The Tell

---

Every quarter, in the routine exhaust of regulatory filings nobody is meant to read closely, a small machine prints the same message in slightly different numbers. A Form 4 here. An automated 10b5-1 disposal there. A few hundred thousand shares, a few tens of millions of dollars, and a name at the top belonging to one of the three men who built CoreWeave out of a failed crypto-mining outfit and rode it to one of the most spectacular debuts of the artificial-intelligence era.

Read one filing and it is nothing. Read all of them together — the way the insider-tracking firm Washington Service does, the way a forensic accountant would — and a different picture resolves. Since CoreWeave's stock came unlocked in August 2025, the company's billionaire co-founders have sold more than $2.3 billion of it. Over the same stretch, across hundreds of insider transactions, the number of shares any insider has bought on the open market is, as best the filings show, zero.

That is the whole story in a single line, and it is worth sitting with before the noise of the bull case drowns it out. The people with the most complete view of this business — the men who know exactly what its chips are worth, what its contracts really say, and how fast its hardware is decaying — have spent the months since their lockup expired turning conviction into cash. The people being asked to buy what they are selling are told to keep their eyes on the growth.

The tell

In a poker game the move that matters is rarely the one a player announces. It is the involuntary one — the tell. CoreWeave's tell is the tape of its own insiders.

Chief among the sellers is co-founder Brian Venturo, the chief strategy officer, who has personally offloaded more than $1.1 billion in stock — enough to make him, by Washington Service's tally, the second-largest individual insider seller by value anywhere on the US market this year. Alongside him, chief executive Michael Intrator and co-founder Brannin McBee have steadily pared their own holdings. Together the three cut their combined stake by roughly a quarter and now control about 18% of the company — Intrator alone around 10.4% — down from the founder-dominant register they carried into the IPO. In one recent sixty-day window the filings logged dozens of insider sales; across the past year, on the order of two hundred. The buy column stayed empty the entire time.

The defense is ready to hand, and CoreWeave will offer it: these are 10b5-1 plans, sales scheduled in advance precisely so executives cannot be accused of trading on inside information. That is true, and less exculpatory than it sounds. A 10b5-1 plan governs when a sale executes; it says nothing about why an executive chose to schedule a disposal of that size in the first place. A founder who believes the next five years will dwarf the last does not, as a rule, arrange to dismantle a quarter of his position the moment the lockup lifts. The plan launders the timing. It does not launder the decision.

Nor are the founders alone at the exits. Magnetar, the hedge fund whose early backing helped turn CoreWeave from a curiosity into a contender, has been selling in a steady drip for months — a near-continuous stream of disclosed disposals through the spring, reported to have carried billions out the door and cut its stake below ten percent. The one large holder conspicuously not selling is Nvidia — and that exception, as we will see, has less to do with faith than with the fact that Nvidia is selling CoreWeave something far more lucrative than shares.

The melting machine

To see why the people inside might be in a hurry, you have to see what CoreWeave actually owns. Strip away the language of "AI cloud" and "hyperscale infrastructure" and the company is, at bottom, a landlord renting out a warehouse of graphics chips. Its business is to buy enormous quantities of Nvidia GPUs on borrowed money and rent time on them to companies training AI models.

The trouble is the central, awkward fact of the hardware: it does not age like a building. It ages like fruit.

CoreWeave depreciates its GPUs over a useful life of about six years — a figure it quietly raised from five back in 2023, a change that lowers the depreciation charged against each quarter and flatters the result. Many people who study this silicon for a living think six years is a fiction. Data-center GPUs run flat out at high utilization, and credible estimates of their economically useful life run as short as one to three years before a faster, more power-efficient generation makes the old inventory uncompetitive. Michael Burry — the investor who shorted the housing bubble — has gone public accusing the big AI-infrastructure players of systematically understating depreciation by stretching useful lives, and has put the cumulative gap across the largest five at roughly $176 billion between 2026 and 2028.

CoreWeave's own numbers show this is no abstraction. In 2025 the company recognized about $2.45 billion in depreciation and amortization — close to half of its $5.13 billion in revenue — and that is at the generous six-year assumption. The expense is the accounting system's way of admitting the machines are wearing out. Shorten the useful life toward what the skeptics believe is real and the charge swallows the business whole. It is no coincidence that, even as revenue more than doubled, CoreWeave's net loss widened — to about $1.17 billion in 2025 from $863 million the year before. The faster it grows, the more hardware it must depreciate, and the deeper the hole.

CoreWeave's answer is that its older chips hold their value better than the bears claim — that 2020-era A100s remain fully booked and 2022-era H100s still command most of their original rental prices. Perhaps. But that is an argument about the recent past, made at the peak of a capacity shortage, by people who are simultaneously selling their own stock as fast as the rules allow. The tell, again, undercuts the testimony.

The debt under the ice

A melting asset would be survivable if it were paid for. CoreWeave's is not.

