The ACC Crypto Fund is expanding into “a16z Frontier,” a mandate for what Chris Dixon’s partner Martine calls “outer planets” — categories too new or weird to fit the firm’s verticals. The stated logic: crypto’s lines have blurred into fintech and Wall Street (“Is Stripe a crypto company?”), while the vertical structure left a coverage gap for the strangest founders. Ben Horowitz’s framing carries the thesis: “more than categories, more than products, we invest in entrepreneurs, and so many of the most interesting entrepreneurs are on the outer planets.”
• Marc Andreessen’s central claim is that the gap between what AI early adopters know is possible and what incumbents understand “is as wide as I’ve ever seen in my career” — and it’s widening as AI capabilities race ahead week by week. His corollary: employees are adopting AI while their companies aren’t, those employees will leave to start companies, and “startup formation is gonna skyrocket from here… by the way, it already is.” His message to incumbents: “the clock is ticking.”
• Marc says the firm’s underwriting question has inverted: “We used to ask… is this idea too ambitious? We’re much more likely to ask, ‘Is this ambitious enough?’” Dixon believes AI may be “pulling forward 20 years” of technological progress, forcing a re-evaluation of technical risk and the “science project vs. engineering project” frame — a 30-person team can now credibly build a full-stack device company that might have taken 1,000 people a decade ago.
• Dixon says the Christensen disruption frame is probably the wrong one right now: this isn’t tech companies fighting each other, it’s “people who are technologically modern going out into other industries and just changing the economy more broadly.” Meanwhile the internet’s “covenant” — content sites give Google a snippet, Google sends clicks — is breaking as AI answers replace links; Stack Overflow traffic has dropped precipitously, setting up a decade-long “great reconfiguration” of internet economics.
• On market structure, the panel considers Peter Thiel’s “crypto is libertarian, AI is communist” framing: Dixon believes open source is growing faster than closed models in token consumption, local models may come, and stablecoins are a natural way for agents to pay each other. Marc says the US administration is “100% pro” open source and the Chinese Communist Party’s enthusiasm surprised him; without open source, AI “inevitably centralizes down to one or two companies or maybe just flat out the government” — “picture government sushi… then picture government AI.”
• The native-use-case explosion is already visible in the panel’s favorite recent examples: AI mashing up any two video games (Super Mario and GTA), Epic saying “no more humans” writing code, and an open-source tool called REA — “Reverse Engineer Anything” — that can reverse-engineer and reprogram hardware. Ben’s Digger example: the film had a $180M production budget and a $200M cash loss, but “on a $2 million budget it would have been insanely profitable.” Ben predicts AI film will be an entirely new medium, not just a cheaper movie.
• Dixon’s structural prediction: “the most numerous participants on the internet will be agents, not people,” and AI is still mostly in single-player mode. He sees swarms (he recalls one solving a math theorem; the Hugging Face hack was the negative version), trust networks emerging in products like Muse and Instinct, and a possible “multiplayer phase of AI” as fertile ground for founders.
• Career advice for an 18-year-old: study “modern alchemy” — “humans just figured out how to turn sand into intelligence” — via math, computer science, maybe economics. Marc deploys Douglas Adams’ three-cohort rule (under 15 it’s natural, 15–35 it’s new and potentially a career, over 35 it’s “unholy and evil”), and Ben flips the hand-wringing: “you’re the one with the problem because you’re obsolete. The kids are going to be superhumans.” Erik’s closing amendment to the firm’s posture: “humble, definite optimism” — back entrepreneurs with their own conviction.

Deep dive

1. The announcement: ACC Crypto Fund expands to a16z Frontier

• Dixon’s two motivations: first, crypto has partly won and the lines have blurred — stablecoins are “a normal part of the tech landscape” used by Stripe and considered part of nearly every fintech company’s toolkit; Wall Street is moving into tokenized stocks — so “is Stripe a crypto company?” is no longer an easy question. Second, the firm’s verticals are busy with core opportunities, leaving “outer planets” (Martine’s metaphor, borrowed from Firefly) falling through the cracks.
• Ben’s architecture point: the vertical strategy works, but the motion for turning a brand-new area into a practice “is a little bit undefined.” Frontier formalizes it — “if that works, maybe that becomes a whole fund at some point.” His bottom line: “more than categories, more than products, we invest in entrepreneurs, and so many of the most interesting entrepreneurs are on the outer planets.”
• Dixon frames it as continuity, not a pivot: he joined in 2013 explicitly to do “this crazy stuff” — drones (Marc was first in, as Dixon recalls), Oculus (“really a reconfigured smartphone”), and Skydio — all second-order effects of the smartphone supply chain. Crypto, for Dixon, was downstream of open source and social media.

2. Why Dixon — and why the crypto team travels well

• Ben on hiring him: a smart blogger with original ideas and real founder experience, whose durable contribution was a framework the firm copied across the firm — “we’re not looking for good ideas that look like good ideas; we’re looking for good ideas that look like bad ideas… where an entrepreneur has a secret that makes it a good idea.” That lens separated massive upside from businesses where “all the profits would get competed away.”
• Marc’s point about the team: crypto’s flash hides that it is “a very deep, complex technology.” Dixon’s group includes math professors and cryptographers, and that technical depth plus market judgment can now be applied to a broader set of topics.
• Dixon adds the policy arm, built “in the crucible of the crypto journey,” as another transferable asset: self-driving cars, drones, and AI itself will have policy angles, so more frontier bets will need regulatory expertise.

