Jake Brukhman of CoinFund on Why Your Next AI Bet Will Be the Model Itself, Not the Company

Two days before this episode of un# was recorded, Anthropic shipped Claude Fable 5, billed as the most capable model to date. Within 72 hours, a US government export-control directive forced the company to disable it for foreign nationals worldwide and pull it offline.

Most people read that as an AI-safety story. Jake Brukhman, founder of the digital asset firm CoinFund, read it as a real-estate story. What a government can switch off is not the algorithm, the talent, or the data. It is the physical GPU data center, which cannot be moved, hidden, or decentralized. That chokepoint is the thesis behind everything Jake is now betting on.

"Is this really the world that we want? A world where a single AI company that has the best model can decide who can see what."

For the version of this conversation watch the full episode of un# before reading on.

A contrarian, by inheritance

Jake was born in St. Petersburg and came to the US in 1990 with computer-scientist parents. His father became one of the first programmers on Wall Street, and Jake learned C at 14. Host Sri Misra, a former Tata dealmaker who exited Milk Mantra before founding aarnâ protocol, drew the parallel on air: the same profile as the man who built Ethereum.

"I come from a cultural background with a communist government and a lot of distrust of government. That skepticism was very necessary in the early days of crypto."

He bought his first Bitcoin in 2011, then ignored it for seven years. CoinFund launched in 2015, roughly 29 days before Ethereum's mainnet, when there was no settled idea of a smart contract and no professional capital in the asset class.

"In 2015 there were no professional LPs looking for exposure to crypto. Our first fund was friends and family. Today we have sovereign wealth funds."

DeFi did not die. It became a feature.

Sri opened with the reframe he has made publicly: there is no DeFi anymore, there is on-chain finance. Jake agreed, and added the part purists dislike. Crypto has never been more successful and never more narrow at once.

"Blockchain and crypto have never been more successful than they are now, but they're finding product-market fit in much fewer areas."

What survived is unglamorous: payments, stablecoins, settlement, tokenization. The unlock was regulatory. The GENIUS Act, signed in July 2025, gave stablecoins a federal framework, and the pending Clarity Act would split crypto oversight between the SEC and CFTC. Jake reads tokenization not as a buzzword but as capital efficiency. It collapses settlement from days to near-instant, which is why regional banks are already tokenizing deposits. The bigger prize is liquidity.

"I own a motorcycle and it's valuable, but to turn it into money takes months. Before, I could get a home-equity loan. I couldn't get a motorcycle-equity loan. Now I can."

You could always borrow against your house, never against your Tesla stock to buy a coffee. Tokenization turns static assets into usable capital. This is not theory for Sri, whose aarnâ is launching India's first regulated tokenization platform for financial assets, starting with private credit. On-chain real-world assets have grown from under $3 billion in 2022 to more than $24 billion today, with institutional forecasts running into the trillions by 2030.

Own the model, not the company

Here is the thesis most listeners have not heard. The AI labs are heading toward trillion-dollar IPOs, and by the end of 2026 the industry will have spent about $600 billion on infrastructure. Inference, the act of answering a query, already generates an estimated $60 to $65 billion in annual recurring revenue, a number some analysts project could reach $500 billion to $1 trillion within five years. Almost no one can buy in. Retail cannot access OpenAI or Anthropic, and the private market, in Jake's words, is a bubble waiting to collapse.

"The average investor getting AI access today is much more likely to get it through a digital asset than on the public market."

The mechanism is decentralized training. A model is split across thousands of untrusted machines, so no participant holds the full set of weights. Those fragments are tokenized, and the token entitles holders to the model's inference revenue. Open models cannot capture value once their weights are public; closed models can, but invite the Fable 5 shutdown. The tokenized model is the third path.

"You have this model that's valuable, but it's also publicly owned. There's no one entity who can censor the whole thing or take it offline."

Four years ago, experts laughed at the idea of training frontier models on gaming PCs over home internet. The trajectory since has been steep: from Prime Intellect's 10-billion-parameter run, to Templar's 72-billion model on Bittensor, to a 100-billion run by Macrocosmos, all on commodity hardware. CoinFund has co-led seed rounds across the stack, including Prime Intellect and Pluralis, the one firm splitting the weights themselves.

Jake closed with the story that holds the whole argument. After Google published, then withheld, a result threatening the encryption under Bitcoin and Ethereum, an open swarm of AI agents coordinated by Eigen Labs set out to rebuild it.

"In about a week the agents reconstructed the same algorithm Google kept private. And then they improved it by over 40%."

The bookend writes itself. A government can switch off one company's model overnight. A loose swarm of agents, owned by no one, cannot be switched off at all. That gap is the business Jake has spent eleven years funding.

Watch the full conversation with Jake Brukhman on un# with Sri Misra.