The Collateral Is a Chip: CoreWeave and the First Honest Repricing of the Compute Trade
Hook
The headline arrived the way most headlines arrive now β stripped of everything except the verb. CoreWeave stock slides. AI data center selloff deepens. No number. No date. No source. No cause. Just a downward motion and a comparative adverb, packaged as news and shipped to a crypto publication that, on any other morning, would be writing about a token that verifies data for a chain nobody uses.
We didn't get the percentage. We didn't get the volume. We didn't get the trigger. Was it a macro print, an earnings miss, a short report, a whisper in a group chat that metastasized into a forty-billion-dollar idea about what happens when you mortgage the future to a chip that depreciates every eighteen months?
What we got was a signal. And signals, in a bear market, are worth more than data, because data tells you what happened and signals tell you what everyone is about to believe. In the ledger's silence, the true story whispers. So let me be honest about what this piece is: not a report on CoreWeave, but a forensic reading of a mood β the mood of a market that spent three years telling itself compute was the new oil, and has now, quietly, begun to ask whether oil ever came with a margin call.
The Company That Used to Mine Ether
CoreWeave is not a mystery. It is a publicly traded entity β ticker CRWV β that came to the market in March 2025, in the middle of the most aggressive capital deployment cycle in the history of information technology. It is what the industry has agreed to call a "neocloud": a GPU-first infrastructure provider that rents bare-metal compute at a scale and speed the traditional hyperscalers, for all their engineering depth, cannot match on a per-chip basis.
Its lineage is the part crypto readers should linger on. CoreWeave did not begin as an AI company. It began as an Ethereum miner β Atlantic Crypto β in the years when the value of a GPU was set not by how well it could run a transformer but by how efficiently it could hash. When the Merge ended proof-of-work mining on Ethereum in 2022, a fleet of industrial mining hardware became, overnight, a stranded asset. The smartest operators did not sell the metal. They repurposed it. CoreWeave is the most successful artifact of that pivot: a mining company that discovered, at precisely the right moment, that the same racks, the same power contracts, the same cooling, and the same obsession with GPU utilization could be re-pointed at a market with ten thousand times the demand.
I was in Dubai when that pivot happened, and I did not see it coming. In 2018 I spent forty hours reverse-engineering a yield protocol called Raptor, convinced its interest-rate arbitrage model was the next narrative, and published a three-thousand-word bullish thesis eight days before a reentrancy bug drained two million dollars from its contracts. The lesson I took was not "be more careful." It was that the market rewards the person who reads the sentiment correctly, not the person who reads the code correctly β because the code changes and the sentiment persists. Code is law, but humans write the bugs, and humans write the sentiment too.
CoreWeave is the sentiment of the last three years made physical: a warehouse full of chips that everyone agreed would always be scarce.
What a Headline Without Numbers Actually Says
Here is the discipline that a flash like this demands. The story contains no date, no magnitude, no attribution, and no cause. On its face, that makes it worthless. But worthless information and zero information are different things, and the difference is this: a flash tells you which direction the crowd is leaning, even when it refuses to tell you how far it has leaned. The verb is "deepens." Not "drops." Not "tumbles." Deepens β which is a word about process, not event. It implies that a decline was already underway, that this is the latest increment of an existing move, and that the market has crossed the line from a single bad day into a trend.
That word choice is doing enormous work, and the platform choice is doing more. This was published by a crypto outlet. Not a wire service, not a terminal, not a sell-side desk. A crypto outlet. Which means the people who will read it first, and trade on it first, are the people who have spent the last cycle financializing everything they touch β tokens, memes, identity, and now compute itself. Sentiment is a shifting tide, not a solid ground, and the tide here is being reported by the one audience most prone to panic and most capable of amplifying it.
So let me treat the flash not as an event to analyze but as a signal to interpret. And the interpretation is uncomfortable: the market is beginning to price something that has been unpriceable for three years β the possibility that compute is not infinitely scarce, and that the balance sheets built on the assumption of its infinite scarcity are not as safe as their owners believed.
The Leverage Structure: When the Collateral Is a Chip
To understand why a single headline can move a company like CoreWeave, you have to understand what CoreWeave actually is, and it is not, despite the branding, a technology company in the way that term is usually meant. It does not train frontier models. It does not own a developer ecosystem. It does not sell software that anyone would miss. It buys chips, racks them, wires them with high-speed interconnect, and rents them out under long-term contracts β and it funds the buying with debt secured against the chips themselves.
