Chris Dixon on Blockchains, AI, and the Future of the Internet

Chris Dixon — general partner at Andreessen Horowitz and author of Read Write Own — joins Tyler Cowen in Ep. 240 to argue that blockchains are the one architecture capable of reversing the internet’s consolidation into a handful of corporate platforms, and to map how AI could either accelerate that consolidation or — through open source — check it.

Key ideas

1. The internet consolidated because subsidised corporate networks outcompeted open protocols. The early web was genuinely decentralised — money flowed to the edges, creators kept most of what they earned. Corporate networks (YouTube, Spotify, Facebook) won by subsidising their own hosting costs with venture capital, then locking in network effects. Today five to ten companies capture more than 90 per cent of internet traffic and extract what Dixon calls high take rates — the percentage of money flowing through a system that the intermediary keeps.
2. Blockchains aim to combine protocol-network benefits with corporate-network capabilities. A protocol network (think email or RSS) has no intermediary and no take rate, but cannot subsidise users or coordinate centrally. A blockchain adds smart contracts — self-executing code that holds and deploys funds — so the network can offer the same incentives (subsidies, rewards) a corporation can, while keeping low take rates and community governance. Dixon calls this the thesis-antithesis-synthesis of internet history: read, write, own.
3. Stablecoins are blockchains’ first mass-market product. Dollar-denominated stablecoins — digital tokens whose value is pegged one-to-one to a reserve asset — processed $3.5 trillion in transactions in a single month (Visa’s own dashboard figure). Their killer application is not speculation but payments: invoicing across borders at one second and one penny per transaction, with full digital automation that eliminates invoice fraud. Dixon expects every major bank to issue one once Congress passes stablecoin legislation.
4. AI and blockchain are converging around agent-to-agent payments and open governance. Because AI agents need to transact autonomously but cannot open bank accounts, crypto infrastructure becomes their natural payment layer. More broadly, Dixon sees AI as a centralising force — whoever builds the foundation models accrues immense soft power — and open-source AI as the structural counterweight, the only route by which countries like Peru retain meaningful sovereignty over their own information systems.
5. NFTs and digital ownership are the correct frame for the AI era. Non-fungible tokens — unique digital assets on a blockchain, as opposed to interchangeable fungible ones like currency — are not merely speculative JPEGs. They are the right way to represent ownership of any distinct digital object: game items, housing deeds, creative works, AI-generated media. Dixon’s broader claim is that the internet’s missing primitive is digital ownership, and blockchains supply it.

Content

Protocol networks, corporate networks, and what went wrong

Dixon opens with a historical arc. The early internet ran on protocol networks: open standards — HTTP, email, RSS — with no company behind them, no intermediary taking a cut. A musician with a thousand paying fans at ten dollars a month could earn a living; Kevin Kelly’s famous ‘1,000 True Fans’ essay described this possibility precisely. That world dissolved not through regulation but through economics. Corporate networks subsidised their way to dominance: YouTube, for example, absorbed enormous video-hosting costs using venture capital, offering free hosting to creators who would otherwise have paid. Once a service reached scale, network effects — the dynamic by which a platform becomes more valuable the more people use it — made displacement nearly impossible. The result: 90 per cent of internet traffic now flows through fewer than ten services.

Dixon draws the analogy to open-source software. Linux won 90-plus per cent of operating-system deployments not by out-subsidising Windows but through composability — the property by which any piece of code, once written and published, can be reused by anyone else as a building block. Composability, he says, is to software as compounding interest is to finance: a ratchet with no ceiling. Blockchains inherit composability from open source; the question is whether they can also match the subsidisation and coordination capacity of a corporate network.

Blockchains as a third architecture

Dixon distinguishes three types of internet architecture. Protocol networks (the early web) have good societal properties — low take rates, open access, community governance — but limited competitive tools. Corporate networks (Google, Facebook, Amazon) have powerful competitive tools — capital for subsidies, hierarchical execution, data flywheels — but extract high take rates and centralise control. Blockchain networks, Dixon’s thesis, can in principle combine both: a smart contract (a piece of code that owns and deploys tokens autonomously, subject to no single controller) can subsidise users, reward contributors, and coordinate the network the way a treasury department does for a corporation, without any single party being able to extract rents or de-platform participants.

He is candid about the difficulty. Mastodon, the decentralised social network, failed in his view not because decentralisation is wrong but because its server-federation architecture created coordination costs that no subsidy could bridge. He treats this as an architectural flaw, not a verdict on the category.

Stablecoins: the first mainstream application

Dixon’s most confident current claim is about stablecoins. A stablecoin is a digital token pegged to a fiat currency — most commonly the dollar — and backed by a reserve asset (typically US Treasury bills). Unlike earlier algorithmic stablecoins such as Terra Luna, which collapsed in a classic bank run because they were backed only by their own circular token scheme, asset-backed stablecoins like USDC hold one dollar in reserve for every token issued.

The commercial use case that Dixon finds most compelling is not savings or speculation but cross-border payment rails. Stripe acquired Bridge, a stablecoin infrastructure company, and describes its use case as treasury management and international invoicing — moving money between jurisdictions in seconds, with full end-to-end digital automation that makes invoice fraud structurally impossible. The technical benchmark Dixon cites: transactions on Base (Ethereum’s Layer 2 network) and Solana now achieve one second and one penny — a target the industry had aimed at for years.

Cowen presses on regulatory fragility: what stops unregulated offshore stablecoins from capturing users seeking yield without the 100-per-cent reserve constraint? Dixon’s answer is a combination of reciprocity requirements in proposed US legislation and network effects among legitimate actors — if Visa, Mastercard, Fidelity, and every major bank issue stablecoins through regulated channels, the offshore fringe is marginalised rather than competitive.

