Andreessen Horowitz has formally expanded its crypto-focused fund into a broader frontier technology vehicle called a16z Frontier, with Chris Dixon leading the effort. The announcement came in a special a16z podcast episode featuring Dixon alongside firm co-founders Marc Andreessen and Ben Horowitz, moderated by Erik Torenberg.
The shift moves the fund beyond blockchain-only investing and into a wider set of areas the firm described as the edge of technological possibility, including hardware, embodied robotics, entertainment and gaming, space biology, and decentralized open-source intelligence.
a16z Frontier broadens the old crypto mandate
Dixon said the team has spent the past several years rethinking how frontier technology cycles are evolving. When a16z launched a dedicated crypto fund in 2018, blockchain represented one of the clearest frontier themes. He said that is no longer the whole picture. Generative AI, modern cryptography and advances in compute have started to spill into one another, and many of the strongest founders now work across several of those domains at once.
According to Dixon, the firm is seeing founders who are building decentralized networks while also working on embodied robotics, local AI agents, new compute clusters, synthetic biology and new immersive media formats. He said a16z Frontier is meant to support that kind of company without forcing it into a narrow category.
Andreessen framed the move as part of a longer pattern inside the firm. Over 15 years, he said, a16z has repeatedly changed its structure and investment strategy as technology waves shifted, from cloud and mobile to enterprise software, health and crypto networks. In his view, technology does not advance in isolated verticals. It moves as an interconnected system, and AI is now pulling together decades of work in hardware, physics and software.
Horowitz made a similar point from the startup side. He said the most important breakthroughs often happen in the gray zones between disciplines. A company using open-source foundation models to power humanoid robots while handling data ownership and micropayments onchain cannot be cleanly labeled as only AI or only crypto. He said a16z Frontier is designed to back founders working in those overlaps.
Why “too early” may matter less than before
Torenberg asked whether the old venture capital warning that being too early is almost the same as being wrong still applies. Andreessen said the warning came from a real constraint in earlier eras: startups often had to wait for the surrounding infrastructure to catch up.
He pointed to online video two decades ago. A company could have the right idea, but if households were still on dial-up internet, without fiber broadband, smartphones or cloud-native data centers, the business might run out of money before the environment was ready.
That timing problem has changed, he argued, because the underlying technology stack has compressed. Startups no longer need to lay cable or build chip fabs from scratch. Public cloud infrastructure is already available, open-source base models are widely accessible, and dense mobile computing is everywhere. AI itself is also acting as an accelerator, shrinking tasks such as coding, prototyping, molecular simulation and hardware rendering from years to days or even hours.
For that reason, Andreessen said the bigger risk now is not simply being too early. It is being too timid. A startup that is only six months ahead, he said, can be overtaken by large technology companies or open-source communities before it fully launches.
Dixon agreed and said many people once dismissed smart contracts and decentralized compute as premature. In his telling, those early efforts laid the groundwork for today’s censorship-resistant, high-throughput distributed settlement networks. He said Frontier is willing to take “too early” risk because outsized outcomes usually appear while mainstream opinion is still confused or skeptical.
A widening gap between builders and the public
Dixon described the current moment as the widest gap he has seen between early adopters and public understanding in the history of computing. He said mainstream society, many multinational executives and even many regulators still treat AI as a chat interface that can draft documents or edit resumes. In that view, AI is little more than a productivity add-on.
Inside top developer circles, he said, the reality already looks very different. Builders are using large numbers of autonomous agents to collaborate on complex operating systems. Robots are training in simulation at speeds far beyond human learning cycles. Synthetic biology researchers are using deep neural networks to design highly active proteins that do not exist in nature, and doing so in a matter of weeks.
Andreessen compared that disconnect to the early 1990s, when many people saw the internet as little more than an electronic newspaper and failed to imagine its eventual effect on finance, supply chains and geopolitics. He argued that this kind of perception gap can become a startup advantage, because incumbents often judge new systems through the logic of the old ones.
Who captures AI profits: incumbents or startups?
