Mike Intrator said the quiet part loud, and then said it again for the cameras: the market is being irrational about CoreWeave.
I've heard that sentence before. I heard it from Do Kwon in April 2022, three weeks before Terra's death spiral ate $40 billion. I heard it from the Fomo3D whales in late 2017, right before the last wallet went dormant and the pot froze into a gas-price graveyard. I heard it from a Celsius executive over a steak in Toronto, three months before the withdrawal freeze.
The code didn't break in any of those cases. The code ran exactly as written. What broke was the assumption underneath it โ that the inflow would never stop, that the collateral would hold its value, that the counterparty would pay.
So when the CEO of a company carrying tens of billions in secured debt, sitting on 60%+ customer concentration, and underwriting a GPU depreciation schedule that would make a used-car dealer blush, tells you the market is irrational โ you don't argue. You don't tweet. You open the filings. You read the collateral terms. You count the days to the next refinancing window.
That's what this is. Let me show you the shape.
Context: The Crypto Company Wearing an AI Costume
CoreWeave is not an AI startup that appeared out of nowhere in 2023. Most people covering AI infrastructure don't know this, and it's the single most important thing about the company.
It started as Atlantic Crypto โ an Ethereum miner, 2017 vintage. When ETH mining margins compressed and the Merge loomed, the team did the most crypto-native thing imaginable: they took the one asset that survived the transition, GPU compute, and re-collateralized it into a new narrative. Mining rigs became "AI clusters." The pivot was not technological. It was financial. It was the same move a degen makes when the yield farm dries up โ rotate the principal into the next pool before the APY prints zero.
That origin is not trivia. It's the risk model. The people running CoreWeave learned their leverage discipline in an industry that invented 100x perpetuals, recursive collateral, and the prayer "the liquidity will come back." They are not naive operators who stumbled into leverage. They are the most sophisticated borrowers in the entire AI capital stack. That's precisely the problem.
Now, the category. CoreWeave is a "neocloud." Not a hyperscaler. Not SaaS. The neocloud model is: buy GPUs at scale, wire them with InfiniBand, rent them by the hour to customers who need capacity faster than AWS can provision it. It's a wholesale compute play. Structurally it behaves less like software and more like a data-center REIT โ except the building depreciates. Fast.
The backdrop is the AI capex supercycle. Every hyperscaler and every frontier lab is spending at a rate with no precedent in corporate history. Microsoft, Amazon, Google, Meta โ combined capex guidance running into the hundreds of billions of dollars. That spend is the tide CoreWeave floats on. It is also the tide that can recede.
CoreWeave is the purest, most leveraged, most concentrated way to express a bet on that tide. That's the entire investment thesis, and it's the entire risk, and the two are the same sentence. A pure beta play doesn't diversify a cycle. It amplifies it. When the tide comes in, CoreWeave rises faster than the incumbents because it has no offsetting businesses โ no advertising arm, no enterprise software cushion, no search monopoly. When the tide goes out, the same absence of cushion works in reverse, and the leverage turns the amplifier into a megaphone pointed straight at the downside.
The pitch is real: AI labs and hyperscalers need compute now, and the incumbents can't move fast enough. CoreWeave filled the gap. Revenue exploded. Backlog โ signed contracts for future delivery โ crossed $30 billion at points in 2025. On paper, this is the greatest growth story in infrastructure.
On paper.
Core: Reading the Structure Like a Contract
Here's where I stop being a reporter and start being someone who has audited smart contracts for a living. Because the CoreWeave structure has a shape I recognize. I've seen it in DeFi treasuries, in token vesting cliffs, in the Fomo3D pot right before it locked.
The shape is this: growth funded by debt, collateralized by an asset whose value is set by the party selling it to you, with revenue concentrated in counterparties who are also investors. Each clause is load-bearing. Let me sit on each one.