The company has built its warehouse of decaying chips on a mountain of borrowed money. In the twelve months through March 2026 alone it raised something like $28 billion in combined equity and debt, leaving leverage so heavy the debt-to-equity ratio runs into the double digits, with a $4.2 billion debt maturity falling due in 2026. And it is not slowing down: capital-expenditure guidance for the year runs to $31–35 billion — more, in a single year, than the company will book in revenue — as it races to buy the next generation of chips before the last one finishes melting.

This is the machine entire: borrow tens of billions, spend it on hardware that begins depreciating the moment it is racked, rent that hardware out under multi-year contracts, and use the contracted revenue to service the debt and borrow again for the next batch. It works beautifully so long as three things hold at once — that demand keeps climbing, that the chips hold their value longer than the skeptics fear, and that credit markets keep lending against an asset base that is, by its nature, evaporating. Pull any one and the structure does not bend. It snaps.

The circle

There is a pattern that recurs in the richest corners of the AI boom, and CoreWeave is one of its purest expressions: a handful of giants passing enormous sums among themselves, each transaction validating the next.

Consider Nvidia. It is CoreWeave's supplier — the source of the GPUs that are the company's entire reason to exist and the gatekeeper of who gets them first when they are scarce. It is also CoreWeave's investor, having steadily enlarged a holding that has grown from about 6% to roughly 11% of the company, a stake worth on the order of $4.9 billion. And it is, remarkably, CoreWeave's backstop customer: under a deal disclosed in September 2025 and worth up to $6.3 billion running through April 2032, Nvidia has agreed to buy CoreWeave's unsold compute capacity — to be the buyer of last resort for the very chips it sold CoreWeave in the first place. Nvidia sells the shovels, funds the digger, and promises to buy back any gold he fails to strike. That is not a market. It is a closed loop wearing the costume of one.

The customer concentration on the other side of the ledger is just as vivid. Microsoft has been the anchor since before the IPO, supplying about 62% of revenue in 2024 and roughly 67% in 2025. A single customer providing two-thirds of your revenue is not a moat; it is a hostage arrangement, and it binds both ways. CoreWeave has worked hard to diversify — a reported $21 billion Meta contract running toward 2031, vast commitments from OpenAI, a $6 billion deal with Jane Street, an agreement with Anthropic — and now says no single customer will exceed about 35% of contracted revenue. But the new customers are the same small set of AI labs burning the same venture and corporate billions, several of them busy building their own data centers. The diversification is real. So is the circularity it diversifies into.

The contracted backlog those deals produced is the number bulls reach for first: about $66.8 billion at the end of 2025, swollen toward $95–100 billion by the 2026 signings. It is a genuinely staggering pipeline. It is also a promise denominated in the budgets of a few cash-burning customers and payable only if the machines stay valuable long enough to fulfill it.

From crypto to colossus

It is worth remembering what this company was. CoreWeave began life as a crypto-mining operation — a warehouse of GPUs pointed at Ethereum — until the founders had the timing and the nerve to repoint the same basic bet at artificial intelligence. That pivot was genuinely visionary, and it made them billionaires. It is also useful context for the present moment, because it means the founders have done this before: they have watched a GPU boom inflate on the promise that the machines would mint money forever, and they have seen what became of the people still holding the hardware when the economics turned. They know what the end of a compute gold rush looks like. They are, right now, selling.

The market has begun to flinch. After running up roughly 150% from its $40 IPO price — and briefly nearly 3.5 times that, at a spring peak near $138 — CoreWeave has rolled over to around $102, down about a quarter from the high, with short interest above ten percent and chartists pointing at bearish patterns. The valuation still embeds heroic assumptions: a price-to-sales multiple north of eight on a company that loses money, carries double-digit leverage, depreciates nearly half its revenue away, and leans on three customers and one supplier. The bull case is not stupid. It is a bet that the music plays long enough for the backlog to outrun the depreciation and the debt.

The quietest sell signal

None of this is proof of wrongdoing. The 10b5-1 plans are legal. The six-year useful life sits within the range auditors have blessed. The debt is disclosed, the customer concentration is in the filings, the Nvidia arrangements are public. Everything here is hiding in plain sight — which is exactly what makes it a story worth telling rather than a scandal waiting to break.

But step back and watch what the insiders do with the one resource that cannot lie: their own money. The founders who understand this business better than any analyst — who see the real utilization curves, the real replacement schedules, the real conversations with Microsoft — have spent the months since their lockup expired converting $2.3 billion of conviction into cash, and have not, between them, bought back a single share at any price. The hedge fund that backed them earliest has been carrying billions out the door for months. The only large holder standing pat is the one whose true profit comes from selling CoreWeave the chips, not from owning the stock.

When the people who built the machine are quietly stepping off it, and the only ones still climbing aboard are the ones who cannot see the depreciation schedule, that is not a growth story. It is a tell. And the house, as always, has read it long before the table does.