That last clause is the whole story. CoreWeave's growth engine is not margin. It is leverage. The company borrows against its GPU fleet β with large institutional lenders, Blackstone among the most prominent, structuring facilities in the tens of billions β and uses the proceeds to buy more GPUs, which it then rents out, which produces the revenue that services the debt that bought the chips that generate the collateral that secures more debt. It is a flywheel, and like all flywheels, it is beautiful while it spins and catastrophic when it stops.
Based on my audit experience β and I use that phrase deliberately, because what follows is pattern recognition, not accusation β the structure resembles something the crypto market knows intimately. In 2020, during DeFi Summer, I wrote three blogs at once and coined a phrase I still stand behind: liquidity mining as social contract. The insight was that yield farming was never really about yield; it was about governance, about the illusion of participation, about turning depositors into stakeholders by paying them in a token whose value depended on the deposits that the token attracted. It worked until it didn't. Yield is the bait, liquidity is the trap. CoreWeave's flywheel is a corporate-grade version of the same mechanism, with one crucial substitution: the token has been replaced by a chip, and the chip has a resale market that behaves exactly like a token's price chart β volatile, sentiment-driven, and prone to gaps.
The debt is the part that keeps me up. Interest is a fixed claim on a variable asset. If you borrow against an H100 and the H100's rental value collapses because a newer chip makes it obsolete, your revenue falls but your coupon does not. That is the definition of a duration mismatch, and it is the reason that the phrase "AI infrastructure" and the phrase "credit risk" are, in 2026, the same sentence wearing two different hats.
The Customer Concentration Problem
Now stack the leverage on top of the concentration. Public information through 2025 indicated that a single customer β Microsoft β at one point accounted for roughly sixty percent of CoreWeave's revenue. In the quarters that followed, additional large contracts were signed, with OpenAI reportedly committing to deals in the ten-billion-dollar range and Meta entering the roster, and the company's backlog climbed into the tens of billions, with some estimates pushing toward the higher hundreds of billions on the strength of those mega-contracts. I flag that range deliberately, because the point is not the number. The point is the shape.
The shape is a company whose revenue is concentrated in a handful of counterparties who are, simultaneously, its most credible competitors. Microsoft buys compute from CoreWeave because Microsoft's own capacity is short. The moment Microsoft's capacity is not short β and Microsoft is spending at a pace that guarantees that moment arrives β the reason to buy from CoreWeave evaporates. This is not a partnership. It is a stopgap. And stopgaps have a way of becoming unnecessary exactly when you have finished borrowing against them.
I watched a version of this in 2022, when I spent three months interviewing fifteen former executives from Celsius and BlockFi for a series on the moral hazard of centralized exchanges. The pattern in every interview was identical: a business model that worked only under one assumption β that the asset would keep appreciating β and a management team that had stopped asking whether the assumption could fail. Nobody in those rooms was stupid. They were simply inside the flywheel, where the walls spin so fast you mistake motion for stability. CoreWeave's customers are not going to be interviewed about it later. But the structure rhymes.
The Depreciation Trap
Here is the technical heart of the matter, and it is where the crypto reader's instincts are most useful. In a normal asset-backed lending market, the collateral depreciates slowly. A building, a fleet of trucks, a portfolio of receivables β these decay on schedules you can underwrite. A GPU does not. NVIDIA ships a new architecture roughly annually β Hopper to Blackwell to Rubin β and each generation does not merely improve on the last; it re-prices it. When Blackwell arrives in volume, the rental value of a Hopper-class chip does not gently decline. It steps down, because the marginal buyer of compute will pay for the newer silicon and the older silicon must discount to clear.
This means CoreWeave's collateral is depreciating on a schedule that is set by a company it does not control and cannot influence, at a pace that is roughly twice as fast as the depreciation assumptions that appear in most data-center lending models. If those models assume a five-year useful life and the market reality is closer to three, then the equity in the structure is thinner than it appears, and the debt is riskier than it is rated. This is the GPU depreciation trap, and it is not a prediction of doom. It is simply an observation that the collateral at the bottom of this pyramid is the fastest-decaying asset ever used to secure institutional credit at scale.
