The number arrived in my feed the way a malformed transaction arrives in a block explorer: every field populated, every check green, until I reconciled the balance.
Harvard's endowment, one report declared, had reached $68 billion, powered by "SpaceX and AI bets." Two nouns. One number. One implied conclusion โ that artificial intelligence is now minting institutional wealth at scale, and that the smartest, most patient money in the world is already deep inside it.
I pulled the last four fiscal years from public filings and from memory. FY2023: roughly $50.7 billion. FY2024: roughly $53.2 billion, a reported return near 9.6 percent. To arrive at $68 billion from $53.2 billion in a single fiscal year requires roughly a 28 percent gain, applied to one of the largest and most diversified pools of capital on the planet.
That is not a return. That is an event. And the event ships with no source, no methodology, and no line separating realized gains from paper marks. Just the number and the two names behind it.
When a headline hands you a 28 percent move and two proper nouns as the explanation, you are not reading financial reporting. You are reading a memo whose function is to make a valuation feel inevitable. I have seen this exact pattern before โ in token launch decks, in "audited" contracts that were never audited, in every pitch that needed a number to do the work that evidence should do.
Zero knowledge isn't magic; it's math you can verify. So is an endowment. So let's verify.
Context: How an Endowment Actually Touches AI
A university endowment is not a wallet. It is a liability-matching machine wrapped in a portfolio, engineered to pay out roughly 4 to 5 percent of its value every year to the university while preserving purchasing power across decades. Harvard's is managed by Harvard Management Company (HMC), which allocates across public equities, fixed income, venture capital, private equity, real estate, and hedge funds.
The critical structural fact is this: Harvard almost certainly does not hold SpaceX or any AI company directly. It holds limited partner interests in external venture funds โ Sequoia, a16z, Thrive, and their peers โ which in turn hold equity in those companies. The exposure is indirect, layered, and valued by someone else's last-round mark.
Every layer between the university and the underlying asset widens the gap between what the asset is worth and what the books say it is worth. A direct public equity position marks to a live market price every second. An LP interest in a venture fund marks to the most recent financing round โ which may be months or years old, priced by the same insiders who benefit from a high mark.
So when a report says Harvard "bet" on AI, the accurate sentence reads: Harvard's external managers allocated capital to funds that hold AI-related equity, and those funds report valuations that HMC aggregates into a headline number.
That is a different sentence. It carries risk structure. The original carries none.
There is also a timing problem the report ignores. HMC publishes its annual figures each fall for the fiscal year ending in June. Any "AI bet" gain recognized by mid-year is, by construction, unrealized. The only way AI shows up as cash is if a position was sold. Nobody sold SpaceX in 2024 โ SpaceX is not public, and its shares trade only in periodic tender offers and secondary transactions.
I learned to read structures like this the hard way. In 2018 I spent six weeks inside the Gnosis Safe multisig contracts, compiling them on a local testnet and finding signature malleability flaws that earlier reviewers had missed. The lesson was not about Solidity. The lesson was that a structure can look correct at every visible layer and still hide its real behavior one layer down. An endowment is the same object. The visible layer is the headline. The real behavior is in the marks.
Core: Four Reconciliations
I ran this the way I run a contract review. Four checks, in order of dependency. Each one either passes or it doesn't.
Check One โ The Number Does Not Reconcile
The gap is the whole story. A jump from roughly $53 billion to $68 billion is about $15 billion of value creation in one year. For scale, the entire endowment added roughly $2.5 billion in FY2024. The report asserts, implicitly, a year six times larger than the last disclosed one โ with no disclosure to support it.
Two explanations survive scrutiny. Either the number is wrong, or it uses a different denominator: total assets including pledges and separately managed funds, a gross figure rather than the net endowment HMC reports. Both possibilities mean the "$68 billion" as presented is not comparable to the $50โ53 billion series. A number that cannot be reconciled to its own history is not a data point. It is a rumor with a currency symbol attached.
The AMM model hides its truth in the invariant. So does a balance sheet. Change the invariant โ gross versus net, realized versus marked โ and the same assets tell a different story. The report changed the invariant and never said so. That is the first defect, and it is disqualifying on its own.
Check Two โ The Attribution Collapses Under Decomposition
An endowment return is a composite. Public equities, private credit, real estate, hedge funds, venture. In a year when both the S&P 500 and the Nasdaq ran hard, the single largest contributor to any large endowment's gain is almost always public equity beta โ and the largest single public equity story of 2024 was Nvidia, whose market capitalization added more value in one year than most national economies produce.
So the honest attribution is: Harvard's endowment rose because its diversified portfolio rose, and the largest visible engine was public technology equities, with Nvidia as the headline name. Venture marks โ SpaceX, AI startups โ contributed at the margin, on paper.
The report compressed that composite into two nouns. That compression is not a simplification. It is a substitution. It swaps a diversified, beta-driven result for a concentrated, narrative-driven one, because the narrative is what travels across feeds.
I don't trust a return that cannot be decomposed. If you cannot say how much came from public beta and how much from private marks, you cannot say anything about skill, and you certainly cannot attribute the total to two names. This is the same discipline I applied when I traced Uniswap V2's swap function in 2020 and modeled slippage in Python: the fee was real, the arbitrage was real, but only because I could decompose the flow into its parts. A number you cannot decompose is a number you cannot act on.