The debt is not venture money, and that changes everything. CoreWeave has raised tens of billions through secured facilities โ Blackstone-led loans, the DDTL lines, a stack of notes. Venture money doesn't get margin-called. Debt does. And this debt is secured against the GPUs themselves. That is the crypto move. It's exactly what an on-chain lending protocol does when it accepts an LP token as collateral: you borrow against an asset, then re-hypothecate, then borrow again. The elegance is real. The fragility is realer, and it lives in one number โ the loan-to-value ratio on hardware that depreciates faster than any collateral in financial history.
I want to underline this, because it's the thing the AI crowd keeps missing. GPU-collateralized lending is the crypto leverage playbook in a data-center uniform. The mining farms did it with ASICs. The NFT funds did it with JPEGs. DeFi did it with yield-bearing tokens. Every one of those collateral types had the same nasty habit: when the market turned, the collateral and the borrower's revenue fell together, and the margin call landed exactly when nobody could answer it. Celsius. BlockFi. Three Arrows. The collateral was crypto; the optimism was identical; the margin call was identical; the hole was identical.
The circularity is the part mainstream coverage waves away. Follow the money in a circle. NVIDIA takes equity in neoclouds, CoreWeave included. CoreWeave uses that capital โ plus debt โ to buy NVIDIA GPUs. CoreWeave rents those GPUs to large customers, including OpenAI and Microsoft. Those customers have their own capital relationships with NVIDIA and with each other. Microsoft is both a customer of CoreWeave and, historically, entangled in the other direction. OpenAI is a customer and an investee inside a web of entities that all transact with the same handful of counterparties.
Read that slowly. Capital flows out from NVIDIA, returns as GPU orders, and comes back as rental revenue from customers who are themselves entangled with NVIDIA. That is not automatically fraud. Circularity in a supply chain is normal. But circularity this tight, at this scale, at this leverage, produces a demand signal that is partly a mirror. When the mirror cracks, everyone in the circle sees it on the same day.
I lived through this in 2017. The Fomo3D pot grew because each new entrant's buy fed the payout that attracted the next entrant. The mechanism was beautiful. It was also, structurally, a loop that could only end one way. I broke that story four hours before the majors because I watched the gas โ the gas told me the inflow was slowing before the price did. Nobody wanted to hear it. The pot was still going up.
The concentration is a dependency, not a business. In 2024, Microsoft accounted for roughly 60% of CoreWeave's revenue. One customer. Over 60%. I don't care what industry you're in โ that is not a business, that is a hostage situation. If Microsoft sneezes on its capex guidance, CoreWeave catches pneumonia.
The CEO's defense is the backlog. Thirty billion dollars, take-or-pay, contracted, locked. Sounds bulletproof. But I've covered contracts in crypto long enough to know the rule: a take-or-pay contract is only as strong as the counterparty's willingness and ability to pay, and both are pro-cyclical. When your customer's revenue is under pressure, "take-or-pay" becomes "let's renegotiate." I've watched token projects with "locked" vesting cliffs unlock early after a single phone call. Contracts are paper. Cash is real. And CoreWeave's paper is concentrated in a few names whose spending is itself a function of the AI capex cycle CoreWeave depends on.
Here's the accounting sleight-of-hand nobody explains. Backlog is a signed promise, not a deposit. Revenue recognition stretches across years, and the contract value can be renegotiated, delayed, or defaulted long before it ever touches the income statement. A $30 billion backlog is not $30 billion of cash. It's $30 billion of conditional intention, priced as if it were guaranteed. The loop closes on itself: CoreWeave's revenue depends on AI capex. AI capex depends on the belief that AI revenue will materialize. AI revenue depends on models getting good enough to monetize. And every layer of that stack is priced as if the answer is already yes.
