Hook
Last week I read a sentence I have heard before, in every bull market, in every voice that has ever confused momentum with meaning. Mike Intrator, the chief executive of CoreWeave, told an interviewer that the market's treatment of his company was "irrational." The words arrived not as an argument but as a reflex — the sound a management team makes when the price of its stock stops agreeing with the story it has been telling. I have spent twenty-seven years watching infrastructure get built and, more often, watching it get financed into something it was never designed to survive. When a founder tells you the market is irrational, he is almost never describing the market. He is describing the questions he cannot answer. So I went looking for the numbers behind the sentiment. What I found was not a rebuttal of the doubt. It was a map of exactly where the doubt is right.
Context
CoreWeave is not, at its root, a technology company in the sense its marketing implies. It began in 2017 as Atlantic Crypto, an Ethereum mining operation that understood something the rest of the industry learned later: that the scarce resource was not cleverness but raw, coordinated compute. When Ethereum's merge ended the mining economics, CoreWeave did not disappear. It pivoted, converting its GPU fleets into a cloud business selling accelerated compute to the AI industry, and it did so with an intimacy with NVIDIA that few competitors could match. NVIDIA took an equity stake. GPUs were allocated preferentially. A miner became the reference cloud for the most capital-hungry technology cycle in modern history.
By the time of its 2024 public listing, CoreWeave had assembled a backlog of contracted future revenue that at one point exceeded $30 billion — the number its leadership now waves as proof of substance. Its largest customer, Microsoft, at times accounted for roughly 60 percent of revenue. Its contracts are predominantly "take-or-pay," long-dated commitments that look like stability on a spreadsheet. Its financing is overwhelmingly debt — billions raised against the collateral of the GPUs themselves, in structures led by firms like Blackstone. The listing itself was a stress test the company barely passed; the shares have since traced a violent arc, spiking and slumping in a pattern that reads less like a mature stock and more like a symptom. This is the company whose market treatment is, we are told, irrational. Before we accept that framing, consider what the market is actually pricing. And there is a quieter mismatch worth naming: this story reached me through a crypto news outlet, because CoreWeave's DNA is still mining. The audience that reads it is fluent in leverage, volatility, and tokenization — the very habits of mind that can make a rational doubt look, from the inside, like irrational panic.
Core
I once spent six hundred hours auditing Aave V2's early interest-rate logic by hand, and the lesson I carried out of it was not about Solidity. It was that a system's true risk lives in the gap between what its code assumes and what its contracts actually promise. CoreWeave is the same kind of object wearing a different skin. Read it as a codebase, and three assumptions carry almost all of its weight.
The first assumption is that leverage is growth. CoreWeave's expansion has been financed not by retained earnings but by debt secured against hardware — a model that works beautifully as long as three curves move together: GPU demand rises, GPU prices hold, and refinancing stays cheap. Debt against a depreciating asset is a bet that the asset will outlive the loan. Nobody says this out loud, because saying it out loud is the whole risk.
The second assumption is that a long contract is a real contract. The $30 billion backlog is presented as revenue. It is not. It is a stack of promises, made by customers who are themselves exposed to the same AI capital-expenditure cycle. When a hyperscaler trims its budget — and every hyperscaler is now building its own silicon, Trainium and TPU and MTIA, precisely to reduce the dependency CoreWeave sells — the "take-or-pay" clause becomes a negotiation, not a guarantee. Backlog is a forecast dressed as a fact. The distance between the two is where value goes to die.
The third assumption is depreciation. This is the quiet one, the one that took a famous short-seller to name in late 2025. If you assume an AI accelerator depreciates over six years, your profit looks robust. If the true useful life is three — because Rubin-generation silicon resets the market and the residual value of yesterday's card collapses — then your profit was a story, and your collateral was a mirage. I have watched this movie before, in the DeFi summer of 2020, when a single unexamined assumption in an interest-rate model could have vaporized four million dollars. The mechanism differs. The lesson does not.