Check Three โ The Marks Are Illiquid
Here is the part that matters for anyone treating this story as evidence of AI's real economy.
Venture assets are marked to the last round. SpaceX's valuation reached roughly $350 billion by late 2024 through tender offers and secondary transactions. Those are real prices paid by real buyers โ but they are prices for tiny slivers of a private company, not the price at which the whole entity could be liquidated. The mark is an estimate wearing a number's clothing.
The same logic governs AI startups. A round priced at a $40 billion post-money valuation is not $40 billion of value. It is one negotiated transaction, often with strategic or reflexive motives, extrapolated across the entire cap table. When the next round prices lower, the mark moves down โ and the endowment's "gain" reverses without a single share being sold.
An unrealized gain is a hypothesis. A realized gain is a fact. The report presents hypotheses as facts. That is the core defect, and it is the same defect I found in the Axie Infinity breeding-fee logic in 2021: the surface said one thing, the edge case said another, and only an isolated test revealed which was true. Here, the isolated test is simple โ ask what fraction of the $15 billion is cash. The answer, in any honest accounting, is close to zero.
Check Four โ The Reflexivity Loop
Now the structural insight, the one the report never reaches.
Endowment capital flows to AI through a loop: AI valuations rise, so endowment marks rise, so endowments allocate more to AI, which pushes valuations higher again. This is reflexivity โ the observation feeding back into the thing observed.
The loop is self-reinforcing on the way up and self-accelerating on the way down. It is not a Harvard problem. It is a system problem. University endowments collectively manage hundreds of billions of dollars and are among the most important LP sources in American venture capital. Their allocation preference directly sets how much capital AI startups can raise.
When the report frames the endowment gain as proof of AI's strength, it inverts the causality. The gain is partly a symptom of the same valuation cycle it is cited to validate. The thermometer is being used as the heater.
What an Honest Disclosure Would Look Like
Because the report gives us nothing to verify, it is worth stating what verifiable would look like. Five items. First, the net endowment value, defined exactly as HMC defines it, comparable to the prior four fiscal years on the same basis. Second, the split between realized and unrealized gains, itemized. Third, the percentage of the portfolio in illiquid assets, with the marks dated to the quarter in which they were set. Fourth, the concentration ceiling on any single asset class, disclosed as policy rather than as outcome. Fifth, the named source and the methodology behind every number โ or an explicit statement that the figure is an estimate.
An honest report would look like an audit trail. This one looks like a press release with the source field deleted. That asymmetry is itself the finding. When a claim cannot survive five ordinary questions, the claim is not information. It is positioning.
Contrarian: The Real Risk Is Downstream, and Nobody Prices It
Everyone reading this story is debating whether AI is a bubble. That is the wrong question, and it is the question the narrative wants you to ask.
The buried question is what happens to the institutions on the other end of the loop. If AI valuations correct 30 to 50 percent โ an ordinary move for a frothy private market, not a catastrophe โ the transmission runs in a straight line: AI companies mark down, venture funds mark down, endowment marks fall, university budgets tighten. Harvard's annual operating scale is in the range of $5โ6 billion. Endowment distributions fund a large share of it. Scholarships, research grants, and tuition support sit downstream of a mark-to-market number that no one controls and few disclose.
There is an irony the report cannot see. Universities are simultaneously profiting from AI and being disrupted by it โ their core products are teaching and knowledge production, precisely what large language models commoditize. The endowment's "win" and the university's mission may be in direct conflict. A financial position that benefits from the thing eroding your institution is not a hedge. It is a bet against yourself.
And beneath the financial problem sits a disclosure problem. The report attributes a precise-looking number to no named source, publishes through a crypto-vertical outlet on a topic far outside its beat, and formats every claim as a discrete, sourced bullet. That pattern โ atomic, unattributed, list-shaped โ is the fingerprint of aggregated or machine-generated copy. I watched the same pattern spread through the 2024 ETH ETF coverage, where custody analyses were recycled without anyone reading the actual filings. If this copy becomes the citation for "AI made Harvard $15 billion," the rumor inherits the authority of the institution it misdescribes.
The people who should be most skeptical are the ones least equipped to check. A donor reads the headline. A trustee repeats it. A consultant builds it into a model. Each repetition strips a layer of doubt until the mark becomes a fact by sheer weight of citation.
Takeaway: Watch the Marks, Not the Headlines
The next hard data point is HMC's official annual report, published each fall for the prior fiscal year. That filing will either confirm the number or quietly bury it. Track it.
Then track three signals. Watch the marks on the largest AI rounds, quarter by quarter, and wait for the first down-round in the private market โ that is the moment the loop reverses. Watch SpaceX's path to liquidity, because until there is a real exit, every dollar of "gain" is a hypothesis. And watch the coupling between higher-education finance and the technology capital cycle, because that is where the risk actually lands โ not on a balance sheet in Cambridge, but on a scholarship that quietly does not get funded.
The lesson is not that AI is a bubble. The lesson is that institutional finance is increasingly priced off marks it cannot verify and cannot sell โ and the institutions most exposed are the ones we assumed were the safest.
Zero knowledge isn't magic; it's math you can verify. An endowment should be the same. When a report gives you a number you cannot reconcile to its own history, a cause you cannot decompose, and a mark you cannot liquidate, you are not looking at wealth. You are looking at a position that has not been closed. Markets, unlike universities, eventually require the close.