There is a reason the crypto market should recognize this instantly. We spent a decade lending against assets whose value was determined by a price we could not control β and we called the resulting blowups "black swans" because it was more comfortable than calling them "arithmetic."
The Sociology of Scarcity
Before the arithmetic, the anthropology. Because the reason this selloff matters is not that a stock fell. It is that an entire culture is discovering that its founding myth was a marketing document.
For three years, scarcity was not a market condition. It was an identity. To own compute was to belong to the future. To hold a GPU allocation was to hold a passport to the rooms where the important decisions were being made. This is the same mechanism I spent twenty interviews mapping in 2021, when I stopped covering Bored Ape floor prices and started asking collectors why they bought. The answer was almost never the art. It was the signal. The 10,000 ETH of volume was not a market discovering value. It was a status system discovering a scoreboard. Art without utility is just noise with a price tag, and I said so at the time, and I was called a heretic for it.
The compute market runs on the same fuel. The chip is the ape. The allocation is the rare trait. The data center is the yacht club. When you look at CoreWeave's investor base and its customer list, you are not looking at a cost-benefit analysis. You are looking at a status hierarchy that happens to clear through a stock price. And status hierarchies are the most sentiment-sensitive instruments ever constructed, because their value depends entirely on what the other members of the hierarchy believe β not on any cash flow that exists independent of that belief.
This is why the headline's verb matters so much. "Deepens" is a word about a hierarchy losing confidence in itself. It is the sound of a room deciding, collectively and without a vote, that the thing it was proud to own last quarter is now the thing it is embarrassed to be seen holding. That is not a financial event. That is a cultural one, and cultural events in a concentrated ownership structure are always more violent than financial ones, because the exit is narrow and everyone is running for the same door.
The Infrastructure Triangle: Compute, Power, Capital
CoreWeave's real constraint is not software. It is a triangle with three vertices: compute, power, and capital. The company can buy chips if it has power to run them and capital to buy them with. It can get capital if its collateral holds value. It can get power if it can sign long-term contracts with utilities and data-center developers at prices that keep its unit economics viable. Every vertex depends on the other two, and the whole structure depends on the price of electricity and the cost of debt moving in the right direction.
For three years, they moved in the right direction. Power was available, capital was cheap, and chips were the only thing in short supply. In 2026, the triangle is rotating. Power is becoming the binding constraint β the AI data center buildout is now more a story about substations and cooling than about silicon. Capital is no longer free; the rate environment has made the marginal dollar of debt more expensive, and the credit market has begun to distinguish between a hyperscaler's balance sheet and a neocloud's. And the chips β the chips keep coming, faster than anyone projected, which is precisely the problem for anyone who owns the previous generation.
I do not claim to know the electricity price at any specific CoreWeave site. But I know what a binding constraint looks like, because I have watched crypto's own infrastructure go through it. The Layer 2 scaling story promised decentralization for two years and delivered sequencers that are, in practice, single nodes operated by single companies. The pitch was always "trustless." The reality was always "trust us." Compute infrastructure has the same gap between the pitch and the plumbing. The pitch is elasticity and abundance. The plumbing is a handful of power contracts, a handful of lenders, and a handful of customers.
The Competitive Squeeze
Where does CoreWeave sit in the competitive map? In the middle, and the middle is a hard place to stand.
Above it are the hyperscalers β AWS, Azure, Google Cloud β which are worse at raw GPU cost and delivery speed but vastly better at everything else: storage, networking, databases, AI tooling, and the gravitational pull of a developer ecosystem that makes switching expensive. When a customer chooses CoreWeave over a hyperscaler, the customer is not replacing the hyperscaler. The customer is supplementing it, renting overflow capacity while it waits for its own buildout or its preferred provider's allocation. Overflow demand is real demand, but it is the most elastic demand in the market. It is the first demand to disappear when the buyer's own capacity arrives.
Beside it are the other neoclouds β Lambda, Crusoe, Nebius, and a lengthening list of entrants. Against these, CoreWeave's advantage is not technology. It is scale, and it is access. Access to NVIDIA allocation, secured through a relationship in which NVIDIA is both supplier and shareholder, is the moat. But a moat that depends on a single vendor's favor is not a moat. It is a lease on a moat, renewable at the vendor's discretion, and the vendor sells to your customers and your competitors with equal enthusiasm. NVIDIA is not a partner. NVIDIA is a market. And markets do not have loyalty; they have allocation.