And then there is the circularity, which no honest reader can ignore. NVIDIA invests in CoreWeave. CoreWeave buys NVIDIA GPUs. CoreWeave rents that compute to Microsoft and others who hold their own capital ties to the same ecosystem. Capital becomes procurement becomes revenue becomes valuation. When the same handful of actors are simultaneously the investors, the suppliers, and the customers, the revenue line stops measuring demand and starts measuring a closed loop. This is not fraud. It is something subtler and more dangerous: a demand signal that cannot be falsified from the inside.
Against all this stands a genuine scarcity. AI compute is real, and someone must build the racks, sign the power contracts, and wire the InfiniBand. CoreWeave does this competently. Its engineering — Kubernetes-native scheduling, bare-metal performance, tight adaptation to NVIDIA's reference architecture — is real, and I do not dismiss it. But competence in deployment is not a moat. It is a head start, and head starts get erased. The hyperscalers will commoditize bare-metal AI clusters. The other neoclouds — Lambda, Crusoe, Nebius, Together — will compete on price and flexibility. CoreWeave sits in the middle, squeezed from above by scale and below by hunger, its differentiation borrowed from a relationship with a single supplier that could be rebalanced at any time.
One more thread deserves the light. The bottleneck is shifting from GPUs to electricity and land. Whoever locks long-term power purchase agreements owns the next decade's cost curve. This is where CoreWeave's real hidden asset — or hidden liability — lives, and where the market's skepticism about its capital structure meets a physical reality no marketing can reframe.
None of this is a prediction of collapse. It is a description of exposure — and exposure, unlike fraud, does not require anyone to behave badly. It only requires that the assumptions fail at the same time, which is precisely how every leveraged infrastructure cycle has ever ended.
I have spent the last year building zero-knowledge verification tooling to keep AI-generated spam off decentralized platforms, and the exercise taught me something uncomfortable: verification is expensive, and the temptation is always to replace it with trust. CoreWeave's entire narrative is an invitation to trust — trust the backlog, trust the depreciation schedule, trust the supplier relationship. But trust is what you extend when you cannot verify. In infrastructure, that is not a compliment. It is a confession. Here is what I keep returning to, as someone who translated the Ethereum whitepaper into Portuguese in 2017 and added eighty pages of ethical commentary because I believed the philosophy mattered as much as the protocol. Code is law, but ethics is soul. The same is true of a balance sheet. A debt structure is law — enforceable, precise, indifferent. What it cannot encode is whether the assumptions underneath it will hold when the weather changes. CoreWeave's "irrational market" is the weather changing. And the CEO's response is to insist the sky is clear.
Contrarian
The contrarian reading — the one I would defend — is this: the market is not irrational. It is doing precisely what a market should do with a company whose growth depends on assumptions it has declined to publish. A rational market does not ignore leverage, concentration, and depreciation. It prices them, imperfectly and noisily, but it prices them. What is genuinely irrational is the reverse claim — that a hundred-billion-scale bet on debt-financed hardware, dependent on one supplier and one customer, should be valued as though none of that were true. The word "irrational" is doing the work of a rhetorical shield. Notice what the defense never addresses. It does not publish the net-debt-to-EBITDA ratio. It does not disclose the loan-to-value terms on the GPU-collateralized debt, or the margin-call triggers that could cascade if accelerator prices fall. It does not break out how much of the backlog is executable and how much is negotiable. A company confident in its fundamentals answers these questions with filings, not with adjectives.
I learned during the bear market of 2022 — mentoring ten junior developers in a private Discord while Terra and FTX burned — that evangelism is not shouting during bull markets. It is whispering truth during bear markets. The same discipline applies to reading a press appearance. Code is law, but ethics is soul. And transparency isn't the oxygen of trust — it is the trust. Without it, every assurance is a draft, and every draft is a liability.
Takeaway
So I will not call the market irrational. I will call it early. The questions it is asking — who guarantees the debt, what the collateral is truly worth, how concentrated the revenue has become — will be answered not by the next interview but by the next set of filings, the next capex guidance, the next time a hyperscaler trims a budget. When a system's survival rests on assumptions no one will publish, the honest position is not confidence. It is verification. And verification is the only thing CoreWeave — or any of us — has ever been able to sell.