Open source models & AI value accrual w/ Christian Catalini

Christian Catalini argues the AI market will split between frontier models and open source models, with the best analogy being iPhone versus Android rather than the crypto protocol layer comparison often cited. Frontier labs can command a premium the way Apple does, but distillation, leaked traces, and talent movement make it nearly impossible to sustain model capability as a durable secret. Foundation models cannot function like the Coca-Cola recipe, leaving only a short monetization window before open source catches up. Pharma and cybersecurity are among the few domains where frontier capability commands a lasting premium due to IP protection or defense requirements.

The defensible advantages Catalini identifies are narrow. Proprietary data that nobody else has, such as exclusive Bloomberg-style feeds used for training, is one. The weights of an individual human brain shaped by unique experiences are another, though tools like Claude computer use are getting closer to capturing individual digital behaviors and increasing substitution risk. He argues that terms like agency, taste, and judgment are too vague to be useful because LLMs can already perform what was called judgment six months ago.

His central thesis is that if intelligence becomes cheap and commodified, the new bottleneck will be verification, defined as applying human judgment to determine whether a model output is correct, including EVALs and feedback loops. Regulated industries like financial services and healthcare will demand much more verification of model behavior. He believes crypto primitives around identity, provenance, and digital traces will become important again in this context, and that end-to-end verifiable models covering source data through model behavior will be required, not just verifiable inference. The infrastructure will likely require a mix of decentralized crypto architecture and advanced cryptography, but remains at very early stages.

Catalini sees a growing enterprise desire for control over AI, not just cost reduction. Microsoft has shifted strategy to promising enterprises that intelligence will be theirs to control, and Palantir has come out in favor of AI model sovereignty. He believes there are far more companies like Microsoft than Palantir that will resist conceding to Anthropic and OpenAI, and that foundational labs will likely pursue vertical integration, potentially acquiring law firms and insurance firms to access proprietary data. On hardware, whoever controls inference hardware will have a structural advantage, and orchestration platforms like OpenRouter will route workloads between local inference, cloud, open source, and frontier models depending on task difficulty. He credits Elon Musk with outplaying competitors by advancing compute first and renting it out, and describes Nvidia cultivating open source models as a smart strategic play.

Catalini draws an explicit analogy between open source AI models and Bitcoin, arguing that Western open source models serve the same escape valve function for intelligence that Bitcoin serves for money. He predicts there will be a moment when people view a foundational AI lab as a competitor or threat making disagreeable choices, creating demand for an exit option analogous to what crypto provides. On decentralized inference and training, he sees no demonstrated demand today, and is skeptical the rebel alliance narrative will hold, noting that decentralized social media failed despite users claiming to want privacy and data ownership.

On Libra, Catalini says the project was more distributed than publicly credited, with 26 members each holding a vote and Facebook positioned as one wallet competing on an open standard. What killed it was the interaction between lobbying and regulatory pressure, not regulation alone, because payments revenues made Libra existential for incumbent banks. The project obtained a greenlight from Powell but Janet Yellen told officials it would be career suicide, and a Swiss license was ready to be signed until the US interfered. The strategic error was launching with a big bang and attempting to reinvent money rather than starting small with remittance, making it easy for incumbents to target a single point of failure. The key lesson he draws is that any project with a single point to choke will be taken out.

Catalini is direct that crypto going mainstream means it will be domesticated and there is no other path. Stablecoins will become more boring and more compliant, consumers will use them without realizing it, and the number of countries where regulatory arbitrage remains viable is narrowing. His concern is that if payments are reinvented on a corporate chain, fees will start low and be raised over time, replicating the card network dynamic on different infrastructure. On agentic payments, flows on existing rails for tasks like booking flights will be dominated by incumbents and deals between foundational labs and payment processors. The genuinely interesting use case is machine-to-machine interactions where agents source data, borrow, buy, and rehypothecate in modular ways, though he cautions that most current volume in AI plus crypto remains meme coins, consistent with how every new vertical in crypto opens.