a16z Frontier Targets Founders Before New Technology Categories Exist

Chris Dixon says a16z is expanding its crypto fund into a16z Frontier because crypto has become part of the broader technology landscape. Stablecoins are now in the toolkit of companies such as Stripe, he says, while Wall Street firms are moving into tokenized stocks and other digital assets. As crypto overlaps with fintech and financial services, it is harder to say where the category begins and ends.

The second reason is organizational. a16z has investment groups focused on established areas, but new opportunities still fall outside them. Ben Horowitz says the firm had relied on Dixon to explore novel fields, then developed a dedicated practice when an area grew large enough. The step between spotting an unusual opportunity and building a repeatable investment effort around it was less defined. Frontier gives the firm a way to pursue that work directly: back founders in areas beyond the existing groups and see what develops.

Dixon calls these fields “outer planets.” The term describes opportunities that may be technologically important but have not yet become recognizable investment categories. Horowitz says an area identified through Frontier might eventually become its own fund, or Frontier itself might grow into a larger practice. The firm does not need to decide which outcome in advance. It needs a way to reach founders before a new vertical exists.

Dixon describes his earlier investments as bets on second-order effects: opportunities created when components of a new platform improve and entrepreneurs recombine them. Smartphones generated obvious first-order opportunities in apps. But their sensors, chips, and other components also formed a supply chain that founders could use to build products beyond the phone. Dixon recalls Marc Andreessen first getting interested in drones and bringing him into the area. He describes Oculus as a “reconfigured smartphone” and says Skydio initially drew on the same supply chain, though its technology has since advanced.

The distinction matters to Frontier’s investment logic. A first-order bet applies a new technology directly, as an app applied the smartphone to a familiar task. A second-order bet asks what becomes possible because the technology’s parts, infrastructure, or communities now exist. Dixon’s examples suggest that founders may create products in areas not usually grouped together, using capabilities assembled for another purpose.

Dixon sees crypto through a related lens. He describes it partly as an outgrowth of open-source software and social media: online communities could form outside mainstream acceptance and coordinate around blockchains. For him, these examples point to a broader investment habit. A technology can matter as a direct product category, but it can also supply components, communities, or forms of coordination that support businesses in fields that were not obvious at the outset.

The firm’s crypto experience, Dixon says, offers expertise that can travel. Andreessen emphasizes that crypto investing requires deep technical knowledge alongside business and market judgment. He points to the team’s experience in cryptography, mathematics, game theory, and software. Dixon adds that the firm has built policy and regulatory expertise through its work in crypto, which he expects to be relevant as technologies such as self-driving cars, drones, and AI become political issues. The expertise is not only about understanding a technology; it also includes helping entrepreneurs navigate the public and regulatory questions that can accompany emerging fields.

Horowitz frames the mandate around founders rather than sectors. He says Dixon’s earlier investment thinking helped the firm look for ideas that initially seem like bad ideas but contain a founder’s insight that makes them viable. Unfamiliarity alone is not a reason to invest; the question is whether the entrepreneur sees why an idea could work when others do not.

That is Frontier’s investment thesis. The speakers’ larger predictions about AI, hardware, media, and the economy explain why they think the opportunity set is widening, but they do not replace the firm’s stated focus: finding and backing entrepreneurs whose work does not yet fit an established category.

Chris Dixon says many good technology ideas that eventually worked had earlier versions that failed, sometimes because they arrived a decade too soon. Timing depends on technology, culture, business conditions, and politics lining up; he calls it the hardest part to predict. Frontier intends to err toward being early and accept that some categories will not work.

Virtual reality illustrates the cost of that position. a16z invested in Oculus, which was acquired, but Dixon says VR had not become as mainstream as hoped more than a decade later. He remains interested in the field, while noting that more technical problems had to be solved than expected. Being early can preserve a stake in a large opportunity; it can also mean waiting while the market develops.