Below it β or rather, beside it in a different geography β are the sovereign and national compute efforts, the programs that states are funding to guarantee their own access to silicon. Those programs are not competing on price. They are competing on control, which is a different game entirely. They are also the clearest illustration of the deeper split that the crypto market has been arguing about for years: the difference between systems built for surveillance and systems built for privacy. A state that wants to see every inference is building a different machine than a market that wants to price every inference. They cannot both be right about what compute is for.
Valuation as Narrative
Now the part that no flash headline will ever give you: what the market is actually paying for.
CoreWeave's valuation, at its peak, implied a set of assumptions that were heroic even by the standards of a heroic market. A price-to-sales multiple that dwarfed the data-center and cloud comparables. A growth rate that required the backlog to convert on schedule and the schedule to be renewed at similar terms. A margin structure that assumed the company could keep pricing power in a market where it is, structurally, a wholesaler. None of those assumptions is impossible. All of them are simultaneously required. That is what it means to buy a narrative: you are not buying a probability distribution. You are buying the conjunction of every good thing that has to happen for the story to be true.
I learned this the hard way in 2018. Raptor's thesis was not wrong because its math was wrong. It was wrong because the math was a conjunction of assumptions that a reentrancy bug turned into a contradiction. The protocol worked exactly as designed, right up until it didn't, and the design had assumed that no one would write the bug. Every bull run is a myth waiting to be debunked, and the debunking almost never comes from the place the bulls are watching. It comes from the assumption they forgot they were making.
For CoreWeave, the forgotten assumption is depreciation. Every valuation model on the sell side contains a line item for the useful life of a GPU, and almost every one of those line items is a number chosen for convenience rather than observed in a market. There is no public tape for used H100s. There is no reference rate for compute-hours. The single most important input to a multi-billion-dollar credit structure is estimated by consensus, which is another way of saying it is estimated by nobody. When that estimate is revised β and it will be revised, because NVIDIA's roadmap is not a secret β the entire valuation compresses at once, and the compression is not gradual. It is a step function, like the chip itself.
The Contrarian Angle: What If the Selloff Is Not About CoreWeave At All?
Here is where I stop reading the tape and start reading the room.
The consensus interpretation of a headline like this is straightforward: a company got too expensive, a bubble is deflating, and the smart money is heading for the exits. Every bull run is a myth waiting to be debunked, and the debunking is always presented as wisdom arriving late to a party it should have skipped. That is the comfortable story. It is also, I think, the wrong one.
Consider what is actually being repriced. The selloff is described as an "AI data center selloff" β not a CoreWeave selloff. That framing matters. It suggests the pressure is on the sector, not the company, which means the market is not making a judgment about CoreWeave's execution. It is making a judgment about the sustainability of the entire compute buildout. And that judgment, at its core, is a judgment about whether the demand for compute is as infinite as the narrative assumed.
Now consider the counterintuitive possibility. What if the demand is real, but the financing is the fiction? What if the chips will all be used β every last one β and the companies that bought them on leverage will still fail, because the cost of the leverage exceeds the return on the chips? That is not a bubble story. That is a capital-structure story. It is the difference between a market that was wrong about the future and a market that was right about the future but wrong about how to pay for it.
I have seen this exact confusion before, and it is the reason I stopped writing hype in 2018. After Raptor, I realized that most "bubble" analysis is really just price analysis wearing a costume. The question is never "is this asset overvalued." The question is "who is holding the bag when the music stops, and did they know it." For CoreWeave, the bag is not held by retail day-traders. It is held by the lenders who accepted a fast-decaying chip as collateral, by the shareholders who underwrote growth on the assumption of perpetual scarcity, and by the customers whose contracts are only as good as their own balance sheets. That is a concentrated, institutional bag. And concentrated institutional bags do not pop. They unwind, slowly, in private, in amended credit agreements and renegotiated take-or-pay terms that nobody announces.
