What the Sellers Know

There is a category of market information that never appears in an earnings release or an analyst model, and it is often the most important information available: what the people with the best seats are actually doing, as opposed to what they are saying. In the first week of July, two of the most consequential actors in artificial intelligence did things that deserve to be read this way — carefully, without hysteria, and without the reflexive comfort of either the bull or bear catechism.

Meta announced plans to build a cloud business selling its AI computing capacity to outsiders. OpenAI, having confidentially filed for an IPO in early June, is reportedly considering delaying that offering into 2027 — while simultaneously discussing a transfer of roughly five percent of its equity to the United States government.

Each of these developments has a perfectly benign explanation. Each also has a less benign one. The intellectually honest exercise — the only one worth doing — is to hold both readings at once and ask which the evidence favors. What we should not do is what much of the commentary has done: declare a “top” in AI, as if tops were legible in real time. They are not. Tops are a retrospective genre. But revealed preference is available in the present tense, and right now the revelations are worth our attention.

The landlord’s confession

Start with Meta. The company told investors it would spend as much as $145 billion this year on data centers and chips — a sum larger than the GDP of most nations, deployed on the premise that its own AI ambitions required every watt of it. Now, per Bloomberg’s reporting, Meta is standing up a business to sell that capacity to others: hosted access to its models in an arrangement resembling Amazon’s Bedrock, or raw compute in the manner of CoreWeave.

The charitable reading is straightforward, and it is not stupid. Capacity at this scale arrives in enormous, indivisible increments, timed against demand projections that are inherently uncertain. Training clusters sit idle between runs. Selling the trough is what a rational operator does. Amazon Web Services — the most profitable accident in corporate history — was born of precisely this logic. On this view, Meta is not confessing anything; it is simply maturing into a utility.

But the charitable reading has to survive contact with two facts. The first is that Meta, unlike Google and OpenAI, has generated no visible external demand for its own AI products. It does not break out revenue for Meta AI or its Llama models, and its executives speak mostly of internal uses. The infrastructure was built, at least in part, to serve a commercial AI franchise that has not yet materialized. When the models don’t sell, the real estate becomes the product. That is not diversification. That is a pivot, however elegantly dressed.

The second fact is stranger: Meta is simultaneously one of the largest renters of other people’s compute, with commitments approaching $48 billion to CoreWeave and Nebius. A company paying tens of billions to lease capacity while announcing it has surplus capacity to sell is engaged in something — capacity arbitrage across chip generations and geographies, perhaps, or narrative management for a shareholder base visibly anxious about the capex line. The market’s reaction was telling in both directions: Meta’s stock rose roughly eight percent on the news, while CoreWeave and Nebius each fell about twelve. Investors rewarded Meta for finding a revenue story and punished the pure-play landlords for acquiring a competitor with infinitely deeper pockets. Read together, the message is that the market has begun to reprice compute itself — from scarce strategic asset toward rentable commodity. Commodities have margins. Scarce strategic assets have moats. These are different investment theses, and the industry has been priced on the second one.

And Meta is not alone. Weeks earlier, SpaceX — via its absorbed xAI operation — leased the entire capacity of its Colossus 1 facility to Anthropic for a reported $1.25 billion per month, then signed further capacity deals with Google and Reflection AI. A pattern is forming, and it is worth stating plainly: the companies that built the largest AI infrastructure in human history, explicitly for their own use, are becoming landlords of it. When the biggest buyers of compute become sellers of compute, that is information about their internal estimates of marginal return — estimates we are not otherwise permitted to see.

The company that would not be counted

Now consider OpenAI. On June 8th, the company confidentially filed its S-1, with a public debut penciled in for the second half of this year. Within weeks, reporting emerged that the offering may slip to 2027. The stated logic is valuation: Sam Altman is said to want a trillion-dollar debut, and the aftermarket behavior of SpaceX — which spiked above $225 before settling back near its $150 offer price — suggested the public market might not underwrite that number today.

Is the delay about reticence to expose the numbers? I suspect that framing is too simple and also not simple enough. Valuation management and disclosure avoidance are not separate motives; they are the same motive viewed from different angles. The numbers that a public OpenAI would have to defend, quarter after quarter, reportedly include a cash burn on the order of $27 billion this year and $63 billion next, with breakeven not expected before 2030. A private company gets to narrate those figures. A public company gets interrogated on them, every ninety days, under oath of a sort. Delaying the IPO does not merely wait for a friendlier tape — it extends the period during which the most important financial claims in the technology sector remain unfalsifiable. I have written before about deferred disclosure as an externality of this era: the cost of not knowing is borne by everyone allocating capital in the sector, while the benefit of not being known accrues entirely to the issuer. The June filing followed by a June delay is that dynamic performing itself in public.

Then there is the government stake. Per the Financial Times, OpenAI is in early discussions to transfer roughly five percent of its equity — about $42.6 billion at the last private mark — to a federal vehicle modeled on the Alaska Permanent Fund, framed as a mechanism for sharing AI’s gains with the American public. The question I keep being asked is whether this is an attempt to purchase the administration’s blessing ahead of an IPO that might otherwise falter. That reading is available, and one should not be naïve about the transactional instincts of the current White House. But I think the sharper reading is defensive rather than promotional. Senator Sanders has introduced legislation proposing a one-time fifty percent equity levy on large AI companies. Against that backdrop, a voluntary, passive, five percent stake is not generosity. It is an insurance premium — a regulatory put, purchased with dilution, that buys alignment with the sovereign and a counter-narrative to confiscatory proposals. Five percent without voting rights is cheap relative to fifty percent with board seats. One need not doubt the sincerity of the wealth-sharing rhetoric to notice that the structure is optimized for protection, not philanthropy.

Two caveats, because honesty requires them: the talks are preliminary, no agreement exists, and any such transfer would likely require congressional approval. And there is a live counterexample — Anthropic has reportedly resisted government ownership altogether, proposing instead a dividend funded by future sector taxes. The industry has not converged on selling equity to the state. One flagship is considering it. That, too, is information.

What this is, and what it is not

So: is this “the top”? I want to be disciplined about what can and cannot be claimed.

It is not a top call. The benign readings are real. Excess-capacity monetization created AWS. IPO timing is managed by every issuer that has ever existed. Governments have taken stakes in strategic industries before without those industries collapsing. Anyone telling you these events prove the cycle has peaked is practicing astrology with a Bloomberg terminal.

What can be claimed is narrower and, I think, more useful. For three years, the AI trade has rested on a single load-bearing assumption: that demand for compute is effectively infinite relative to supply, and that every dollar of buildout will be absorbed by internal need. That assumption was never tested, because the parties with the data — the hyperscalers and the frontier labs — had no reason to reveal it. In the past month, they have begun to reveal it, not in words but in transactions. Meta’s internal demand projections evidently no longer absorb Meta’s internal supply. SpaceX’s did not either. OpenAI evidently prefers another year of unaudited narrative to a public accounting at today’s prices and is willing to spend forty billion dollars of shareholder equity buying political shelter for whenever the accounting comes.

None of these actors believe AI is finished. All of them, judging by their behavior rather than their keynotes, believe something more specific: that the marginal return on the next dollar of compute is lower than the marginal return on the last one, and that the wise move is to start selling what they were recently only buying — capacity, equity, exposure — to whoever still believes otherwise.

In every cycle I have invested through, that moment has a name. It is not the top. It is the moment the smart money starts distributing to the faithful. The distance between that moment and the top is unknowable. The direction of the information is not.

The honest position, then, is neither panic nor complacency. It is simply to insist on the oldest discipline in markets: when the people who know the most begin behaving differently than they speak, believe the behavior.