There's a warehouse somewhere in the American Northeast where the ghosts of Ethereum still hum in the wiring. In 2017, that building was a crypto mining operation—row after row of GPUs chewing through cheap electricity to mint coins that most of the financial world had never heard of. The company was called Atlantic Crypto back then. It ran on the same faith that powered every mining farm of the era: that raw compute, pointed at the right problem, would eventually be worth more than the electricity it consumed. By 2025, the same physical bones, rewired and rebooted, are reportedly producing $5.1 billion in annual revenue. And the company that owns them, now called CoreWeave, has just told the market to expect somewhere between $12.4 billion and $13.2 billion in 2026.
Say that number slowly, because it deserves the weight. Five point one billion dollars today. A midpoint of twelve point eight billion next year. That is an implied jump of roughly 150 percent—the kind of curve that normally belongs to a vaccine maker during a pandemic, or a startup riding one terrifying wave of demand that everyone quietly assumes will never break. And here is the part that should make anyone who still cares about decentralization sit up straight: the physical and financial plumbing that crypto built—the data halls, the power contracts, the GPU procurement relationships, the sheer unnatural willingness to lose money for years on the bet that compute would someday be the scarce resource—has been absorbed into the most centralized computing project in human history. Most of us let that transaction clear without ever reading the terms.
To understand why CoreWeave's guidance matters beyond its own balance sheet, you have to understand what the company actually is. It belongs to a category the industry awkwardly calls “neoclouds”—specialized GPU infrastructure providers that rent compute by the hour, mostly to AI labs, hyperscalers, and enterprises training or serving machine learning models. It is not a SaaS company. It is not a platform marketplace. It is, stripped of its software veneer, a data center and power company with a Kubernetes control plane bolted on top.
The revenue trajectory tells the story more cleanly than any pitch deck. In 2024, CoreWeave did roughly $1.9 billion. In 2025, it reports $5.1 billion. That is about 165 percent growth. The new 2026 forecast of $12.4 billion to $13.2 billion implies growth of 143 to 159 percent—so growth is decelerating, but only slightly, and from an already absurd base. When I was auditing early Ethereum smart contracts for a boutique consultancy back in 2017, the projects I evaluated raised $10 million and promised the world. CoreWeave is now forecasting more revenue in a single year than the entire market capitalization of most crypto networks that survived the last cycle. The scale has shifted categories. We are no longer comparing apples to apples; we are comparing an orchard to a nation's grain reserve.
And here is the historical rhyme that keeps nagging at me. CoreWeave did not appear from nowhere. It came out of the crypto mining diaspora—the same pool of capital and talent that, when mining Ethereum stopped being profitable after the Merge, needed to find a new place to point its GPUs. That migration is the single most underreported infrastructure story of the decade. The same people who were mocked as reckless speculators built the physical foundation that every major AI lab now stands on. In 2022, when the market fell 70 percent and FTX collapsed, I watched that diaspora scatter. Some of them folded. Some of them pivoted to AI and never looked back. CoreWeave is what the survivors became.
I need to be precise about the mechanics here, because the numbers only make sense once you see the machine underneath them. A forecast like $12.8 billion is not a sales target in the ordinary sense. It is closer to a physical capacity statement. To hit it, CoreWeave does not primarily need better salespeople or a slicker funnel. It needs GPUs delivered, power secured, and buildings commissioned. In the AI compute business, your revenue ceiling is set by how many Nvidia chips you can physically acquire and how much electricity you can buy to run them. The financial projection is downstream of the supply chain, not upstream of it.
This is why the moat is real and shallow at the same time. CoreWeave's advantage—its relationship as one of Nvidia's favored cloud partners—is genuine. When demand outstrips supply by a wide margin, privileged access to allocation is worth more than any software feature you could ship. But it is also fragile, because the relationship lives entirely on Nvidia's side of the table. A company whose primary competitive edge is a supplier's goodwill is, structurally, a tenant in its own house. The same partner that anoints you can rebalance you, and there is no contract language strong enough to prevent it.