The depreciation question is where it hits the income statement. This is the one the short-sellers found, and it's the one that matters most, because it touches profit directly rather than indirectly. CoreWeave depreciates its GPUs over a useful-life assumption somewhere around six years. If that's right, the profit numbers are roughly what they claim. If it's wrong โ if a GPU's economically useful life, in a market where NVIDIA ships a new architecture every 18 months, is really three or four years โ then CoreWeave has been systematically understating cost and overstating profit. The gap doesn't hit cash flow immediately. It hits later, as a write-down, as a refinancing that can't clear, as collateral that no longer covers the loan against it.
The code didn't lie about the depreciation schedule. The people who set it did. Or, more charitably, they set it the way every leveraged operator sets assumptions โ toward the outcome that keeps the machine running.
The IPO itself was the first confession. Think about what an IPO actually is: a moment when a private balance sheet is forced into daylight. CoreWeave's post-listing price action was wild โ sharp run-ups, brutal retraces, the kind of volatility that tells you the market is genuinely split on the company's worth. Management reads that split and calls it irrationality. I read it differently. Extreme two-way volatility is not proof the market is wrong. It's proof the market is undecided, and undecided markets are usually undecided because the disclosure is incomplete. If the numbers were clean and the structure were simple, the price wouldn't whip like that. The whipsaw is the information.
The customer is quietly building the replacement for the supplier. I've sat in the rooms. King West, Toronto, the kind of dinner where you learn more in two hours than in two quarters of filings. The neocloud category is crowded and getting more crowded. Lambda. Nebius. Crusoe. Together. And above all of them, the hyperscalers โ AWS, Azure, Google Cloud, Oracle โ building dedicated AI clusters and, critically, their own silicon.
Read the second-order effect. The hyperscalers are CoreWeave's customers and its competitors. Microsoft buys capacity from CoreWeave because it can't provision fast enough โ and simultaneously builds its own capacity and designs its own chips to reduce exactly that dependency. Every dollar Microsoft spends on internal silicon is a dollar that eventually doesn't go to CoreWeave. The customer is methodically constructing the replacement for the supplier. That is not a durable moat. That's a countdown.
Stack the comparison honestly. CoreWeave wins on pure AI focus โ it's the most single-minded compute play in the market. It loses on scale, on capital, on ecosystem, on software integration, on customer diversification, and on balance-sheet leverage. The hyperscalers have low concentration risk and deep distribution; CoreWeave has the opposite of both. The smaller neoclouds โ Nebius, Lambda, Crusoe โ have pricing flexibility CoreWeave lacks because they carry less debt. So CoreWeave is squeezed: too leveraged to win a price war, too small to out-scale the incumbents, too concentrated to survive a single customer's capex cut.
CoreWeave's edge is the NVIDIA relationship, bare-metal performance, and speed. Two of those three are copyable within a couple of years. The third โ the NVIDIA relationship โ is not a technology. It's a favor. And favors get reallocated when the patron's incentives shift. I've seen this in crypto too: the project that got the exchange listing because of a relationship, the protocol that got the market-maker because of a phone call. The moment the relationship stops paying for the counterparty, it evaporates. There's no code enforcing it.
The new bottleneck is power, not silicon. Here's a thread most AI-infra analysts are still not pulling. GPUs are becoming less scarce; electricity and land are becoming more scarce. The large neoclouds are now in a land-and-power race โ long-term power purchase agreements, substation interconnects, water rights for cooling. CoreWeave's ability to lock long-term, low-cost power is simultaneously its hidden moat and its hidden liability. A twenty-year PPA is a beautiful asset if demand holds and an anchor if it doesn't. It's the same structure as a token with a multi-year lock: wonderful in the bull case, a noose in the bear case.
And the energy angle drags in an ESG dimension. The more these clusters scale, the harder the carbon and grid-strain debate bites, and in ESG-sensitive capital pools that debate translates directly into financing friction. The same leverage that makes CoreWeave flexible in a boom makes it fragile in a squeeze, because the marginal lender starts asking questions it never asked before.