AI and the internet’s implicit covenant

Dixon introduces the idea of an implicit covenant between the internet’s distribution layer (Google, Facebook) and its content layer (everyone else). For 25 years the deal was: let us index and excerpt your content; in return, we send you traffic. That equilibrium was already tilted toward the distributors — Rupert Murdoch sued Google; Facebook cut news in Canada rather than pay — but it was functional.

Generative AI breaks it. If users get the answer from Claude or ChatGPT rather than clicking through to Stack Overflow or a cooking site, the traffic-for-content exchange collapses. Dixon points to Stack Overflow — a site he sat on the board of — whose traffic has fallen roughly 80 per cent since AI coding assistants like Copilot and Cursor absorbed its knowledge. He calls Stack Overflow the canary in the coal mine and worries that the open web’s long tail of blogs, specialist sites, and independent creators will atrophy to nothing, leaving an internet that structurally resembles 1970s broadcast television: four channels, no serendipity.

His tentative diagnosis for media: barbelling. Technology repeatedly hollows out the middle of a market while extremes grow. Retail saw Amazon and LVMH win while Sears and Kmart died. Media has seen TikTok dopamine hits and three-hour podcasts flourish while the 30-minute sitcom shrank. AI will likely accelerate this: mass-scale AI-generated content at one pole, premium handcrafted human work at the other — live concerts, limited-edition books, bespoke journalism. He draws the analogy to photography, which both replaced representational painting and enabled an entirely new medium, film. AI may eliminate illustration as a profession while spawning a native AI creative form that could not have existed before.

AI, geopolitics, and open source as structural counterweight

Cowen puts the geopolitical challenge starkly: if Peru hands its education system, treasury management, and national defence to American AI companies, has it in any meaningful sense retained sovereignty? Dixon’s answer hinges on open weights — the publicly released parameters of a model like DeepSeek or Llama that allow any government or institution to download, fine-tune, and modify the model locally. With open weights, Peru can at least adjust the model; without them, it is entirely dependent on whoever trained it.

Dixon expects open-source AI to become one of the two dominant competitive paradigms — alongside proprietary models — precisely because governments and large institutions will demand it. DeepSeek’s competitive parity with frontier models, achieved by a team of roughly 150 people, signals that the tricks of model training propagate quickly. Foundation models may end up like PyTorch or statistical libraries: commoditised infrastructure that many actors can run, with value accruing to applications and integrations rather than to model weights themselves.

On AI and politics more broadly, Dixon observes that technology always has first-order and second-order effects. The automobile’s first-order effect was faster travel; its second-order effects were suburbs, highways, and trucking. Social media’s first-order effect was sharing lunch updates; its second-order effect was restructuring information flows such that insurgent political movements — Trump, Bernie Sanders — could bypass institutional gatekeepers. AI’s second-order political effects, Dixon thinks, will be similarly transformative and are barely visible yet. His one policy preference: strong open-source AI, so that the models shaping political information are auditable and pluralistic rather than controlled by a handful of firms.

NFTs, digital ownership, and the wallet as identity

Dixon defends NFTs — non-fungible tokens, meaning unique digital assets as opposed to interchangeable currency tokens — against their caricature as speculative JPEGs. His argument is architectural. The internet has no native ownership primitive: your Twitter followers are owned by Twitter, your game items are owned by the game studio, your professional reputation on LinkedIn is owned by LinkedIn. A crypto wallet breaks this dependency. The user holds a persistent inventory of digital assets — tokens, credentials, game items, creative works — that they carry across services. The service provides the interface; the wallet provides the ownership layer.

He sees this as increasingly important as AI generates more digital content. NFTs could function as the property-right system for AI-created works, and as the payment and identity layer for AI agents transacting with each other. Cowen half-jokingly suggests AIs will design their own superior tokens; Dixon thinks it is quite possible once AI systems acquire emergent capabilities and the legal infrastructure to hold and transfer assets.

On the question of whether AI agents will prefer Bitcoin or stablecoins, Dixon floats the case for Bitcoin made by David Marcus (former head of Facebook’s Libra project): Bitcoin is the only crypto asset that is credibly globally neutral — no country perceives it as a national instrument — which may make it the natural currency of AI-to-AI transactions, where volatility matters less because machines can settle in milliseconds.

Venture capital, philosophy, and the limits of intelligence

Asked about AI’s implications for a16z itself, Dixon distinguishes the picking function of venture capital (evaluating companies, which AI could plausibly do well) from the high-touch relational function (persuading founders to take your money, then helping them navigate crises). He doubts AI can fully substitute for the latter in the near term, but acknowledges the question is live — he and Marc Andreessen have discussed building a blockchain-based endowment that uses AI to allocate and distribute grants, modelled loosely on the Fast Grants initiative.

Dixon’s philosophical reading surfaces in the conversation’s final third. He studied analytic philosophy in graduate school before dropping out, and has spent recent months working through continental philosophy — currently struggling with Heidegger’s Being and Time, guided by a friend and by lecture videos (he recommends Bryan Magee and Princeton’s Michael Sugrue). The move in philosophy that has most influenced his thinking is Kant’s transcendental argument: rather than deducing a conclusion from premises, Kant asks what preconditions must hold for the premises themselves to be possible. Dixon finds this kind of meta-level reasoning — arguing from what must be true rather than from what appears true — genuinely useful when evaluating technological claims where evidence is sparse and the question is structural.

Related

• Chris Dixon — speaker; general partner at a16z, author of Read Write Own
• Tyler Cowen — host
• Token Economics — concept; stablecoin take rates and incentive design within blockchain networks
• Tokenisation — concept; NFTs and the digital ownership layer Dixon argues the internet lacks
• What Makes a Great Investor — theme; Dixon’s venture-capital framing of the AI-as-mobile-vs-internet question