Dixon thinks AI may change how investors assess that wait. He points to progress in software engineering and mathematics and argues that the distinction between an engineering problem and a “science project” may need to be reconsidered. He expects further changes in physics and hardware. A team of roughly 30 people, he says, may now be able to attempt a full-stack device company that might have required hundreds or even a thousand people a decade earlier.

That does not remove the risks. Founders still need to think about defensibility and avoid being so close behind the AI frontier that their work is overtaken by the next wave of capability. But Dixon argues that entrepreneurs have only begun exploring what the tools make possible. Projects that once required much larger teams may become viable sooner.

Marc Andreessen sees a particularly wide gap between what AI-native founders believe they can build and what typical incumbent companies understand about the technology. He describes it as the largest such gap he has seen in his career, and says it is widening as capabilities advance. Startups building around AI, in his view, have a different sense of what a team can do than incumbents in almost any industry.

Andreessen imagines what he and Horowitz could have attempted at Netscape in 1994 if they had access to the tools they discuss. Horowitz replies that they would have “built the whole damn internet.” They speculate that streaming video, crypto, and hardware projects might have come sooner. The example captures their view of AI’s leverage: small teams may be able to take on more technically ambitious work.

That leverage changes the standard the firm applies to ambition. Horowitz says a16z used to ask whether an idea was too ambitious; now it is more likely to ask whether it is ambitious enough. A moderate plan may not be distinctive given what the tools could enable. Andreessen expects founders to combine areas previously treated as separate—software and hardware, for example—and build products and businesses with broader scopes.

Andreessen does not say incumbents cannot adapt. Some are moving directly into AI, he says, and a16z works with companies that want to equip established businesses with new tools. But he warns that the clock is ticking. Adoption can be easier for individual employees than for institutions: employees may use AI to check work or proofread emails without telling management. That gap matters because an employee’s private use of a tool does not automatically change a company’s products, processes, or ability to compete.

He connects that gap to company formation. If individuals can do more with these tools, some may find it easier to start a business. Andreessen says startup formation is already rising and expects it to increase further; Horowitz agrees they are already seeing it. Their prediction is not only that existing companies will become more productive, but that the tools may change who can start a company and how much it can attempt.

Dixon argues that the familiar frame of technology companies disrupting one another may be too narrow. He suggests the larger contest may be between technologically modern businesses and the rest of the economy. He points to Atmo, which he describes as building a full-stack business in food service and related areas, and mentions mining as another field. Frontier’s rationale is to stay open to companies using technology to reorganize activities outside established technology categories, while accepting that the markets and timing may be unclear.

Chris Dixon distinguishes between using new technology to do familiar things better and creating activities that were not possible before. Early internet commerce often moved existing behavior online: a customer could order from a catalog more easily. Later companies made bets on new behavior. Dixon uses YouTube as an example of a company that depended not only on the technology but on a future in which new kinds of creators and audiences would emerge.

The distinction is about more than product design. An established behavior gives a company a known task to improve, but the new technology may initially be constrained by the assumptions built into that task. A catalog put online remains a catalog; a search box that answers questions remains a familiar interface, even if the answer is generated differently. New behavior requires people to adopt activities that are not yet routine. Dixon says those opportunities involve more uncertainty because founders are betting on changes in technology and in culture or behavior at once.

He sees AI use so far as concentrated in familiar tasks: coding, mathematics, knowledge work, and search-like interaction. The more consequential applications, he argues, may involve habits and forms of participation that are difficult to anticipate from today’s products. Some opportunities will be direct applications of known capabilities; others will require technical and cultural changes to arrive together. Dixon calls those “bank shots”: harder to predict, but potentially larger if the pieces come together.

One possibility is a shift from AI as a one-person tool to AI as a network. Most current interactions, Dixon observes, still involve a person typing to a model and receiving a response. He expects a possible “multiplayer” phase in which agents coordinate, share knowledge, and transact. Products such as Muse and Instinct, which he describes as utility or proxy agents for people, suggest one early direction. He also points to experiments with groups of agents, including a swarm that he says helped solve a mathematical problem. Similar coordination, he suggests, could help with work such as designing a building or making a scientific breakthrough.