Second, this is a capital-intensive business dressed in software language. CoreWeave's margin structure resembles a utility, not a SaaS company. It buys expensive GPUs, depreciates them over a few years, pays enormous power bills, and finances the whole thing with debt. Gross margins may sound impressive by industrial standards, but they will never look like software margins, because the cost of goods sold here is a physical asset that wears out and becomes obsolete on a schedule set by Nvidia's next architecture launch. When I built out tutorials for OpenLedger Academy demystifying yield farming, I learned to compare liquidity pools to community gardens—shared resources that need constant tending and honest accounting. A GPU cloud is the opposite. It is a garden where every tomato plant costs a million dollars and goes rotten in three years regardless of how well you water it. The metaphor matters, because it forces you to look at the depreciation line instead of the growth line.
Third, the customer list is the risk. AI compute demand is currently dominated by a handful of enormous buyers—frontier labs and hyperscalers. That means a small number of contracts likely account for the overwhelming share of revenue. High concentration is not unusual in enterprise infrastructure, but in a business with this much fixed cost, a single lost anchor tenant is not a bad quarter; it is an existential event. The forecast of $12.8 billion quietly assumes that the biggest customers keep expanding, that none of them decide to build their own clusters, and that financing markets stay open the entire time. All three assumptions can break at once, and they are correlated when they break. That correlation is the thing nobody prices until the moment they have to.
There is one part of the CoreWeave story the flash news consistently skips: its acquisition of Weights & Biases, the machine learning experiment-tracking and MLOps platform. On a pure numbers basis it looks like a bolt-on. Strategically, it is the only thing standing between CoreWeave and commodity status. Think about what a GPU cloud actually sells. It sells raw compute. Raw compute is, over time, a commodity—the thing every hyperscaler, every competitor, and every well-funded challenger can eventually offer. What is not a commodity is the workflow layer: the tools developers use to track experiments, manage model versions, and orchestrate training pipelines. If CoreWeave can move from “we rent you GPUs” to “we host the workflow your entire team depends on,” it converts a procurement decision into a habit. That is the difference between a supplier and a platform, and it is the difference between a business that survives the next Nvidia generation and one that gets repriced like a memory chip maker.
But here is where my old audit instincts kick in. I have spent years watching crypto projects claim to be building ecosystems while their actual governance sat in the hands of three anonymous multi-signature wallets. The tell was always the same: a grand ambition narrative on top, a much smaller structural reality underneath. CoreWeave's software story is more credible than those, because W&B is a real product with real users and real retention. But it is still early. The company is, for now, primarily a resource seller with an ecosystem aspiration. The aspiration has not yet been stress-tested by a real downturn, and downturns are where every unproven narrative goes to die.
This is the part where the crypto-native origin of CoreWeave becomes more than trivia. CoreWeave exists because of a governance event. When Ethereum moved to proof-of-stake, it rendered an entire generation of mining hardware and mining capital obsolete overnight. That capital did not disappear. It looked for the next place where raw compute would matter, and it found AI. In a strange way, the Merge was one of the largest capital reallocations in tech history, and almost nobody describes it that way. We treat it as a consensus mechanism change. It was also a forced liquidation of an entire industry's physical plant, and the beneficiaries were not the Ethereum community. They were whoever happened to own the right buildings.
But notice what was not preserved in that migration: the decentralization ethos. Mining was, in theory, a decentralized activity—anyone with hardware could participate, and the network's security came from that openness. The GPU cloud business that replaced it is the opposite. Access is gated by Nvidia allocations, capital is gated by debt markets, and power is gated by utility agreements that take years to negotiate. The crypto industry spent a decade building infrastructure to escape gatekeepers, and then handed that infrastructure to a new set of gatekeepers who happen to wear data center badges instead of banking ones. The gate changed uniforms. The wall did not move.
I keep coming back to the phrase I use when I talk about DAO governance: democracy isn't a download—it's a transaction where every voice holds weight, and transactions have counterparties. The crypto miners who built these data halls never voted to become the foundational layer of centralized AI. There was no governance proposal, no token holder vote, no community discussion about whether the physical commons of crypto should be repurposed to train models owned by five companies. It just happened, because capital flows to wherever the return is highest, and no smart contract was ever written to prevent it. The absence of that contract is the whole story.