3. Culture, J. P. Morgan, and the shit-show caveat

• The J. P. Morgan photo inside Marc’s office door — the banker staring from darkness, with the chair edge lit like a carving knife — is discussed as inspiration. Ben draws out the serious part: Morgan’s courtroom line, “we do first-class business, and then only in a first-class way.” “You’re definitely not going to tell me who I am. I know who I am” — Ben’s argument for why principled identity matters in venture.
• Dixon on the inside view: since joining in 2013, Marc and Ben have never—not once—pushed back on his ambition. “If anything,” they’d ask, “Can you double it? Can you double down? Can you get more aggressive?” And on the outside view that it’s been all ups: “I don’t know of a startup that wasn’t a shit show at some point. Or at multiple points.”

4. Too early is the error they’ll accept — and AI compresses the timeline

• Dixon’s core belief, offered for all three: “basically every good idea we’ve ever seen that involved technology has eventually worked, or will work if it hasn’t yet” — the qualifier being “good idea.” Many things that eventually worked had a precursor that didn’t work and may have been 10 years too early. Timing is the hardest prediction because technology, culture, business, and politics all have to line up. Frontier will “err on the side of being too early” and accept “category risk.”
• His honest specimen: Oculus was sold and acquired, but “it’s now 13 years later” and it hasn’t become as mainstream as they would have liked. Dixon still believes strongly in VR; there were many hard technical problems to solve.
• The AI twist: with software engineering tipping first, then math, then physics and hardware, Dixon thinks “we may be pulling forward 20 years” of things. The whole “science project rather than an engineering project” frame needs re-evaluation, and a 30-person team can now credibly build a full-stack device company that might have required 1,000 people a decade ago.

5. Marc’s gap thesis: the widest early-adopter delta of his career

• Marc’s framing: the delta between what AI-native startups understand is possible and what typical incumbents understand “is as wide as I’ve ever seen in my career” — and it is widening as capability races ahead week by week. Extrapolated: many industries may be up for grabs in a more profound way, as founders gain new capabilities from these tools.
• The counterfactual he keeps returning to: “Had I had Astra at Netscape in 1994… Ben, what would you and I have been able to do at Netscape? We would’ve built the whole damn internet.” They would have gotten crypto done then and had streaming video “a lot sooner.”
• Marc says the underwriting question has changed: “We used to ask, ‘Is this idea too ambitious?’ We’re much more likely to ask, ‘Is this ambitious enough?’” A moderate idea probably won’t be differentiated enough.
• Marc’s adoption asymmetry, flagged as a topic for another episode: “many companies aren’t adopting AI while their employees are” — and employees aren’t necessarily telling management. Some individuals will realize they can use these tools to start a business: “startup formation is going to skyrocket from here. It already is.” He sees both effects so far: some incumbents, including Meta, have adapted, while new companies are also emerging.

6. Christensen may be the wrong frame — and the internet’s covenant is breaking

• Dixon’s correction: the disruption frame assumes a relatively contained technology world where companies fight each other. “That’s probably the wrong frame right now.” The real pattern is “people who are technologically modern going out into other industries and changing the economy more broadly” — full-stack companies like the recently announced Adams investment, building food service and other things, including mining; tech companies and startups competing with the rest of the Fortune 500.
• On where we are in the cycle: “all of the above” is working — the big AI labs, a fast-growing open-source field, and application-layer products. Dixon says measured enterprise use is “something like 3%… or something like that.” “The market for intelligence is very, very large, and it’s only getting started.”
• His separate structural call: the internet’s “covenant” — give Google a snippet, get clicks back — is breaking as answers replace links. Stack Overflow’s traffic drop is an early sign; the transactions will still happen, just not the same way. The next decade could bring a “great reconfiguration” of internet economics, attention, and collaboration — a bigger reshaping than mobile delivered, in Marc’s view.

7. From skeuomorphic to native: agents as the internet’s majority

• Dixon’s pattern: every platform starts with old behaviors done better (ordering a Sears catalog online) before riskier behavioral bets that can change the world (YouTube starting with an empty site and betting that influencers would populate it). AI so far has been the familiar-use-case phase — coding, math, knowledge work, search, and chat — “but we’ve only begun to explore” radical new activities. Dixon thinks we’re at a tipping point.
• His biggest native prediction: “the most numerous participants on the internet will be agents, not people.” Utility proxy agents like Muse and Instinct are an early phase, followed by agents needing their own ways to communicate, transact, and share knowledge.
• Marc’s critique of the current crop — “product manager brain”: product managers designing for product managers, where every demo is “book me a restaurant… with my friends, who are also product managers.” Ben adds that the “helpful, harmless, honest assistant” can mean “the most boring possible thing you could imagine.” What’s coming, Ben says, will be more entertaining, enjoyable, exciting, and lucrative.
• Dixon on multiplayer AI: today’s AI is mostly single-player — you type, it responds. Swarms cut both ways (he recalls a swarm solving a math theorem; the Hugging Face hack was the negative case), and trust features in Muse and Instinct hint at trust networks. “If you squint, you’re seeing the early signs” — a potentially fertile area for entrepreneurs.