The other contrarian point is about the signal's source. This was published by a crypto outlet, and that is not incidental. The crypto market has spent two years trying to financialize compute β GPU-backed tokens, decentralized compute networks, "depin" projects that promise to aggregate idle silicon. CoreWeave is the TradFi version of that thesis. If crypto is writing about CoreWeave's slide, it is partly because crypto's own compute narratives are watching CoreWeave as a proxy. When the proxy falls, the derivative narratives fall harder, because they have no cash flow to fall back on. The tide that is going out here is not just the AI capex tide. It is the tide of the entire "everything can be financialized" religion, and CoreWeave happens to be standing on the wet sand.
The Crypto Connective Tissue
I keep being asked why a crypto editor writes about a GPU cloud company. The answer is that the two are the same trade wearing different tickers, and the connective tissue is thicker than most people admit.
Start with the oracles. DeFi's most persistent technical weakness is oracle feed latency β the gap between when a price moves in the world and when the chain knows it. That gap is where liquidations happen, where the "decentralized" price is revealed to be a handful of node operators reporting a number they read somewhere else. Chainlink solved the decentralization problem by concentrating it: a set of nodes that are decentralized in name and coordinated in practice. Now look at compute. CoreWeave's entire valuation depends on a price β the price of GPU rental β that is set in a market with no public oracle at all. There is no Chainlink for H100-hours. There is no reference rate for compute. The "price" of compute is whatever the last large contract implied, which means the collateral backing tens of billions in debt is valued by an oracle that does not exist. In the ledger's silence, the true story whispers β and here the silence is not a metaphor. There is literally no tape.
Now look at the sequencers. Layer 2's dirty secret is that "decentralized sequencing" has been a PowerPoint for two years while the actual sequencers remain single nodes run by single entities. The pitch is decentralization. The plumbing is centralization. CoreWeave's pitch is elasticity and independence from the hyperscalers. The plumbing is dependence on one chip vendor, one or two lead lenders, and three or four mega-customers. The pattern is identical across the two industries: the story is about removing intermediaries, and the reality is about replacing a visible intermediary with an invisible one.
And then the deepest connection of all, the one that runs through every crypto cycle: the financialization of a real asset into a speculative one. Compute is real. Electricity is real. The demand for inference is real. But when you mortgage real assets to buy more real assets and securitize the whole stack, you have built a structure whose behavior is governed by sentiment, not physics. Sentiment is a shifting tide, not a solid ground. The physics of a data center does not care whether the market is bullish. The structure built on top of it cares about nothing else.
The 2026 Question: An Economy With No Readers
I have one more layer to add, and it is the reason I am writing this in 2026 rather than 2024.
Last year I published a piece called The Silent Market, built on a study of ten thousand AI-agent interactions on-chain. The finding that stopped me was not the volume. It was the composition: roughly seventy percent of those transactions were micro-payments for data verification, executed by software, for software, with no human in the loop and no human-readable narrative attached. The agents did not need a story to transact. They needed a signature and a fee. And that changed how I think about every asset whose value is derived from human attention.
CoreWeave is a bet on the compute layer of that agent economy. But here is the contradiction the bulls have not resolved: the more the economy is driven by agents, the less it depends on the human narratives that currently set the price of the assets that build it. The chips will still be needed. The data centers will still run. But the sentiment premium β the part of the valuation that exists because humans find the story exciting β is precisely the part that an agent-driven market will not pay. Agents do not buy status. Agents do not chase scarcity myths. Agents do not care that the founder used to mine Ethereum.
This is the deepest bear case, and it has nothing to do with demand. It is that the compute trade is priced for human enthusiasm, and the compute economy is being built for something that has none. When the two meet, the valuation normalizes toward utility β and utility is a much smaller number than narrative. The people who bought CoreWeave because it was the most exciting way to own the future may discover, in the end, that they were early to a market that will never be excited about anything.
The Risks That Matter
Strip away the noise and three risks dominate, in order of severity.
The first is a financing-chain rupture. If the market's appetite for GPU-backed debt narrows β because rates rise, because a lender marks down collateral, because a credit committee reads the same depreciation math I just walked through β then CoreWeave's expansion engine stalls at exactly the moment its fixed costs are highest. This is not a probability I can assign with confidence; it is a mechanism I can describe. Watch the bond spreads. Watch the credit ratings. Watch the quarterly cash burn. Those are the vital signs.
