We didn't get a demand curve. We got a stage.
That was my first thought watching the clip make the rounds โ Jensen Huang, in his element, pushing back on the most consequential assumption in the entire AI capital stack: that compute demand comes from two laboratories and a handful of hyperscalers, and that everything else is decoration. His rebuttal was emphatic and structurally vague. Demand is broadening. It's enterprises. It's inference. It's sovereign AI. It's the long tail.
I have no problem with the claim. I have a problem with the fact that no one on that stage, no one in the audience, and no one reposting the clip with fire emojis can actually check it.
Here is what kept me up. The company at the center of this โ CoreWeave โ was not born as an AI company. It was born as a crypto miner. Atlantic Crypto, Ethereum, 2017, a converted warehouse full of GPUs pointed at a proof-of-work chain that paid for itself in months and then stopped paying. When the mining margin evaporated, the same founders pointed the same racks at a different workload and became the most important neocloud on earth: a public listing, a multi-billion-dollar debt stack, an NVIDIA relationship that is simultaneously supplier, shareholder, and marketing department, and a customer list you could count on one hand with fingers left over.
So when I say I have seen this movie, I mean literally. I have been the guy with racks and a narrative. In 2020 I ran three yield aggregators into DeFi Summer and watched two million dollars of TVL materialize and then fifteen percent of it vanish through a contract I had not audited. The lesson was not audit your contracts, although yes, obviously. The lesson was that in a bull market the fastest-growing metric is always the one nobody can independently verify.
That is the whole argument. CoreWeave is the cleanest test case we have for how AI compute demand gets priced, and crypto is the place where that pricing gets amplified into something that no longer resembles an underwriting decision. If you hold GPU DePIN tokens, decentralized compute marketplaces, AI-agent infrastructure, or anything that has been repriced upward by the sentence AI needs compute, this piece is your audit.
Let me start with what the company actually is, because the narrative has drifted far from the ledger.
CoreWeave is a GPU cloud. Not a model lab, not a research organization, not a framework company. It rents NVIDIA GPUs, wired with high-speed interconnect, scheduled with Kubernetes and Slurm, delivered as bare metal, priced by the hour. The technical stack is competent and entirely conventional. The moat is not algorithmic. The moat is allocation โ the ability to get high-end NVIDIA silicon ahead of competitors, finance it at scale, and stand it up in weeks rather than quarters.
That distinction matters enormously for how you read the demand-breadth argument. Jensen Huang is not making a technical claim. There is no architecture, no training method, no data-engineering breakthrough in any of this. He is making a market-narrative claim, on behalf of a company that is simultaneously his customer, his portfolio holding, and his demand-proof exhibit. That does not make him wrong. It makes him a stakeholder, and stakeholders do not count as independent evidence.
The financial architecture underneath is where the real analysis lives. CoreWeave's model is built on take-or-pay contracts โ customers commit to paying for capacity whether or not they use it โ funded by debt secured against the GPUs themselves. As reported through 2024 and into its 2025 listing, Microsoft has accounted for the substantial majority of revenue, with public figures around sixty percent, and a large OpenAI commitment was announced alongside the IPO. I would want the current 10-Q before quoting any of those numbers as settled, but the shape is not in dispute: a small number of counterparties, long durations, and a capital structure that converts signed paper into physical iron.
That structure is why the demand-breadth question is existential rather than academic. If demand really is broadening into enterprise inference and sovereign deployments, then the concentration is a temporary artifact of being early, and the contracts are the seed of a diversified book. If demand is still two labs and a hyperscaler, then CoreWeave is a bond with a cloud logo on it, and the spread between what it earns and what it owes is a function of three counterparties' capital expenditure plans.
Now the crypto part, which is where I actually live.
The AI-compute narrative has been imported wholesale into this industry over the last eighteen months, and the import has been remarkably sloppy. GPU DePIN networks list H100 counts in their pitch decks. Compute marketplaces quote aggregate throughput. Agent platforms describe autonomous economic actors that need to rent inference. Every one of these projects is, implicitly, making the same bet Jensen Huang made on that stage โ that compute demand is broad, durable, and will keep flowing down to anyone who can aggregate supply.
Almost none of them have done the work to check whether that is true. And here is the uncomfortable symmetry: the same opacity that makes CoreWeave's demand breadth hard to verify makes every GPU-token supply claim impossible to verify. You cannot audit a GPU from a blockchain. You can audit a signature. The gap between those two things is where an entire sector's valuation currently sits.
I learned this the hard way in a regulatory sandbox in 2024, working on a decentralized identity protocol. The elegant part of DIDs is attestation: a credential signed by an issuer, verifiable without contacting the issuer. The ugly part is that attestation proves a claim was made, not that the claim is true. A signature saying this entity controls ten thousand H100s is cryptographically perfect and epistemically worthless if nobody checks the racks. I spent weeks building a visual guide to explain that distinction to regulators, and the whole time I was thinking about how badly my own industry needs the same document.