This is the same lesson I keep relearning across every corner of this space. The Lightning Network has been half-dead for seven years because routing failure rates and channel management complexity made it a niche tool instead of a payment rail. Layer 2 rollups were supposed to make blockspace abundant, and instead the blob data is already trending toward saturation, which means the gas fee relief was a season, not a settlement. Code, in each case, worked exactly as designed. The design just did not account for what humans actually do with markets. CoreWeave's infrastructure worked exactly as designed too. The design just did not account for the fact that a permissionless origin story can end up serving a permissioned outcome.
Let me try the opposite argument now, because steelmanning is how you avoid becoming the person who has been bearish since 2013. The bullish case is simple and, in the short term, probably correct. AI compute demand is real, and it is not a crypto-style narrative that can evaporate when sentiment turns. Enterprises genuinely need GPUs. Governments genuinely want sovereign AI capability. If CoreWeave can deliver capacity, the revenue follows, and the $12.8 billion forecast is achievable. From this view, the crypto origin is irrelevant—a fun footnote for journalists, not a structural fact. The company is simply the right asset in the right place at the right time, and complaining about its centralization is like complaining that a power plant is centralized.
I think that argument is half right, and the half that is wrong is the half everyone ignores. Here is the pragmatism test: if your business depends on a single supplier's allocation, a single financing market's mood, a single region's power grid, and a handful of customers' expansion plans, then your growth is not a property of your company—it is a temporary condition of the environment. That is exactly the trap the crypto industry fell into in 2021, when everyone confused a liquidity cycle for product-market fit. CoreWeave may be about to make the same mistake at ten times the scale, with infrastructure instead of tokens. The symptoms look different. The disease is identical.
This is also where the comparison to crypto's own infrastructure ambitions bites hardest. Projects like decentralized compute networks—Render, Akash, and their many imitators—promised to turn idle GPUs into a permissionless marketplace. On paper, they solve exactly the centralization problem CoreWeave embodies. In practice, they have not come close to matching the throughput, reliability, or enterprise trust of a centralized provider. Which raises an uncomfortable question: if decentralized compute cannot win even when the demand is exploding and the money is unlimited, when exactly is it supposed to win? The honest answer may be that the market never actually wanted decentralization in compute. It wanted capacity. And capacity, it turns out, centralizes. We built the cathedral for a different god and then filled it with the one we already had.
There is one more ceiling that the flash headline never mentions, and it is the one that will decide whether any of this is durable: energy. The real bottleneck for AI infrastructure is not chips alone. It is the ability to secure power and grid interconnects, which takes years of permitting and billions of dollars of capital, and which is subject to regulators who do not care about Nvidia's roadmap. A data center is a power purchase agreement with a building attached. CoreWeave's forecast is not really a bet on software or even silicon. It is a bet that the electricity keeps flowing and the permits keep getting signed. That is a physical and political constraint, not a technical one, and it is immune to every optimization the industry can throw at it.
So what should you actually track from here? Not the $12.8 billion number—that is the headline, and headlines are what get optimized for. Watch the RPO, the remaining performance obligations that reveal how much of that forecast is contractually locked versus aspirationally hoped. Watch the debt line, because a business funded by borrowing to buy depreciating assets lives or dies on the cost of that borrowing. Watch the customer concentration, because the fastest way to know if the moat is deepening is to see whether revenue is spreading across more buyers or piling into fewer. Watch the power contracts and the export-control landscape, because those are the two places where the ceiling gets set without anyone sending a press release.
The $12.8 billion forecast is a real number, and CoreWeave earned the right to make it. But it was built on infrastructure that crypto assembled and then quietly surrendered, and it now depends on gatekeepers that no token holder ever approved. The compute war was never voted on. It was financed. And that is exactly why it is worth reading the balance sheet as carefully as the press release—because the story of decentralization's future may not be written in a whitepaper at all, but in a depreciation schedule that nobody in the community ever thought to sign.
What happens next will tell us whether crypto was ever a destination, or just a staging ground. If the diaspora that built these halls never turns around to build a permissionless alternative, then we will have our answer: that the industry's greatest export was not decentralization, but the physical capacity to host someone else's moonshot. And the check for that export has already cleared. The question is whether anyone is still watching the account.