On a Tuesday morning in Chengdu, I opened a filing summary I had been avoiding for eleven days. David Tepper's Appaloosa Management had allocated 40% of a $7.7 billion fund to three AI stocks. Three. Out of the thousands of listed companies in the United States alone. The number that stayed with me was not 40. It was three.
I have spent most of my adult life inside governance systems that are built to prevent exactly this. In 2020, during DeFi Summer, I ran a governance working group at MakerDAO and reviewed more than five hundred voting proposals. One of the recurring fights was over concentration — how much of the collateral portfolio could sit in a single asset type before the system became a single point of failure dressed up as a strategy. We lost that argument more than once. Whales prefer conviction to diversification, and conviction is easier to market.
So when a macro fund that has survived every regime since 1993 decides that two-fifths of its capital belongs in three tickers, I do not read it as a stock tip. I read it as a governance signal about who will own the substrate the rest of us build on.
David Tepper does not need an introduction to anyone who traded through 2009 or 2020. Appaloosa Management made its name on distressed debt and macro dislocations, and Tepper's public record includes buying beaten-down financials when the consensus had already written them off. He is often described, a little breathlessly, as one of the greatest traders of his generation. The description is not wrong, and that is precisely why the filing matters.
The source of the report I read was a brief carried by Crypto Briefing, a publication whose center of gravity is digital assets rather than institutional equities. That is worth stating plainly, because it shapes what the article actually contains. The brief highlights the 40% allocation and characterizes the bet as a play on tech infrastructure. It does not name the three companies. It does not say whether the position was newly built or inherited from price appreciation. It does not tell us which quarter the filing covers, or whether Tepper still holds it.
Those omissions are not incidental. They are the entire analysis.
I have a habit, developed over years of reading secondary reporting on governance votes, of going back to primary sources. The primary source here is a filing, and filings are better than summaries for a simple reason: they contain the parts that do not fit the story. A summary tells you the percentage. A filing tells you the share count, the option overlay, the puts, and the fact that a portion of the position may be hedged in ways that reverse its apparent direction. Based on my audit experience, the gap between what a secondary report says a position is and what the filing shows the position to be is frequently large enough to change the conclusion.
Here is the mechanical problem. Any US fund managing more than $100 million must disclose its long equity positions on Form 13F within 45 days of quarter end. The rule was designed for transparency, but it produces a strange artifact: by the time the investing public reads a position, the manager may have already closed it. A 45-day lag during a period when the Nasdaq can move 15% is not a disclosure regime. It is a historical document.
I have written elsewhere about how digital artifacts need to be read as historical documents rather than living claims, and the same discipline applies here, for the same reason. A 13F is a snapshot of a decision that has already been made and may already have been unmade. Treating it as a current signal is like treating a signed lease as proof that someone still lives in the apartment.
The first thing to separate is active conviction from passive drift. If three positions each rose 70% while the rest of the book rose 12%, a 25% allocation becomes a 37% allocation without a single buy order. The 40% figure in a headline cannot distinguish between a manager who chose concentration and a manager who simply did not rebalance. The distinction is not academic. One signals forward-looking conviction; the other signals administrative inertia. Anyone who wants to verify this can pull the prior quarter's 13F, compare share counts rather than dollar values, and look for increases in the number of shares held. Dollar figures move with price. Share counts move with decisions.
The second thing is the arithmetic of the downside. In conventional institutional practice, a single-name position is usually capped somewhere near 10% and a sector exposure near 30%, though these are conventions rather than rules and macro funds break them routinely. At 40% in a single theme, a 25% drawdown in those three names costs the fund roughly 10% of total assets. That is not a mark-to-market nuisance. That is a letter to limited partners, a redemption conversation, and possibly a headline that changes the fund's fundraising trajectory for two years.
The third thing is what "tech infrastructure" actually means in 2026. The phrase points away from application software and toward the physical and near-physical layers: advanced packaging, memory bandwidth, datacenter power, cooling, interconnect. Capex guidance from the major cloud operators has run well above 30% year over year, and the spending is not primarily on models. It is on buildings, transformers, and the machines inside them. When a macro manager says infrastructure, he is describing a category of asset that behaves less like a technology stock and more like a utility with a monopoly problem.
A note on where the bottleneck actually sits. Every conversation I have had with infrastructure operators in the last eighteen months has circled back to the same constraint, and it is not silicon. It is interconnection — the queue to connect a new load to a transmission network, which in parts of the United States now runs several years. Power purchase agreements signed long in advance, transformers with lead times measured in quarters, water rights for cooling in arid regions. A fund that buys three large-cap names is not buying that. It is buying the balance sheets of companies that can afford to wait in the queue. The distinction matters because the queue is a physical limit, and financial capital does not shorten it.