Proof of GPU is not proof of compute, and proof of compute is not proof of demand. Three separate claims, routinely collapsed into one sentence in a token whitepaper.
Let me go deeper on each, because this is where the information gain is.
Start with the depreciation problem, which I consider the single most consequential unaudited opinion in the AI trade. A GPU cloud's economics depend entirely on how long it assumes its hardware stays economically useful. Assume six years, and the annual depreciation charge is gentle, margins look robust, and the debt is comfortably serviceable. Assume three years โ closer to the actual cadence at which each NVIDIA generation resets the price-performance frontier โ and the same business looks like a treadmill that has to refinance every cycle just to stand still.
A GPU is not an asset. It is a depreciating option on a workload that may not exist in four years. โ Root: The depreciation schedule is the most consequential opinion in the entire AI trade, and almost nobody outside the filings is auditing it.
I have a very specific scar here. In 2020 I was running yield aggregators and treating TVL as a balance sheet. It was not a balance sheet. It was a commitment that could walk out the door in a single block, and when a minor exploit drained fifteen percent of my liquidity, I discovered that the number I had been reporting to my community had never measured anything durable. I wrote a post-mortem about imperfect innovation and it resonated precisely because everyone in DeFi recognized the delusion. Contracted backlog in a neocloud is the same species of number. It measures commitment, not cash. It is a promise with a duration attached, and promises get renegotiated when the promiser's own funding environment changes.
Which brings me to the second layer: the customers themselves are financed by the same narrative they are financing. If a large share of your revenue comes from AI labs whose own spending is underwritten by venture capital and sovereign wealth enthusiasm, then your revenue quality is a function of someone else's fundraising cycle. That is not a criticism unique to CoreWeave โ it is the structure of the entire compute buildout. But it does mean that the demand-breadth claim is not just a growth question. It is a duration-matching question. Long-duration debt against short-duration narrative is the classic failure mode, and I have watched it destroy crypto lenders twice.
Third layer: the concentration of allocation. CoreWeave's competitive position depends on continued preferential access to NVIDIA silicon. That access is the moat, and it is also the dependency. NVIDIA benefits from having a visible, fast-growing, GPU-pure neocloud that proves demand exists beyond the hyperscalers โ it is a demonstration asset as much as a customer. But NVIDIA will not sacrifice its relationships with AWS, Azure, and Google to protect it, and those three are building custom silicon precisely to reduce the rent they pay. The moment custom accelerators reach adequate capability for the majority of inference workloads, the price of rented NVIDIA capacity compresses, and the neocloud with the most debt and the least diversified customer base absorbs the compression first.
I have been watching an almost identical structure play out in rollups. A Layer 2 sequencer is, functionally, a single centralized node with a privileged position in a supply chain, marketed with a decentralization roadmap that has been a slide deck for two years. The operator captures the ordering revenue, the token absorbs the governance story, and the actual decentralization is perpetually one upgrade away. Neoclouds are the same shape with different hardware. A single operator, a privileged allocation, a narrative of eventual openness, and a customer base that has no practical alternative at the current price. If you understand why sequencer decentralization keeps slipping, you understand why compute decentralization keeps slipping, and it has nothing to do with the technology being hard.
Now the part I find genuinely interesting, and where I think most people are looking the wrong way.
Suppose Jensen Huang is right. Suppose demand really is broadening โ enterprise inference, sovereign deployments, industry models, agent workloads. What happens to crypto compute tokens in that world? The intuitive answer is that they moon. The correct answer is that they get squeezed, because broadening demand at the margin is demand for reliability, contractual guarantees, compliance posture, and data residency. That is the opposite of what a permissionless aggregator of heterogeneous consumer GPUs sells. Enterprises buying inference for a regulated workflow do not want a spot market. They want an SLA and an entity to sue.
This is the same trap that swallowed the RWA narrative for three years. The pitch was that tokenization would bring trillions of dollars of real-world assets on-chain. The reality is that the institutions holding those assets do not need a public chain to do a private transfer, and the ones who do engage are using permissioned environments that look nothing like the original vision. Compute is following the identical path. The demand that is actually broadening โ the enterprise, the sovereign, the regulated inference buyer โ is structurally hostile to the decentralization premium. They will pay more for less openness, every time, and they will call it risk management.
Which means the honest framing for decentralized compute is not we are the future of AI infrastructure. It is: we serve the segment that cannot get allocation from the incumbents, at a discount, with worse guarantees. That is a real business. It is a niche business. It is not a hundred-billion-dollar business, and the tokens are priced as if it is.
Here is my contrarian read, and it is going to annoy people on both sides.