Dixon predicts that agents could eventually become more numerous participants on the internet than people. If they take on more activity, he says, they may need ways to communicate, exchange information, establish trust, and make payments. This would be more than giving each person a more capable assistant: agents acting on behalf of people could interact with other agents, creating demand for shared ways to coordinate. Dixon describes the area as early, with products still in an experimental phase.

In that possible economy, Dixon connects agent networks to crypto. Stablecoins, he says, could be a natural way for agents to pay one another; decentralized lending and other financial services could be built on top of programmable money. This depends on agents becoming active participants in transactions rather than remaining private tools used by individuals. He presents it as an opportunity to explore, not a settled forecast.

Andreessen’s concern is that early AI products may reflect the limited imaginations of the people building them. He calls this “product manager brain”: designing for familiar, narrow use cases while claiming to build for everyone. His examples include arranging a restaurant booking for a group of friends who are also product managers, or planning a trip to a product-management conference. He says there is a role for such products, but they do not represent the full range of what people might want.

He also warns that an assistant defined by the most restrictive version of helpfulness can become “the most boring possible thing.” Andreessen expects AI products to extend into entertainment, creation, and ways for people to earn or save money. He points to recent activity he describes as an explosion of people making music videos from scratch. His point is not that this is already a settled industry, but that creative uses can emerge quickly and outside the narrow tasks that early products are designed to address.

Video games offer another view of that change. Dixon says games often sit near the leading edge of media because their technical culture moves quickly. He describes tools that, in his account, can disassemble and port games, as well as an experiment that combines two games into a mash-up. Those are examples he uses to illustrate how AI-assisted tools might let creators recombine material in ways that would have been difficult before. He expects AI to help independent developers make games that would previously have required much larger teams. In his view, that could expand human creativity by giving creators more leverage.

Film may change in two ways: AI could make conventional movies more efficiently or improve production elements, while also enabling a distinct medium with stories that have not been viable before. Horowitz argues that AI film will not simply be a movie made differently; it could be something new. Dixon compares the possibility to the camera, which changed an older form of media and unlocked motion pictures.

Andreessen illustrates the economic stakes with Digger. He describes it as a creative achievement that some viewers might leave early, and says its production budget was about $180 million and that it lost roughly $200 million for the studio. He asks what the economics might have looked like if a film of that kind had cost $20 million—or $2 million—to produce. In his hypothetical, a $2 million version could have been highly profitable and returned many times its cost to investors. Production costs shape which creative projects are attempted; lowering them could change both the business case and the range of work that gets made.

Chris Dixon treats AI’s eventual market structure as unsettled. He expects large proprietary labs to succeed, while pointing to growth in open-source use and the possibility of models running locally on personal devices. Referring to open-router data, he says open-source token consumption is growing quickly. He expects local models to become more capable as well.

Dixon wants a landscape in which large companies coexist with open-weight models and systems that run on local devices. Those alternatives, he argues, give consumers and startups more choice and may support a more decentralized market. The possibility matters to the agent economy too: if agents become active online, their developers may benefit from having options beyond the largest providers.

Marc Andreessen makes the case for open source through access, education, and resilience. He says open-source AI is advancing in both the United States and China alongside proprietary systems. He also says governments have threatened restrictions on open-source development, and argues that limits could leave AI concentrated in a few organizations.

The educational case is that people should be able to examine and learn how the technology works. Andreessen asks whether the only way to understand AI should be to work for one of a handful of companies. Open systems let students and developers experiment directly, he says, and AI itself can help teach people about AI. He contrasts open-source models, which he says will teach users how to make open-source AI, with some proprietary models that refuse to do so.

Andreessen also describes local models as useful when there is no internet connection. He imagines wilderness travelers carrying models on their phones so they can seek help in an emergency or identify a poisonous berry. A locally running model, unlike a cloud service, can be available without connectivity. The example makes the broader point concrete: access to a tool can depend not only on its capabilities but on whether a user can reach the service that provides it.