This is where my own work intersects, and not comfortably.
In 2025, I spent six months designing the governance structure for CivicChain, a DAO focused on municipal data sovereignty. My job was to mediate between regulators and developers, translating statutory language into enforceable clauses. About four months in, I had a conversation with a city official that I still think about. She asked a simple question: if the data lives in someone else's datacenter, whose law applies? I had prepared answers about jurisdictional arbitrage and cryptographic access control. None of them survived contact with the question.
She was asking about sovereignty. And sovereignty, in the end, is a function of who owns the building.
That is why a hedge fund allocation is a governance event. Equity holders in the infrastructure layer receive residual claims, board influence, and information rights that are enforceable in court. Token holders in an "AI-adjacent" protocol typically receive a governance vote on a parameter they cannot change and a token whose value depends on a narrative they do not control. I watched this asymmetry up close during the MakerDAO debates, where algorithmic neutrality repeatedly produced outcomes that fell hardest on the smallest collateral holders. The system was fair in its mechanics and regressive in its effects. That is not a bug in code. It is a design choice that nobody wanted to name out loud.
The fourth thing is where the honest Bitcoin-AI convergence actually lives, and it is not where the marketing says it does. For two years I have watched projects describe themselves as Bitcoin Layer 2s when the underlying architecture was an Ethereum bridge with a rebranded front end. The Bitcoin community's own developers have been notably unmoved by most of it. What has genuinely changed is less glamorous: mining operators converting sites to high-performance computing, reusing their interconnection queues, their substations, and their relationships with grid operators to host inference and training workloads. That is not a narrative. That is a capital allocation decision made by people who already owned the hard part — power and land.
The fifth thing is that verification is the only place where crypto has a defensible claim on the AI stack. If inference is concentrated in a handful of datacenters owned by three or four firms, then proving that a model ran as claimed becomes a first-order problem — for auditors, for regulators, for anyone who needs to trust an output they did not compute. Zero-knowledge machine learning, trusted execution environments, and proof-of-inference systems are technically hard, commercially unproven, and almost entirely absent from the current token narrative. They are also the only pieces of this puzzle where the crypto toolkit offers something the incumbent stack cannot replicate at the same cost. I say this as someone who has audited enough of these systems to be skeptical of most of them. The schemes that work are boring. The ones that raise the most money are not. Curating the soul in a world of derivative clones is what this industry has always actually been doing, whether or not it admits so in the pitch deck.
Now the part that the headline gets wrong.
Everyone reading that filing sees conviction. I see something closer to a confession. A macro fund's edge has historically come from reading the cost of capital, the direction of rates, the fragility of a currency peg. When the most respected macro manager of his generation concentrates 40% of his book in three equities, one plausible reading is that he has decided macro alpha is no longer reliably available and that the only remaining asymmetric bet is structural. That is not a bullish statement about AI. It is a bearish statement about the tradable macro landscape.
There is a second blind spot, and it is the one that matters most to the people I write for. The crypto market has produced an enormous population of what I can only describe as clones — tokens that borrow a thesis from a sector they do not operate in, attach a governance vote to it, and sell exposure to a capex cycle whose contracts, permits, and power agreements they will never hold. Curating the soul in a world of derivative clones is not a metaphor here. It is a description of the asset class. When the person who actually owns the datacenter announces his position, derivative claims on that same thesis do not become more valuable. They become legible as what they always were.
And a third blind spot, quieter: concentration is not the same as understanding. A 40% position tells you about the size of a belief. It tells you nothing about whether the belief is correct. Some of the largest, most confident positions in financial history were also the most wrong, and they were wrong in exactly the way that concentration makes possible — slowly, then completely.
So what do I do with a 45-day-old snapshot of someone else's conviction?
I file it under context, not instruction. I pull the prior quarter's share counts. I watch whether other large managers move in the same direction, because a single filing is an anecdote and four filings pointing the same way is a regime. I ignore the tokens that will inevitably rebrand around this headline by Friday.
I have spent a career curating the soul in a world of derivative clones, and it has taught me one durable lesson: the narrative is always the cheapest layer to build and the first layer to collapse. And I keep asking the question that city official asked me, because it has not gotten easier. If the compute layer consolidates into the hands of three firms and a handful of funds, what exactly is decentralized about the systems we build on top of it? We can verify the outputs. We can refuse to custody the keys. We can write governance that does not pretend to be neutral when it is not. What we cannot do is vote our way into owning the building.
The question worth sitting with is not whether Tepper is right. It is what the rest of us are still building, when the ground beneath the building is no longer ours to govern.