The demand-breadth debate is a distraction, because both outcomes are bearish for the highest-beta expressions of the trade. If Jensen is wrong โ if demand really is two labs and a hyperscaler โ then the entire neocloud complex is a concentrated credit bet dressed as growth equity, and the crypto tokens that borrowed the narrative are the furthest out on the risk curve and get liquidated first. If Jensen is right โ if demand broadens โ then compute becomes commoditized at the margin, margins compress toward the cost of capital, and the value migrates to whoever controls power, land, transformers, and grid interconnection. Notice that none of those things are on a blockchain.
That is the blind spot. Everyone is arguing about whether the demand exists. Almost nobody is arguing about where the profit lands if it does. In a genuine compute-abundance scenario, the scarce inputs are electricity, substations, cooling water, and permitted land โ not GPU hours and certainly not tokenized GPU hours. I have watched this exact error before, in 2021, when a community of five thousand NFT holders and I discovered that the value of a digital residency right was entirely dependent on the physical jurisdiction behind it. The art was on-chain. The right to live somewhere was not. When the floor dropped eighty percent, the people who stayed were the ones who had understood which part was real.
The pragmatist's test for any decentralized compute project is therefore brutally simple. Ask what happens to the business if NVIDIA doubles allocation to hyperscalers tomorrow and rental prices fall forty percent. If the answer is we pivot to being a marketplace, the project was a spread trade, not an infrastructure company. Ask what the customer is actually buying โ compute, or an option on compute they cannot get elsewhere. If it is the second, your revenue is a function of someone else's supply constraint, which means your revenue disappears precisely when the constraint eases, which is precisely when you were planning to scale.
And ask the question nobody wants to answer on stage: who verifies the demand? Not the CEO of the supplier. Not the founder of the marketplace. Not the token holder voting on a governance proposal. Somebody has to look at the racks, the utilization curves, the contract terms, the cancellation clauses, and the counterparty's own funding runway. That is unglamorous work, and it is the only work that matters.
I want to be fair to the other side, because there is a version of this where I am wrong and it matters.
If inference demand genuinely explodes โ not training, inference, at consumer scale, embedded in every application and agent that runs โ then the shape of the market changes in a way that favors distributed supply. Inference is latency-sensitive and geographically distributed. It wants to be near users, not in a single campus in Virginia. That is a real structural argument for smaller, edge-adjacent, heterogeneous compute providers, and it is the strongest case the DePIN crowd has. I have built agent infrastructure myself, on a chaotic testnet with multiple model providers wired together, and I can tell you from direct experience that the bottleneck for autonomous agents is not raw FLOPS. It is reliable, cheap, low-latency, verifiable inference at the edge, plus a legal identity to sign the contract. I wrote an essay arguing for digital personhood on economic grounds and got invited to Lisbon to defend it, and the thing I kept saying on stage is that agency without verifiable execution is just a chatbot with a wallet.
But note what that argument actually requires. It requires verification infrastructure that does not exist yet at production quality. It requires that the distributed supply be auditable to a standard an enterprise will accept. And it requires that the economics work when the incumbent can simply build edge points of presence with cheaper capital. Those are solvable problems. They are not solved problems, and the market is currently priced as though they are.
So what do I actually watch? Four things, and none of them are conference keynotes.
First, the share of NVIDIA data center revenue that is not attributable to the largest hyperscalers and labs. That is the closest thing to an independent read on demand breadth, and it is buried in disclosure that requires work to extract. Second, GPU rental spot prices for current-generation parts. If breadth is real, spot prices should hold as supply expands. If breadth is narrative, spot collapses the moment a new generation ships. Third, backlog conversion โ the ratio of contracted commitments that actually convert to billed revenue, quarter over quarter, and any disclosure about renegotiated or deferred terms. Fourth, and most important for my own readers: whether any decentralized compute network publishes audited utilization data, not capacity. Capacity is a claim. Utilization is a business.
I will be honest about my bias. I am a decentralization believer who has spent thirteen years watching decentralization claims get made and then quietly redefined. I wrote a forty-page manifesto in 2017 and handed out five hundred printed copies at a hacker space in Tallinn because I genuinely believed code could function as law and that this would distribute power. Some of that belief survived. A lot of it got rewritten by watching what people actually do when incentives arrive.
The thing that survived is narrower and, I think, harder. Sovereignty is not a token. It is the ability to verify the thing you depend on. If you cannot audit the compute, you do not own the compute โ you are renting a narrative from someone who can change the terms. That applies to a neocloud with a concentrated customer base and a depreciation schedule that is really an opinion. It applies to a GPU network that lists capacity it cannot prove. And it applies to the entire AI trade, which has spent two years pricing an unauditable demand curve as though it were a disclosed fact.
Jensen Huang may well be right that demand is broadening. I hope he is. But the question I would put to him, and to every founder building on his answer, is not whether demand is broad. It is: who gets to check? Because a market that cannot answer that question is not pricing infrastructure. It is pricing a story, and stories reprice fastest at exactly the moment everyone is most certain.