Asked whether AI economics will mostly favor existing technology giants or produce a new generation of major startups, Horowitz said the incumbents clearly hold early advantages. He named Microsoft, Google, Amazon and Nvidia as companies with distribution, compute budgets and enterprise relationships already in place. Plugging AI models into existing products such as office software or search is the fastest route to revenue.
Still, he said technology history repeatedly shows the innovator’s dilemma. Large companies are built around old channels, old organizational structures and old profit engines. When a new paradigm requires them to disrupt those profit engines, hesitation follows. He used the early internet as an example: telecom companies controlled the fiber and phone networks, yet companies such as Netscape, Amazon, Yahoo and Google emerged as the defining internet winners.
Dixon said the same logic applies even more strongly in a future agent economy. If the internet becomes a network where autonomous agents allocate resources and settle transactions with one another, then centralized business models based on card payments and ad clicks stop fitting the system. Agents do not need banner ads, he said. They need low-friction micropayments and cryptographic credit onchain.
Why an agent internet points back to crypto rails
Dixon said that within the next three to five years, humans could become a minority among active participants in online interactions, with tens of millions of AI agents acting on behalf of people in asset allocation, data retrieval, code deployment and supply-chain bidding.
That creates a basic identity and payments problem. How does an agent prove who it is, and how does it settle across borders without a traditional bank account? Dixon said centralized banking systems are not built to issue debit cards to autonomous code. In his view, that is why decentralized protocols and AI agents fit together so naturally. Agents need their own cryptographic keys, their own onchain wallets and smart contracts that let them coordinate and settle in permissionless networks.
Andreessen extended the idea into social networks. He said the dominant social platforms of the last 20 years were built around person-to-person following graphs. Future social systems could be mixed environments where agents have their own personalities, knowledge bases and economic goals, debate ideas, create content and even form decentralized collaborative organizations.
Andreessen’s critique of “Product Manager Brain”
One of Andreessen’s sharpest comments in the discussion was aimed at what he called “Product Manager Brain.” He described it as a bureaucratic mindset that grew out of the mature mobile internet era, where innovation gets reduced to retention curves, multivariable A/B testing and tiny interface changes meant to squeeze out marginal gains in click-through rates.
He said that approach may help extract some profit in mature software markets, but it does not produce frontier breakthroughs. In fact, he argued, it can suppress them. If Steve Jobs had relied on focus groups and A/B tests while building the iPhone, Andreessen said, the result would likely have been a BlackBerry with a smaller keyboard rather than a new computing form factor.
For him, major platform shifts come from taste, artistic instinct and engineering obsession, especially in AAA games, film storytelling and new hardware devices. He said AI should not be used mainly to generate more dashboards for product managers. It should put powerful creative tools in the hands of small teams with unusually strong taste. In that scenario, a 10-person studio could build an interactive universe with visual detail and world-building depth beyond top Hollywood productions.
Horowitz agreed, saying business is not just an algorithmic loop. It still depends on reaching human curiosity and emotion, and founders need to keep making conviction-driven decisions instead of handing all judgment to dashboards.
Open-source AI as a check on concentration
Torenberg also raised Peter Thiel’s old line that “Crypto is libertarian, AI is communist,” a shorthand for the idea that blockchains distribute power while AI naturally centralizes it. Andreessen and Dixon both rejected the idea as a fixed outcome.
Andreessen said the argument had force in an earlier phase of deep learning, when progress depended heavily on centralized data centers and the servers of a small number of large companies. If only a handful of corporations and governments controlled the most capable machine intelligence, he said, AI could become an extreme instrument of centralization.
But he argued that open-source AI has changed the trajectory. He pointed to Meta’s open-source Llama family and to global communities working on architecture improvements and quantization. In his account, advanced intelligence is being compressed quickly enough that models with capabilities comparable to supercomputing systems from a few years ago can now run on laptops and even smartphone chips. Open source, he said, means the code is visible, the weights are not locked inside a single gatekeeper, and anyone can fork, fine-tune and deploy.
