We spend a good part of our research time on two questions that sound different but are really the same question twice. First: if generative AI is a change of technological paradigm rather than another hardware cycle, how should long-term capital be positioned? Second: if money, securities, and collateral are migrating from 1970s settlement rails to public blockchains, who actually earns from that migration? This note summarises the framework we use to answer both — one philosophy, applied to two different wars.
Paradigms, not cycles
Every few decades the economy rebuilds its production stack around a new base technology — steam, railways, electrification, the microprocessor. The AI build-out has the anatomy of such a shift, not of a cycle: hyperscaler capital expenditure is running at over $600 billion a year, data-centre projects are now measured in gigawatts rather than megawatts, and the competition between platforms is existential — nobody can afford to stop spending. On the settlement side the picture rhymes: stablecoins have passed $320 billion and keep growing even in months when crypto prices fall, tokenised real-world assets are setting records month after month, and regulators have moved from resisting the new rails to laying them — tokenised equities can now trade on national exchanges interchangeably with ordinary shares.
Paradigms have a mathematical signature: returns are distributed by power laws. A handful of names capture most of the value, and nobody can reliably say in advance which ones. The honest response is a wide basket rather than a sniper shot — and a discipline that never trims a winning position to zero.
The lens: arsenals versus gladiators
During a gold rush the most reliable money was made not by prospectors but by the sellers of shovels, jeans, and whisky. We call such companies arsenals: they are paid for the volume of the conflict, not its outcome. An arsenal sells to every side at once, sits at a chokepoint with no route around it, and converts that position into pricing power. A gladiator, by contrast, is a side in the war — its revenue depends on its own model, platform, or protocol winning. Gladiators offer the fifty-fold outcomes; they also populate the graveyard.
Every name we look at faces the same three questions: Who pays this company? Will they still be paid whichever side wins? And what, specifically, would have to happen for the payments to stop? The answers come from the revenue mix in percentages — not from the investor presentation.
The barbell
The lens becomes a portfolio through a barbell: a heavy core of arsenals — roughly two-thirds of the book, anchored by a small “bedrock” of highest-conviction chokepoints — and a light tail of gladiators, split into many positions of a fraction of a percent each. Stability comes from the core; convexity comes from the tail; and nothing heavy is allowed to sit in the middle without a clear answer as to what, exactly, it is paid for.
Rails you can only own through equity
The crypto-native reflex says: if you want to own a protocol, buy its token. The most important observation of the second book is that the companies at the heart of the new financial world have no token and never will. Circle, the issuer of USDC, has no token — its entire economics lives in a listed share. The same is true of the largest fiat on-ramp and institutional custodians. This is structural, not accidental: regulated financial businesses — dollar liabilities, brokerage, custody — cannot distribute their economics through a protocol token; licences and supervision demand a classical corporate form. The irony of the era is that the issuers of tokens do not have tokens. For an investor, equity is therefore the only door to the cash flows of the rails.
The second axis: decoupling from the coin price
Crypto equities need one more test that AI names do not: how far is the revenue decoupled from the price of Bitcoin? We sort every name into three classes. Class A — rail revenue: paid for infrastructure adoption that grows through coin cycles — stablecoin float, tokenisation of assets, custody basis points, compliance subscriptions. Class B — volume revenue: paid for activity; volatility in either direction feeds it, but a deep multi-year winter compresses it. Class C — price revenue: treasury wrappers, miners, ETF shells whose value is simply a function of the coin price. Class C scores zero in our map — anyone who already holds the coins directly owns the same exposure cheaper and without corporate risk.
Discipline over forecasts
The machinery around the theses matters more than the theses themselves. Conviction is scored 1–10 per name and normalised into percentages, so the book self-adjusts whenever a name enters or leaves. Profits are taken by a fixed ladder — a slice of the remainder at pre-set gain levels — which returns the entire original outlay while still leaving roughly half of the position running for the power-law tail. Averaging down is forbidden: if conviction falls, the position leaves the book rather than getting cheaper. And there is no leverage anywhere — a portfolio that cannot be liquidated by force is the only kind that can genuinely hold a paradigm bet for a decade.
Two wars, one set of rules: pay for chokepoints, size the lottery honestly, let the ladder do the selling, and never let anyone else decide when you exit.