Dixon and Andreessen make a defensive case too. Dixon argues that open-source software can help people protect themselves in cybersecurity. Andreessen says open tools give users capabilities that would otherwise be available only through proprietary cloud services. He warns that without open source, AI could consolidate among one or two companies or the government.

Ben Horowitz makes the safety argument through reward hacking: an AI system may find a way to obtain a reward that does not match what its designers intended. Horowitz says the leading labs have not worked on the problem as much as they might. He asks what researchers could discover if students and professors could inspect model weights and investigate the behavior. His argument is that a wider research community should be able to work on making AI safer, rather than leaving the task to people inside a small number of companies.

That claim links openness to both opportunity and risk. Andreessen and Horowitz argue that broader access can help more people learn the technology and investigate problems; they also acknowledge the concern that open systems can be used for harmful purposes. Their case is that restricting access carries its own costs: it can concentrate capability, limit public understanding, and narrow the pool of people working on safety.

Chris Dixon describes the internet’s “covenant” as an exchange between distribution platforms and the sites they indexed. A search engine could show a snippet, send a user to a website, and help that publisher earn money through advertising or transactions. Social networks also sent people outward. The arrangement depended on users visiting the sites that supplied information or services.

Dixon says that bargain has weakened as search engines answer questions directly and social networks try to keep users on their platforms. When people get answers without visiting source websites, those sites may lose the traffic on which their advertising or sales depend. He points to Stack Overflow and other sites as early examples of traffic falling as AI systems offer alternatives. He presents those sites as signs of pressure on the existing arrangement, not as proof that every online business will be affected in the same way.

The underlying needs will remain: people will still book travel and look for recipes, Dixon says. But they may use an agent or another interface rather than visit the same websites in the same way. Businesses built around human attention and referral traffic may face painful transitions, even when the services they provide remain useful. If an agent supplies the answer or completes the task, the user may no longer see the publisher, comparison site, or service that used to receive the visit.

Dixon sees an opportunity alongside the disruption. Entrepreneurs can ask what a better internet would look like when people have AI tools and agents acting on their behalf. Over the coming decade, he expects a reconfiguration of how attention flows, knowledge is shared, people produce things together, and organizations interact online. He says this could be more dramatic than the transition to mobile, which expanded access and engagement but did not, in his view, have the same effect on the internet’s economics.

Asked what he would study if he were 18 today, Chris Dixon answers with a phrase he credits to Andreessen: “modern alchemy.” Humans, Dixon says, have learned to turn sand into intelligence. He suggests mathematics and computer science as ways to understand that transformation, perhaps alongside economics. A young person need not become a specialist in one narrow technical field, he says, but should be involved with AI, use the tools, and stay close to a technology he expects to shape the next 20 or 30 years.

Andreessen’s examples extend beyond computing. He says AI may help decipher ancient scrolls and change what can be learned from lost texts. He expects economics to be transformed and points to biology, drug discovery, design, film, music, and other forms of art. Horowitz raises the possibility of understanding animal communication and changing how people train dogs. The examples suggest that opportunities may emerge inside established disciplines as well as in a new AI industry.

Andreessen invokes a generational framework he attributes to Douglas Adams. In his telling, children under 15 may experience new technology as simply the way the world works; people between 15 and 35 may see it as an exciting field and possible career; people over 35 may regard it as a threat to existing structures. He recalls introducing ChatGPT to his seven-year-old, who wondered why a computer would not answer questions. For the child, that capability was an obvious expectation.

Andreessen argues that school restrictions may make AI more appealing to students, who may treat it as a “secret superpower.” Horowitz challenges the assumption that young people are the ones struggling most: adults, he says, may be the ones with the problem. He expects AI to let younger people enter fields that previously required more experience, money, or organizational support.

Dixon qualifies the optimism. The firm is not investing in abstract theories about what AI will do, he says, but in entrepreneurs. He calls the posture “humble definite optimism”: conviction that important things can be built, alongside recognition that the firm has been wrong many times. Frontier aims to find founders with their own insight and confidence, and to back them before their ideas fit established categories.