Dixon said that logic closely matches the spirit of cryptography. Cryptography removes privileged financial intermediaries; open-source models can remove privileged intermediaries around intelligence itself. He also said some entrenched interests in Washington and other regulatory centers are lobbying in the name of safety to classify open-source software as illegal, using compliance barriers to protect monopoly positions. He said a16z Frontier intends to keep backing open ecosystems, decentralized compute and open protocols.
The internet’s creator bargain is breaking down
Dixon returned to a theme from his book Read Write Own: the collapse of the internet’s original economic bargain. He said early Web2 platforms were built on an implicit deal. Independent creators, bloggers, news organizations and community contributors supplied large amounts of high-quality content for free, and platforms returned value through distribution, audience growth and monetization opportunities.
Once network effects hardened into monopoly power, he said, platforms began tightening algorithms and forcing creators to pay for reach. In the generative AI era, he argued, model companies are scraping years of human work across the web and then delivering direct answers inside search and chat interfaces, without sending users back to the original sites.
That breaks the old economic loop for creators. If the pattern continues, Dixon said, high-quality original work will dry up and the internet will fill with low-quality machine-generated output. His proposed fix is blockchain-based, verifiable data ownership, where each use of data for inference or training can be traced and compensated automatically. Rebuilding that contract, he said, is one of Frontier’s most important missions.
What could come after the smartphone
On hardware, Horowitz said smartphones have dominated digital life since the first iPhone launched in 2007, but he does not see the current form factor as the endpoint. The problem, in his view, is that the smartphone pulls people out of the physical world and forces them to stare down at a small glass screen.
He said the firm is especially interested in always-on wearable computing, particularly smart glasses and environment-aware hardware. A device with the weight of ordinary glasses, equipped with a forward-facing high-dynamic-range camera and directional audio microphones, could give users a companion AI that shares their first-person perspective and understands the surrounding physical environment.
That system would know what the user is looking at, understand who the user is speaking with and provide assistance at the right moment, without requiring the user to unlock a phone and navigate an app. Andreessen described this as a shift from tool-based computing to environment-aware intelligence. He said spatial audio, retinal projection and haptic feedback are all maturing, and that the category could support a hardware supply chain worth hundreds of billions of dollars.
What they would study if they were 18 today
At the end of the conversation, Torenberg asked what each of them would focus on if they were 18 years old and just entering college or the workforce.
Andreessen said he would avoid standardized fields that train people for routine work likely to be automated by large models. Instead, he urged young people to go back to first principles and the physical world: mathematics, theoretical physics, low-level computer systems, chemistry and materials. At the same time, he said they should become highly fluent with the latest AI tools.
He argued that today’s young builders have an unprecedented productivity multiplier. A person who knows how to use modern generative tools and coding agents can, in some cases, match the output of what used to require a 20-engineer software team. Building real products and learning from code and user feedback matters more, he said, than collecting credentials.
Horowitz added that leadership and responsibility remain scarce no matter how technology changes. He said young people should learn to face failure, gather talented misfits around a mission and hold onto their judgment when the broader public dismisses them.
Dixon’s advice was to keep a geek mindset and stay skeptical of overpackaged consensus. He said people should join new frontiers while they still look rough and unfinished. He placed open-source software 20 years ago, blockchains 10 years ago, and today’s hard-tech and agent world represented by a16z Frontier on the same longer arc of frontier building.
Key lines from the discussion
- Dixon said the key question for founders is no longer whether their ambition is too large, but whether it is large enough.
- The three investors said the gap between what early users understand and what the public and traditional management believe is unusually wide.
- Andreessen used “Product Manager Brain” to criticize optimization-heavy thinking centered on A/B tests and tiny conversion gains.
- Andreessen and Dixon both argued that open-source AI is a critical defense against monopoly power and concentration.
- Dixon said the original creator-platform bargain of the internet has been broken by AI scraping and closed-loop summaries, and that cryptography and onchain ownership are needed to rebuild it.
The podcast episode was released on Oct. 22, 2026 under the title “The Frontier Ideas a16z Wants to Fund Now | Marc Andreessen, Ben Horowitz & Chris Dixon.”