The security and isolation layer is the quiet risk. CoreWeave doesn't train or publish models, so it dodges the direct obligations that the EU AI Act places on model providers โ its regulatory exposure is genuinely lower than a frontier lab's. But it's a multi-tenant compute platform holding some of the most sensitive training data on earth. One cross-tenant leak, one supply-chain compromise inside the NVIDIA stack, and the trust collapse is total. That's not an "AI safety" risk. It's an infrastructure-security risk, and it's underpriced because it's boring. Nobody writes threads about data isolation until the day they have to.
And there's a crypto-native tail nobody is pricing. CoreWeave's debt could, in principle, be tokenized โ packaged and sold as yield-bearing instruments in the RWA wave. If that happens, you bolt crypto's leverage cycle onto AI's capex cycle, two pro-cyclical engines firing at once. That's not diversification. That's resonance. And resonance is how small cracks become category-wide contagion.
The neocloud debt model is contagious, and that's the real systemic story. If one neocloud borrows against GPUs, that's idiosyncratic risk. If a dozen do it, at the same time, against the same collateral class, with the same optimistic depreciation assumptions, you've built an AI subprime โ a stack of loans whose underlying asset can lose half its value in a single NVIDIA product cycle. When the first borrower misses a margin call, the forced selling of GPUs hits the residual values that all the other loans depend on. That's the contagion vector. That's the thing the sector's cheerleaders refuse to model.
Contrarian: The Market Is Right, and the CEO Knows It
Now the unreported angle โ the one the AI-infrastructure crowd keeps missing because they don't read crypto history.
Everyone is analyzing CoreWeave as an AI company. It isn't. It's a crypto company that survived a bear market by rotating into AI, and it is running the crypto playbook on an AI balance sheet. That's why the source of this story โ a crypto outlet โ is itself a data point. Crypto readers are the only ones who recognize the shape, because we watched it happen three times already.
Here's the contrarian read: the "market irrationality" the CEO is complaining about is the market being more rational than he is. The market has priced the debt. It has priced the concentration. It has looked at the circularity, the depreciation, the NVIDIA dependency, and said: this deserves a discount, not a premium. That is not irrationality. That is the efficient-market hypothesis doing its job โ processing the risk that management wants you to ignore.
The CEO wants you to believe the disagreement is about growth. It isn't. Nobody serious doubts that AI compute demand is real and rising. The doubt is about financing structure โ whether the way CoreWeave captures that demand can survive a slowdown in the demand curve or a write-down in the collateral. Those are different questions, and conflating them is the oldest trick in the leveraged-operator playbook. Kwon did it. SBF did it. Every founder whose balance sheet is load-bearing does it: reframe criticism of your structure as criticism of your vision, then call the critics unsophisticated.
I don't buy it. And here's the tell: if the fundamentals were as bulletproof as the CEO implies, you wouldn't need to call the market irrational. You'd just post the numbers. You'd publish the LTV. You'd disclose the customer-concentration trend. You'd defend the depreciation methodology line by line, publicly, with a straight face. Silence on those points, paired with loudness on the narrative, is the signature of a management team defending a price, not a business.
We didn't get the financials. We didn't get the covenants. We didn't get the customer-by-customer backlog breakdown. We didn't get a single net-debt-to-EBITDA figure. We got a vibe. And vibes, as the 2017 pot taught me, are what you feel right before the gas spikes and the last wallet goes quiet.
The code didn't care that the room was full of believers. It cared about the inflow. And so does the debt.
Takeaway
Watch the refinancing calendar, not the earnings call. The next time CoreWeave has to roll its debt, GPU residual values and LTV terms will decide more about its future than any model release. Watch NVIDIA's allocation decisions โ the moment the patron tilts its favor, the moat is gone. And watch the depreciation debate, because when the first write-down lands it won't be a CoreWeave story. It'll be the whole neocloud category, and then the whole AI capex trade, discovering at once that the collateral was never as solid as the loan against it assumed.
The market isn't irrational. It's just early. And it's seen this movie before โ we all have.