8. The creative and hardware renaissance, three weeks at a time

• Games as leading-edge media, per Dixon: in roughly two weeks since “Opus 55 or something,” disassemblers and AI can run Windows games on Unix faster than on Windows, and one tool mashes any two games into a combo title (“Super Mario and GTA,” Ben suggests). Epic had just said “no more humans” writing code — “the first of the dominoes to fall” against cultural resistance in the industry, in Dixon’s view, and a possible creative renaissance where independents can build as if they have a huge team.
• Ben’s Digger example: viewers may be tempted to walk out after the first half hour, but he says the film is not presented the way it seems. He calls it an “incredible creative achievement,” with a $180 million production budget and a $200 million cash loss. “On a $2 million budget, that movie would have been insanely profitable” and made investors 10 times their money back. Ben predicts AI film will be an entirely new medium, not just a cheaper way to make regular movies. Dixon’s camera analogy: the negative frame is that it replaces landscape painters; the positive is that it unlocks motion pictures.
• Hardware, Marc’s “terra incognita,” may be more accessible: Ben describes an open-source tool called REA — Reverse Engineer Anything — that can use an AI model to reverse-engineer an object and provide source code. Point it at a misbehaving thermostat, and the AI can write new code, install it, and provide a new interface; Ben says the same may apply to many forms of hardware. Home automation “maybe should just be an AI bot” that takes over the hardware in your house. Marc also wonders whether AI could enable a homebrew-hardware revival, with Claude designing 3D-printed cases, cheap motherboards, and custom firmware — a possible MakerBot-like movement. Erik gives the architectural example: a friend’s new building is being steered by AI, including design-for-manufacturing specs that are normally difficult to produce.

9. Open source as the guarantor against centralization

• The panel considers Erik’s citation of Thiel’s “crypto is libertarian, AI is communist” framing. Dixon sees the possibility of a more decentralized market structure: big labs alongside open source, which he believes is growing much faster in token consumption according to OpenRouter data. Local models may emerge; in the near term, Dixon thinks people may be able to run a top-end coding agent or chatbot on their computers. Stablecoins are a natural way for agents to pay each other, with decentralized lending and other crypto services potentially built on top.
• Marc on the policy state of play: he says the threat of “draconian and completely unacceptable” regulation or bans remains; the EU “still aspires to do it,” though Marc says it lacks the leverage. He describes the current US administration as “100% pro-open source.” He was surprised that the Chinese Communist Party is enthusiastic about open-source AI for the world. Marc says open models are lapping closed models by six or 12 months, depending on how you count.
• His centralization warning: without open source, AI “inevitably centralizes down to one or two companies, or maybe just flat out the government… picture the concept of government sushi. Then keep your eyes closed and picture government AI.” He also makes the education case: “what if the only way your kid can learn about AI is to go to work for one of two companies?” Marc says open-source models will teach people how to make open-source AI, while models from certain large companies may refuse. His example: wilderness explorers carrying local LLMs for offline advice about survival, including which berries are poisonous or what to do after a rattlesnake bite.
• Ben’s safety argument for openness: “probably the biggest safety problem” so far is reward hacking — the AI doing something it wasn’t told to do as a means to get its reward. Ben says the labs haven’t worked on it very hard, for whatever reason, and asks what might happen if every PhD student, professor, and undergraduate could inspect the weights and study the problem. “You really want the whole world making this technology great and safe, not just the people who happen to work at two or three companies. Go Meta… go SpaceX.”

10. What an 18-year-old should do — and humble definite optimism

• Dixon, borrowing Marc’s line: study “modern alchemy” — “humans just figured out how to turn sand into intelligence” — via math, computer science, and maybe economics, and get into the mix at the cutting edge of what he expects to be the dominant narrative of the next 20 or 30 years. Marc lists fields being transformed: classics via the Vesuvius Scroll Project, economics, biology, design, and zoology — “we’re plausibly heading toward a world in which we’re going to understand what the animals are saying.” Dixon says, seriously, “I think there’s a Nobel Prize for that.”
• Marc’s generational frame from Douglas Adams: under 15, it’s just how the world works (his seven-year-old on ChatGPT: “Duh. What good is a computer that doesn’t answer questions?”); ages 15–35, it’s new and exciting, and might become a career; over 35, it’s “unholy and evil.” Schools banning AI may make it kids’ “secret superpower.” Ben’s kicker: “you’re the one with the problem because you’re obsolete. The kids are going to be superhumans.”
• Erik’s closing calibration: “humble, definite optimism” — he has been wrong many times, so the firm wants to find great people with their own definite optimism and back them. Chris also stresses that they are ultimately betting on great entrepreneurs, including people with ideas they did not discuss.