The second is customer-concentration backlash. The mega-customers are building their own capacity at a pace that will, within a small number of years, remove the reason to outsource. If Microsoft, OpenAI, and Meta all bring their own clusters online and reduce their reliance on third-party neoclouds, CoreWeave's backlog becomes a schedule of contracts that will not be renewed. Watch the self-build capex guidance from the hyperscalers. That number is the real threat to the neocloud thesis, and it is published every quarter.
The third is a sector-wide repricing. If the AI-bubble narrative takes hold across the market, the entire compute complex reprices together β the operators, the data-center REITs, the power names, the equipment makers. In that scenario, the question of whether CoreWeave is a good company becomes irrelevant, because the question the market is asking is whether the sector is a good sector. And sectors do not trade on merit. They trade on mood.
The Opportunities Nobody Is Pricing
I am a contrarian by trade, which means I am contractually obligated to point out that every panic contains an opportunity, and that the opportunity here is not the obvious one.
The obvious opportunity is the oversold bounce β the bet that the selloff is pure sentiment and the fundamentals are intact. I am skeptical of this one, not because CoreWeave is a bad company, but because I do not yet have the evidence to distinguish a sentiment-driven selloff from a fundamental reassessment. The company's own response, and the next earnings report's disclosure of revenue quality and customer concentration, will settle it. Until then, buying the dip is not analysis. It is a coin flip with a narrative attached.
The less obvious opportunity is the dispersion trade β the recognition that compute is a genuine long-term need, but that the risk-adjusted way to own it is not through the highest-leverage operator. The picks-and-shovels of the AI buildout include power producers, cooling specialists, and networking suppliers, many of which carry far less leverage and far more diversified customer bases than the neoclouds. If the demand for compute is real β and I believe it is β then the durable returns accrue to the least leveraged part of the chain, not the most. The market has spent three years rewarding leverage. The next three years may reward its absence.
The third opportunity, the one that requires patience, is consolidation. The neocloud sector is a set of undifferentiated suppliers competing on inputs they do not control. That is not a stable equilibrium. It resolves either through failure or through acquisition, and both outcomes reward the patient buyer who waits for the price to reflect reality rather than narrative. I do not know which neoclouds survive. I know that the sector, as currently structured, cannot all survive, and that the moment of resolution is when the real money is made.
What I Am Watching
I track signals the way I tracked Telegram groups in 2018 β not for what they say, but for who is saying it and why now.
At the company level, the single most informative datapoint will be CoreWeave's own response to the slide, and the next quarterly disclosure of revenue quality. A management that addresses the depreciation question directly is a management that understands its own structure. A management that deflects is a management that is still inside the flywheel.
At the peer level, the key question is synchronization. If the other neoclouds β and the data-center REITs, and the power names β fall together with CoreWeave, the problem is sector beta and the whole complex is repricing. If CoreWeave falls alone, the problem is company alpha and the story is specific. That distinction is the difference between a mood and a diagnosis, and it can be read in a single day's tape.
At the macro level, I am watching the credit market, because that is where this story will actually resolve. Equity prices are noise. The willingness of lenders to refinance GPU-backed debt at a spread that keeps the flywheel spinning is signal. If that willingness contracts, everything else follows. If it holds, the selloff is a shiver, not a fever.
And at the structural level, I am watching the hyperscaler capex guidance and the electricity-allocation policies that will determine where the next generation of data centers can physically be built. Those are the slow variables. They do not move the tape this week. They determine the tape for the next five years.
The Takeaway
So what does the headline actually mean, the one with the verb and the adverb and nothing else?
It means the market has started to ask the question it should have asked three years ago: not whether AI will change everything, but who will be left holding the chips when the financing that bought them comes due. CoreWeave is not the answer to that question. It is the question, expressed as a balance sheet β the highest-leverage, most concentrated, most sentiment-sensitive way to own the compute thesis. When a structure like that slides, it is not telling you that compute is worthless. It is telling you that the market has begun to price the difference between owning a thing and owing for it.
We didn't get the number. We didn't get the cause. But we got the direction, and in a bear market the direction is the only data that matters. The tide is going out. The question now is not who is swimming. It is who was never wearing anything at all.
Every bull run is a myth waiting to be debunked. This one will be debunked not by a crash, but by a credit committee, in a room with no windows, reading a depreciation schedule it should have read a year ago. And somewhere in that room, in the ledger's silence, the true story will whisper β the one about the chip that was collateral, the debt that was faith, and the compute that was never as infinite as the story required.
